Medical clean air source generator with pressure regulation function and regulation method of medical clean air source generator

Abstract

The invention provides a medical clean air source generator with a pressure adjusting function, and belongs to the field of medical instruments. Comprising a cabinet, and an air compressor, a condensation mechanism, a gas-liquid separation mechanism, a drying system, a filtering system and a gas storage tank which are connected in sequence. The control module and the pressure adjusting module are arranged, an optimization model is arranged in the control module, a future demand curve can be predicted according to historical gas consumption data, the optimal pressure set value is generated in real time, and then the pressure adjusting module dynamically adjusts the output gas pressure. By means of the design, the problem of pressure fluctuation caused by a traditional fixed pressure or simple feedback mode is solved, and continuous and stable supply of the medical air source is ensured.

Description

A medical clean gas generator with pressure regulation and its regulation method Technical Field

[0001] This application belongs to the field of medical device technology, specifically relating to a medical clean air source generator with pressure regulation. Background Technology

[0002] In the medical field, clean air sources are a crucial foundation for ensuring the normal operation of medical equipment and patient safety. Traditional medical clean air generators typically employ simple compression, cooling, and filtration processes, and mainly consist of basic components such as air compressors, condensers, filters, and air tanks.

[0003] For example, Chinese patent CN116557270A provides a clean gas generator. In this patented technology, the gas pressure relies solely on the buffer of the gas storage tank, and the output pressure cannot be dynamically adjusted according to the load at the gas-consuming end. The control is not precise enough, making it difficult to meet the diverse gas pressure requirements of different medical devices. It also lacks intelligent predictive and control functions, and cannot dynamically adjust according to actual gas demand, leading to energy waste and low operating efficiency. Furthermore, existing clean gas dehumidification technologies generally use heating and vaporization. This dehumidification technology is prone to incomplete dehumidification, easily forming water mist after the clean gas cools. A humid environment is prone to bacterial and mold growth, affecting medical efficacy and posing safety hazards. Therefore, there is an urgent need to design a clean gas generator that can dynamically adjust pressure and achieve high dehumidification efficiency. Summary of the Invention

[0004] To solve the above problems, the technical solution adopted in this application is as follows: The first aspect of this application provides a medical clean gas generator with pressure regulation, comprising: a cabinet for support; an air compressor disposed above the cabinet for drawing gas; a condensing mechanism connected to the outlet of the air compressor for cooling the gas; a fan disposed on one side of the condenser for cooling the condenser; a gas-liquid separation mechanism disposed at the rear end of the condenser for separating liquid; a drying system disposed at the rear end of the gas-liquid separation mechanism for drying the gas; a filtration system disposed at the rear end of the gas storage tank for filtering the gas; a gas storage tank disposed at the rear end of the drying system and connected to the filtration system for storing and outputting gas; a control module for predicting future gas demand and generating an optimal pressure setpoint; and a pressure regulation module for receiving instructions from the control module and adjusting the output gas pressure in real time.

[0005] Furthermore, the gas-liquid separation mechanism includes: a separation cylinder, comprising two sets of separation cylinders, the lower part of which is provided with a conical liquid collecting section for collecting liquid; a tangential air inlet, located on the upper side wall of the separation cylinder, for tangentially introducing the mixed gas and liquid into the cylinder and forming a rotating airflow; a gas outlet, located at the top of the separation cylinder and connected to a drying system; a centrifugal separation component, consisting of a tangential air inlet and a cylindrical inner cavity of the separation cylinder, for achieving primary gas-liquid separation through centrifugal force; a wire mesh demister, located inside the separation cylinder and on the airflow path between the tangential air inlet and the gas outlet, for capturing residual liquid droplets in the gas after passing through the centrifugal separation component; and an air inlet switching valve, located at the beginning of the tangential air inlet, which controls one and only one separation... The cylinders are connected for replacing and discharging liquids; a drain valve, located at the bottom of the conical liquid collection section, is used to automatically or manually drain accumulated liquid; an exhaust switching valve is used to merge the gas outlets of the two separation cylinders into a single total gas outlet; when the exhaust switching valve is used, its operation is synchronized with the intake switching valve; detection sensors, located at the upper end of the separation cylinders, are used to detect the liquid level; the detection sensors, the intake switching valve, and the drain valve are electrically connected to the control module; the controller module is configured to: control the intake switching valve to switch its conduction path based on the feedback signal from the detection sensors, and control the drain valve of the currently non-working cylinder to open for liquid drainage, thereby enabling the two separation cylinders to work alternately and drain liquid online, ensuring uninterrupted gas-liquid separation.

[0006] Furthermore, the air compressor is a swing piston type oil-free air compressor, which is installed in the cabinet through a spring shock absorber, with a noise level ≤60dBA and an exhaust volume ≥100L / min. The bottom of the gas-liquid separation mechanism is connected to an automatic drain solenoid valve. The fan is an axial flow fan, which is located on the air inlet or air outlet side of the condensation mechanism. The drying system is a dual-tower adsorption drying device, which is connected to a switching valve group and is controlled by the control module to alternate between adsorption and regeneration states.

[0007] Furthermore, the filtration system includes three stages of filtration: the first stage is a 0.3μm micron-level filter, the second stage is a 0.01μm terminal filter, and the third stage is an ultraviolet sterilization filter, so that the final output gas meets the ISO 8573-1 Class 0 cleanroom standard.

[0008] Furthermore, the control module is deployed on the edge computing unit, and its built-in prediction model is the Transformer time-series fusion model, which is used to predict the gas demand curve in future periods based on historical gas consumption data sequences.

[0009] Furthermore, the pressure regulation module includes a precision pressure regulating valve, a first pressure sensor for monitoring the pressure of the gas storage tank, and a second pressure sensor for monitoring the pressure at the output end, forming a closed-loop pressure control system.

[0010] The second aspect of this application provides a method for regulating a medical clean gas generator with pressure regulation, comprising the following steps: Step S1, collecting historical gas consumption data, equipment start / stop signals, and environmental parameters to construct a time-series feature vector; Step S2, inputting the data from Step S1 into a Transformer time-series fusion model deployed on an edge AI chip; Step S3, the model outputs a gas demand curve for the next 1-3 minutes and an optimal pressure setpoint, and uses quantile prediction to provide uncertainty estimation; Step S4, receiving model instructions through a high-speed servo pressure regulation module and implementing closed-loop feedback control in conjunction with a piezoelectric microsensor; Step S5, a reinforcement learning module performs online fine-tuning of the model based on pressure error, energy consumption, and equipment lifespan loss to optimize the pressure regulation strategy.

[0011] Furthermore, the time-series feature vector of step S1 includes: gas tank pressure, output flow rate, gas temperature, compressor current, filter differential pressure, ambient temperature and humidity, time encoding, static encoding of department type, and equipment start / stop mask, totaling 9 dimensions; the sampling frequency is 100ms, the historical window length is 180 steps, and the prediction window length is 18 steps; the data is normalized by moving mean and standard deviation.

[0012] Furthermore, the Transformer temporal fusion model in step S2 specifically includes the following sub-steps: Step S21, a static feature encoder, which encodes the department type into a context vector; Step S22, a variable selection network, which calculates the importance weights of each feature through Softmax and automatically selects key sensor signals; Step S23, an LSTM encoder, which extracts historical temporal context features; Step S24, a multi-head self-attention mechanism, with 4 heads and a model dimension of 128, which captures pressure fluctuation patterns; Step S25, a gated residual connection, which fuses the attention output with the original features through learnable gating coefficients to prevent gradient vanishing; Step S26, a quantile output layer, which predicts the P10, P50, and P90 quantile values ​​of gas demand in the next 18 steps.

[0013] Furthermore, the reinforcement learning module in step S5 employs the Proximal Policy Optimization algorithm, specifically including the following sub-steps: Step S51, the state space is defined as a 9-dimensional vector: ,in: For the pressure of the gas storage tank, For output flow, For gas temperature, For compressor current, For filter pressure difference, For environmental humidity, For time encoding, Static coding for department types The device start / stop mask is used; Step S52, the action space is a discrete valve opening adjustment amount, ranging from [-5%, +5%], with a step size of 0.5% for a total of 21 actions; Step S53, the reward function is: In the formula: Set a value for the target pressure. Let be the actual pressure at time t. For pressure deviation, It is the square of the change in valve opening. For indicator functions, The AI ​​model predicts the pressure at time t, where the first term is the pressure error penalty, the second term is the energy consumption and valve wear penalty, the third term is the overpressure or underpressure safety penalty, and the fourth term is the prediction accuracy reward; Step S54, Algorithm Execution Flow: Collect the strategy execution trajectory every 4 hours, with a trajectory length of 1440 steps; Calculate the advantage function using the generalized advantage estimation algorithm. In the formula In exchange for a discount, Let be the state-value function, where the discount factor γ = 0.99 and the GAE parameter λ = 0.95; the trajectory data is divided into mini-batches of 256 data points each, and 10 rounds of iterative training are performed; the policy update magnitude is limited by a pruning coefficient ε = 0.2, and the objective function is: In the formula, For the clipping function, The minimum value function is used; Step S55, parameter optimization: The Adam optimizer is used to update the policy network parameters, with an initial learning rate α = 3 × 10⁻⁶. -4 During the initial training phase, Ornstein-Uhlenbeck exploration noise was added to the action space. After training was completed, the system was switched to deterministic policy execution.

[0014] Furthermore, the closed-loop feedback in step S4 adopts a feedforward-feedback composite control: In the formula, the feedforward weights The feedback coefficient is 0.8. The integral coefficient is 0.3. The error is 0.05. The difference between the set pressure and the actual pressure.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: 1. This application provides a medical clean gas source generator with pressure regulation. Through a control module with a built-in pressure regulation optimization model, this application can predict the gas demand curve for future periods based on historical gas consumption data sequences. Compared with traditional fixed pressure or simple feedback regulation systems, this invention can generate the optimal pressure setpoint in real time, avoiding the problem of unstable gas consumption caused by pressure fluctuations and ensuring the continuity and stability of the medical gas source.

[0016] 2. This application provides a medical clean air source generator with pressure regulation. This application adopts a swing piston type oil-free air compressor with noise controlled at ≤60dB(A), which is far lower than the noise level of traditional air compressors, providing quieter operating conditions for the medical environment; meeting the demand for high-flow medical gas and improving the efficiency of equipment use.

[0017] 3. This application provides a medical clean air source generator with pressure regulation. The gas-liquid separation mechanism of this application is connected to an automatic drain solenoid valve at the bottom, which can promptly drain condensate, preventing moisture from entering the subsequent drying and filtration systems and extending the service life of the equipment. Compared with traditional equipment with manual drainage or no drainage design, this significantly improves the efficiency and reliability of gas-liquid separation. Attached Figure Description

[0018] Figure 1 is a front view of the device of this application; Figure 2 is a front sectional view of the device of this application; Figure 3 is a rear view of the device of this application; Figure 4 is a rear sectional view of the device of this application; Figure 5 is a right view of the device of this application; Figure 6 is a structural diagram of the gas-liquid separation mechanism of this application; Figure 7 is a front view of the gas-liquid separation mechanism of this application; Figure 8 is a top view of the gas-liquid separation mechanism of this application; Figure 9 is a sectional view of Figure 7 along section AA; Figure 10 is a flow chart of the adjustment process of this application; Figure 11 is a diagram of the pressure adjustment method in this application.

[0019] In the diagram: 1. Cabinet, 2. Air compressor, 3. Condensation mechanism, 4. Fan, 5. Gas-liquid separation mechanism, 51. Separation cylinder, 511. Conical liquid collection section, 52. Tangential air inlet, 53. Gas outlet, 54. Centrifugal separation component, 55. Air inlet switching valve, 56. Drain valve, 58. Detection sensor, 59. Wire mesh demister, 6. Drying system, 7. Filtration system, 8. Air tank, 9. Control module, 10. High-speed servo pressure regulation, 11. Casters. Detailed Implementation

[0020] The present application will be further described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present application.

[0021] Example 1, as shown in Figures 1 to 9, provides an embodiment of a pressure-regulated medical clean air source generator. Using a rocking piston oil-free air compressor 2 as the core component, it integrates a shock absorption system, a drying system 6, a filtration system 7, and an air storage tank 8. In this application, a 750W rocking piston compressor is used as the air compressor 2, with post-processing and filtration. The air is stored in a 15L air storage tank 8 and, after pressure regulation, delivered to the endoscope cleaning equipment or dental chair. This application includes a cabinet 1 and internally integrated components such as the rocking piston air compressor 2, condenser 3, cooling fan 4, pressure switch, gas-liquid separation mechanism 5, solenoid valve, drying system 6, filtration system 7, air storage tank 8, and control module 9, as well as pressure regulating valves, pressure gauges, safety valves, and other major components. The rocking piston compressor is mounted on the cabinet 1 with spring shock absorption. Combined with noise-reducing cotton pasted inside the cabinet 1, the noise of the rocking piston compressor during operation can be reduced to an extremely low level. The swing-type piston compressor draws ambient air from the environment through an air intake filter. The high-pressure air generated by the piston structure enters the condenser 3. A fan 4 is installed next to the condenser 3 to draw ambient air into the cabinet 1, cooling the condenser 3 and the cabinet 1. The high-pressure, high-temperature, and high-humidity gas generated by the compressor head is cooled accordingly, and some of the gaseous water becomes liquid water, which then enters the gas-liquid separation mechanism 5 (filtration accuracy 5μm) through the pipeline. The gas-liquid separation mechanism 5 separates the liquid from the gas. The liquid water is discharged through the drain valve under the separation mechanism 5, thus discharging the liquid water from the equipment. The separated compressed gas enters the subsequent drying system 6, which contains a drying adsorbent to adsorb the gaseous gas in the high-pressure gas. The water drying system 6 switches between cyclic drying and adsorption via a switching control board and a solenoid valve. The dried gas enters the gas storage tank 8 through a one-way valve. The gas storage tank 8 is installed on the base plate of the cabinet 1. When the pressure in the gas storage tank 8 reaches the preset pressure switch, the machine head stops. The solenoid valve of the gas-liquid separation mechanism 5 discharges the water in the pipeline to the outside of the cabinet. The compressed air generated by the machine head is temporarily stored in the gas storage tank. When the customer uses it, the compressed air in the gas storage tank 8 will enter the second-stage filter (0.3μm) through a hose and a pressure regulating valve. The treated compressed air will then enter the external third-stage filter (0.01μm) through a hose output device and an exhaust ball valve for use by the endoscope cleaning equipment or dental chair.

[0022] Control module 9 is an embedded system based on an industrial computer or a high-performance programmable logic controller (PLC), specifically including: a central processing unit (CPU): used to run the pressure regulation optimization model and core control logic; a memory: storing historical operating data, trained model parameters, real-time data, and control programs; data acquisition interfaces: a pressure sensor interface: real-time acquisition of the outlet pressure of air compressor 2, the pressure of air tank 8, and the outlet pressure of filter system 7 (i.e., terminal pressure); a flow sensor interface: real-time acquisition of the terminal air flow rate; an equipment status interface: receiving the operating status (such as start / stop, fault signals) of equipment such as air compressor 2, drying system 6, and fan 4; a clock module: providing accurate timestamps for establishing the time series of data; control output interfaces: a high-speed servo pressure regulation module 10 control interface: outputting the optimal pressure setpoint (digital or analog signal); an air compressor 2 start / stop and loading / unloading control interface: serving as an auxiliary adjustment means; and a human-machine interface (optional but recommended): a touch screen, used to display the prediction curve, pressure setpoint, system status, and allowing operators to input preset air consumption plans or adjust model parameters. Together with the high-speed servo pressure regulation module 10, a prediction-optimization-execution closed loop is formed. The control module 9 provides intelligent setpoints, while the high-speed servo pressure regulation module 10 is responsible for high-precision, fast-response physical regulation. With major energy-consuming equipment such as the air compressor 2, the optimization strategy generated by the control module 9 can achieve energy savings on a larger time scale by coordinating the pressure setpoints with the loading / unloading of the air compressor, avoiding frequent start-stop operations.

[0023] In this application, the external dimensions of cabinet 1 are: width * thickness * height = 450 * 590 * 590 cm. The overall color of cabinet 1 is white with a silver sheen, and the panel has blue patterns. The overall structure is compact and highly integrated. The front of cabinet 1 is equipped with a pressure regulating LCD control screen, a gas tank pressure gauge, and an output pressure gauge. The back of the equipment has a drain outlet, an exhaust outlet, a power switch, and a precision filter. The side cabinets are secured with triangular door locks and door hinges. The bottom and top are detachable, and the left and right side cabinet doors are also detachable, allowing for individual disassembly, maintenance, or replacement of parts. The bottom is equipped with locking swivel casters 11, allowing for flexible movement and silent pushing.

[0024] The gas-liquid separation mechanism 5 includes two sets of parallel-connected separation cylinders 51. Each separation cylinder 51 has a conical liquid collecting section 511 at its lower part for collecting liquid. A tangential air inlet 52 is provided on the upper side wall of the separation cylinder 51 to tangentially introduce the mixed gas and liquid into the cylinder and form a rotating airflow. Each separation cylinder 51 has a gas outlet 53 at its top, which is connected to the subsequent drying system 6. The interior of each separation cylinder 51 constitutes a centrifugal separation assembly, specifically formed by the tangential air inlet 52 and the cylindrical inner cavity of the separation cylinder 51. Specifically, a spiral groove is provided on the inner wall of the separation cylinder 51, which is tangential to the inlet of the tangential air inlet 52, forming a spiral rotating downward airflow, generating centrifugal force to separate the gas and liquid. Its function is to achieve primary gas-liquid separation through centrifugal force. Inside each separation cylinder 51, a wire mesh demister 59 is installed along the airflow path between the tangential air inlet 52 and the gas outlet 53 to capture residual fine droplets in the gas after separation by the centrifugal separation component. The wire mesh demister 59 is a multi-layer stainless steel wire mesh structure with a wire diameter of 0.1–0.25 mm and a porosity of 97%–99%. It is installed on a support ring below the gas outlet 53 to intercept and condense the fine droplets, causing them to fall into the conical liquid collection section 511.

[0025] To achieve alternating and uninterrupted operation of the two separation cylinders 51, the mechanism includes the following control and switching components: an intake switching valve 55 is located at the beginning of the tangential intake port 52, configured to allow only one separation cylinder 51 to receive air at any given time. A drain valve 56 is installed at the bottom of each conical liquid collecting section 511 for automatically or manually draining accumulated liquid. An exhaust switching valve is used to combine the gas outlets of the two separation cylinders 51 into a single general gas outlet 53; the operating state of the exhaust switching valve is synchronized with that of the intake switching valve 55. A detection sensor 58 is installed at the upper end of each separation cylinder 51 for real-time detection of the liquid level within the cylinder. The detection sensor 58, the intake switching valve 55, and each drain valve 56 are electrically connected to the control module 9. The control module 9 is configured to automatically control the air intake switching valve 55 to switch the airflow path based on the liquid level signal fed back by the detection sensor 58, and simultaneously control the drain valve 56 corresponding to the currently non-working separation cylinder 51 to open for draining. Through this control logic, the two separation cylinders 51 are automatically alternated between the separation working state and the draining maintenance state, thereby ensuring the continuous and uninterrupted operation of the entire gas-liquid separation process.

[0026] The air compressor 2 is a swing piston type oil-free air compressor, which is installed in the cabinet 1 through a spring shock absorber. The noise level is ≤60dBA and the exhaust volume is ≥100L / min. The bottom of the gas-liquid separation mechanism 5 is connected to an automatic drain solenoid valve. The fan 4 is an axial flow fan, which is set on the air inlet or air outlet side of the condensation mechanism 3. The drying system 6 is a double-tower adsorption drying device, which is connected to a switching valve group and is controlled by the control module 9 to alternate between adsorption and regeneration states.

[0027] The filtration system 7 comprises three stages of filtration: a first stage is a 0.3μm micron-level filter, a second stage is a 0.01μm terminal filter, and a third stage is an ultraviolet sterilization filter. The final output gas meets the ISO 8573-1 Class 0 cleanliness standard. The filtration system 7 is located at the rear end of the gas storage tank 8 and is used to filter the gas. The filtered gas is then delivered to medical equipment.

[0028] The control module 9 is deployed on the edge computing unit, and its built-in prediction model is a Transformer time-series fusion model, which is used to predict the gas demand curve for future periods based on historical gas consumption data sequences. The Transformer time-series fusion model is an encoder-decoder architecture, in which the encoder is used to extract historical gas pressure, flow rate, and gas consumption period characteristics, and the decoder is used to generate a gas consumption prediction sequence for a specified future time step, which is then converted into the optimal pressure setpoint of the gas storage tank 8 by the dynamic programming algorithm of the control module 9.

[0029] The high-speed servo pressure regulation module 10 includes a precision pressure regulating valve, a first pressure sensor for monitoring the pressure of the gas storage tank, and a second pressure sensor for monitoring the pressure at the output end, forming a closed-loop pressure control system.

[0030] This application embodiment also includes a human-machine interaction module, which is connected to the control module 9, for displaying real-time pressure, gas quality information, predicted gas consumption curves and equipment status, and allowing users to set target pressure ranges or receive maintenance alarm information.

[0031] The motor of the swing piston oil-free air compressor 2 is a permanent magnet synchronous servo motor, which is electrically connected to the control module 9. The control module 9 adjusts the speed of the servo motor according to the demand curve output by the prediction model to achieve variable load matching of compressed air volume.

[0032] Example 2, as shown in Figures 10 and 11, also provides an embodiment of a method for regulating a medical clean air source generator with pressure regulation, comprising the following steps: Step S1, collecting historical gas usage data, equipment start / stop signals, and environmental parameters to construct a time-series feature vector; sensing and aggregating the data, storing the data in a memory to form a historical database; the system collects multi-dimensional data from sensors such as air compressors, air tanks, and filters at a high frequency of 100ms in real time, and accurately constructs a time-series vector containing 9-dimensional features as described in the claims. Normalization processing ensures the stability of subsequent model inputs.

[0033] Step S2: Input the data from Step S1 into the Transformer time-series fusion model deployed on the edge AI chip; extract key features from historical data, and predict future gas demand based on the extracted features, including but not limited to working hours, flow rate change rate, and medical equipment characteristics; Step S3: The model outputs the gas demand curve for the next 1-3 minutes and the optimal pressure setpoint, and uses quantile prediction to provide uncertainty estimation; the core optimization objective is to "minimize the total system energy consumption while meeting future predicted demand." System energy consumption mainly comes from air compressor 2 and drying system 6, which are strongly correlated with output pressure; Steps S2 and S3 are the core prediction and decision-making stages, and the processed data is input into the Transformer time-series fusion model deployed on the edge AI chip. This model not only outputs the gas demand prediction curve for the next 1-3 minutes (P50 median), but also provides the uncertainty range (risk boundary) of the prediction through the quantile output layer (P10, P90), enhancing the robustness of the system. The model finally calculates the optimal pressure setpoint based on the prediction results, which serves as the control command.

[0034] Step S4: The high-speed servo pressure regulation module receives model instructions and combines them with piezoelectric microsensors to achieve closed-loop feedback control. The high-speed servo pressure regulation module (such as a precision pressure regulating valve) receives the optimal setpoint instruction and combines it with the piezoelectric microsensor (second pressure sensor) installed at the output end to perform millisecond-level closed-loop feedback, accurately and quickly adjusting the pressure of the final output gas to meet the immediate needs of medical equipment.

[0035] Step S5: The reinforcement learning module fine-tunes the model online based on pressure error, energy consumption, and equipment lifespan loss to optimize the pressure regulation strategy. It compares the difference between predicted and actual flow rates, as well as the actual effect of pressure regulation. Using this deviation data, it periodically (e.g., daily at dawn) performs offline incremental training on the LSTM prediction model to better adapt it to the hospital's actual gas usage patterns, achieving continuous evolution of model performance. It continuously monitors the regulation effect (e.g., pressure error, air compressor energy consumption, valve actuation frequency, etc.) and evaluates the long-term costs (performance, energy consumption, lifespan) of the current regulation strategy through the reinforcement learning module. This module periodically fine-tunes the parameters or control strategy of the Transformer model online, enabling the system to adapt to changes in the actual operating environment and equipment wear, achieving continuous optimization.

[0036] The time-series feature vector of step S1 includes: gas tank pressure, output flow rate, gas temperature, compressor current, filter differential pressure, ambient temperature and humidity, time encoding, static encoding of department type, and equipment start / stop mask, totaling 9 features; the sampling frequency is 100ms, the historical window length is 180 steps, and the prediction window length is 18 steps; the data is normalized by moving mean and standard deviation.

[0037] The Transformer temporal fusion model in step S2 specifically includes the following sub-steps: Step S21, a static feature encoder, which encodes the department type into a context vector; Step S22, a variable selection network, which calculates the importance weights of each feature through Softmax and automatically selects key sensor signals; Step S23, an LSTM encoder, which extracts historical temporal context features; Step S24, a multi-head self-attention mechanism, with 4 heads and a model dimension of 128, which captures pressure fluctuation patterns; Step S25, a gated residual connection, which fuses the attention output with the original features through learnable gating coefficients to prevent gradient vanishing; Step S26, a quantile output layer, which predicts the P10, P50, and P90 quantile values ​​of gas demand in the next 18 steps.

[0038] The reinforcement learning module in step S5 employs the Proximal Policy Optimization algorithm, specifically including the following sub-steps: Step S51, the state space is defined as a 9-dimensional vector: ,in: For the pressure of the gas storage tank, For output flow, For gas temperature, For compressor current, For filter pressure difference, For environmental humidity, For time encoding, Static coding for department types The device start / stop mask is used; Step S52, the action space is a discrete valve opening adjustment amount, ranging from [-5%, +5%], with a step size of 0.5% for a total of 21 actions; Step S53, the reward function is: In the formula: Set a value for the target pressure. Let be the actual pressure at time t. For pressure deviation, It is the square of the change in valve opening. For indicator functions, The AI ​​model's predicted pressure at time t is calculated using four terms: first, a pressure error penalty; second, an energy consumption and valve wear penalty; third, an overpressure or underpressure safety penalty; and fourth, a prediction accuracy reward. By penalizing errors, the AI ​​model is forced to learn how to finely adjust the valves, ensuring the actual pressure quickly and smoothly tracks the target value and minimizes fluctuations. This is crucial for medical equipment (such as ventilators and anesthesia machines), as pressure instability directly impacts treatment effectiveness and patient safety. Mechanical wear on valves and actuators is positively correlated with their frequency and amplitude of action. Penalizing large variations can significantly reduce equipment wear, extend the lifespan of critical components, and reduce maintenance costs and downtime risks. Overpressure can damage downstream precision medical equipment or tubing; underpressure can lead to insufficient gas supply to medical equipment, endangering patients. This penalty is typically set with a very high weight to ensure the AI ​​model maintains the pressure within an absolutely safe range during any learning or operational process, giving it the highest priority. Accurate prediction means the AI ​​can anticipate whether the pressure will rise or fall, allowing for preemptive valve adjustments rather than scrambling to correct errors after they occur. This can significantly reduce overshoot and improve the system's response speed and overall stability.

[0039] Step S54, Algorithm Execution Flow: Collect the policy execution trajectory every 4 hours, with a trajectory length of 1440 steps; calculate the advantage function using the generalized advantage estimation algorithm. In the formula In exchange for a discount, Let be the state-value function, where the discount factor γ = 0.99 and the GAE parameter λ = 0.95; the trajectory data is divided into mini-batches of 256 data points each, and 10 rounds of iterative training are performed; the policy update magnitude is limited by a pruning coefficient ε = 0.2, and the objective function is: In the formula, For the clipping function, The minimum value function is used; Step S55, parameter optimization: The Adam optimizer is used to update the policy network parameters, with an initial learning rate α = 3 × 10⁻⁶. -4 During the initial training phase, Ornstein-Uhlenbeck exploration noise was added to the action space. After training was completed, the system was switched to deterministic policy execution.

[0040] The closed-loop feedback in step S4 adopts a feedforward-feedback composite control: In the formula, the feedforward weights The feedback coefficient is 0.8. The integral coefficient is 0.3. The error is 0.05. The difference between the set pressure and the actual pressure.

[0041] This application utilizes a control module with a built-in pressure regulation optimization model to predict future gas demand curves based on historical gas consumption data sequences. Compared to traditional fixed pressure or simple feedback regulation systems, this invention can generate optimal pressure setpoints in real time, avoiding gas instability caused by pressure fluctuations and ensuring the continuity and stability of the medical gas supply.

[0042] Of course, the above embodiments are not intended to limit this application, and this application is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this application should also fall within the protection scope of this application.

Claims

1. A medical clean air generator with pressure regulation, characterized in that: include: Cabinet (1), used for support; An air compressor (2) is located above the cabinet (1) and is used to extract gas. A condensing mechanism (3) is connected to the outlet of the air compressor (2) and is used to cool the gas. A fan (4) is located on one side of the condenser (3) and is used to cool the condenser (3). A gas-liquid separation mechanism (5) is located at the rear end of the condenser (3) and is used to separate liquid. A drying system (6) is located at the rear end of the gas-liquid separation mechanism (5) and is used to dry the gas. A gas storage tank (8) is located in the drying system (6) and is used to store the dried gas. A filtration system (7) is located at the rear end of the gas storage tank (8) and is used to filter the gas. The filtered gas is then delivered to the medical equipment. A control module (9) has a built-in pressure regulation optimization model and is used to predict future gas demand and generate the optimal pressure setting value. A high-speed servo pressure regulation module (10) is used to receive instructions from the control module (9) and adjust the output gas pressure in real time.

2. The medical clean air generator with pressure regulation according to claim 1, characterized in that: The gas-liquid separation mechanism (5) includes: a separation cylinder (51), which consists of two sets, with a conical liquid collecting section (511) at the bottom for collecting liquid; a tangential air inlet (52), located on the upper side wall of the separation cylinder (51), for tangentially introducing the mixed gas and liquid into the cylinder and forming a rotating airflow; a gas outlet (53), located at the top of the separation cylinder (51) and connected to the drying system (6); and a centrifugal separation assembly (54), consisting of... The cylindrical inner cavity of the tangential air inlet (52) and the separation cylinder (51) is used to achieve primary gas-liquid separation by centrifugal force. An air inlet switching valve (55) is located at the head end of the tangential air inlet (52), controlling only one separation cylinder (51) to be open, used to replace and discharge liquid respectively. A drain valve (56) is located at the bottom end of the conical liquid collection section (511), used to automatically or manually discharge accumulated liquid. An exhaust switching valve is used to connect the two separation cylinders (51, 52, and 51). The gas outlets of 51) converge into a total gas outlet (53); when an exhaust switching valve is used, its working state is synchronized with the intake switching valve (56); a detection sensor (58) is set at the upper end of the separation cylinder (1) and is used to detect the position of the liquid level; a wire mesh demister (59) is set inside the separation cylinder (51) and located on the airflow path between the tangential air inlet (52) and the gas outlet (53) to capture residual liquid droplets in the gas after passing through the centrifugal separation component (54); the detection sensor (58), the intake switching valve (55) and the drain valve (56) are electrically connected to the control module (9); the control module (9) is configured to: control the intake switching valve (55) to switch the conduction path according to the feedback signal of the detection sensor (58), and control the drain valve of the currently non-working cylinder to open for draining, thereby realizing the alternating operation and online draining of the two separation cylinders to ensure that the gas-liquid separation process is uninterrupted.

3. A medical clean air source generator with pressure regulation according to claim 1, characterized in that: The cabinet (1) is equipped with casters (11) around its lower perimeter. The filtration system (7) includes three-stage filtration: the first stage is a 0.3μm micron-level filter, the second stage is a 0.01μm terminal filter, and the third stage is an ultraviolet sterilization filter. The final output gas meets the ISO 8573-1 Class 0 cleanliness standard. The air compressor (2) is a swing piston oil-free air compressor, which is installed in the cabinet (1) through a spring shock absorber. The noise level is ≤60dB(A), and the exhaust volume is ≥100L / min. The bottom of the gas-liquid separation mechanism (5) is connected to an automatic drain solenoid valve. The fan (4) is an axial flow fan, which is located on the air inlet or air outlet side of the condensation mechanism (3). The drying system (6) is a double-tower adsorption drying device, which is connected to the switching valve group and is controlled by the control module (9) to alternate between adsorption and regeneration states.

4. A medical clean air source generator with pressure regulation according to claim 1, characterized in that: The control module (9) is deployed on the edge computing unit. Its built-in prediction model is the Transformer time series fusion model, which is used to predict the gas demand curve in the future period based on the historical gas consumption data sequence.

5. A medical clean air source generator with pressure regulation according to claim 1, characterized in that: The high-speed servo pressure regulation module (10) includes a precision pressure regulating valve, a first pressure sensor for monitoring the pressure of the gas storage tank, and a second pressure sensor for monitoring the pressure at the output end, forming a closed-loop pressure control system.

6. A medical clean air source generator with pressure regulation according to claim 1, characterized in that: It also includes a human-machine interaction module, which is connected to the control module (9) to display real-time pressure, gas quality information, predicted gas consumption curve and equipment status, and allows users to set target pressure range or receive maintenance alarm information.

7. A medical clean air source generator with pressure regulation according to claim 4, characterized in that: The Transformer time-series fusion model is an encoder-decoder architecture. The encoder is used to extract historical gas pressure, flow rate and gas usage period features, and the decoder is used to generate a gas usage prediction sequence for a specified time step in the future. The sequence is then converted into the optimal pressure setting value of the gas storage tank (8) by the dynamic programming algorithm of the control module (9).

8. A medical clean air source generator with pressure regulation according to claim 2, characterized in that: The wire mesh demister (59) is a multi-layer stainless steel wire mesh stacked structure with a wire diameter of 0.1 to 0.25 mm and a porosity of 97% to 99%. It is installed on the support ring below the gas outlet (53) to intercept and condense tiny droplets so that they fall into the conical liquid collection part (511).

9. A medical clean air source generator with pressure regulation according to claim 3, characterized in that: The motor of the swing piston oilless air compressor (2) is a permanent magnet synchronous servo motor, which is electrically connected to the control module (9). The control module (9) adjusts the speed of the servo motor according to the demand curve output by the prediction model to achieve variable load matching of compressed air volume.

10. A method for regulating a medical clean air source generator with pressure regulation, characterized in that: The process includes the following steps: Step S1: Collect historical gas usage data, equipment start / stop signals, and environmental parameters to construct a time-series feature vector. The time-series feature vector in Step S1 includes: gas tank pressure, output flow rate, gas temperature, compressor current, filter differential pressure, ambient temperature and humidity, time encoding, static encoding of department type, and equipment start / stop mask, totaling 9 dimensions. The sampling frequency is 100ms, the historical window length is 180 steps, and the prediction window length is 18 steps. The data is normalized using moving average and standard deviation. Step S2: Input the data from Step S1 into a Transformer time-series fusion model deployed on an edge AI chip. The construction of the Transformer time-series fusion model in Step S2 specifically includes the following sub-steps: Step S21: Static feature encoder to encode department type into a context vector; Step S22: Variable selection network to calculate the importance weights of each feature using Softmax. Step S23: Select key sensor signals; Step S24: LSTM encoder extracts historical time-series context features; Step S25: Multi-head self-attention mechanism, with 4 heads and a model dimension of 128, captures pressure fluctuation patterns; Step S26: Gated residual connection, fusing attention output with original features through learnable gating coefficients to prevent gradient vanishing; Step S27: Quantile output layer, predicting the P10, P50, and P90 quantile values ​​of gas demand in the next 18 steps; Step S3: The Transformer time-series fusion model outputs the gas demand curve for the next 1-3 minutes and the optimal pressure setpoint, and uses quantile prediction to provide uncertainty estimation; Step S4: The high-speed servo pressure regulation module receives model commands and combines them with piezoelectric microsensors to achieve closed-loop feedback control; Step S5: The reinforcement learning module performs online fine-tuning of the model based on pressure error, energy consumption, and equipment lifespan loss to optimize the pressure regulation strategy.

Images