Visual monitoring system for condensate water level of micro hyperbaric oxygen chamber

Abstract

The invention relates to the technical field of water level monitoring, in particular to a visual monitoring system for the condensate water level of a micro-hyperbaric oxygen chamber, which adopts the technical scheme that an internal image of a water receiving box is acquired at a preset frequency, and water mist and reflective interference are eliminated by adopting a dynamic illumination compensation and frequency domain filtering technology; a U-Net deep learning model is integrated, a water level line is segmented through multi-scale feature fusion, foam and impurity artifacts are recognized in combination with time sequence analysis, a water level value is dynamically corrected, graded alarm is triggered according to a water level threshold value, an air conditioner drainage system is linked, traditional image processing and the deep learning model are combined, and detection robustness under complex working conditions is improved. Real water levels and interferents are effectively distinguished through time sequence analysis and physical characteristic modeling, the real-time requirement under the high-pressure environment is met through the model compression and hardware acceleration technology, and absolute reliability of equipment operation is ensured through linkage of a grading early warning mechanism and a drainage system.

Description

Technical Field

[0001] The present invention relates to the technical field of water level monitoring, and in particular to a visual monitoring system for the condensation water level of a micro-hyperbaric oxygen chamber. Background Art

[0002] Micro-hyperbaric oxygen chambers are widely used in health care, medical assistance, sports recovery, and plateau acclimatization. Compared with medical-grade hyperbaric oxygen chambers, micro-hyperbaric oxygen chambers have lower pressure and higher safety, making them suitable for non-medical scenarios. They are also easy to operate and can be used in places such as homes, gyms, or beauty salons. Today's micro-hyperbaric oxygen chambers are equipped with air conditioning, which directly discharges the air conditioning condensate through pipes to the outside of the chamber. The gas pressure inside the chamber will also leak out along the condensate pipe to the outside of the chamber, causing pressure leakage, which will affect the use effect. Currently, the condensate pipe is connected to the water collection box in the chamber. Staff members manually open the inspection door regularly to check the water level in the water collection box and empty it. This not only wastes manpower, but also causes the condensate in the water collection box to overflow due to staff negligence.

[0003] In view of this, we propose a visual monitoring system for the condensation water level in a micro-hyperbaric oxygen chamber to solve the existing problem. Summary of the Invention

[0004] The object of the present invention is to provide a visual monitoring system for the condensation water level in a micro-hyperbaric oxygen chamber to solve the problems raised in the above background technology.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a visual monitoring system for the condensation water level in a micro-hyperbaric oxygen chamber, comprising: an image acquisition module, a preprocessing module, a water level detection module, an abnormality discrimination module, and a control module; wherein the image acquisition module includes a pressure-resistant and fog-proof wide-angle camera, which captures images of the interior of the water collection box at a preset frequency; the preprocessing module uses dynamic illumination compensation and frequency domain filtering technology to eliminate water mist and reflection interference; the water level detection module integrates a U-Net deep learning model, and segments the water level line through multi-scale feature fusion; the abnormality discrimination module combines time series analysis to identify foam and impurity artifacts, and dynamically corrects the water level value; the control module triggers a graded alarm according to the water level threshold and links the air conditioning drainage system.

[0006] Furthermore, the visual acquisition module obtains the cabin pressure value in real time through the embedded pressure sensor, dynamically adjusts the camera acquisition frequency, and triggers the emergency sampling mode based on the water level change rate.

[0007] Furthermore, the preprocessing module includes a dynamic illumination compensation subsystem and an intelligent frequency domain filtering subsystem; among them, the dynamic illumination compensation subsystem adopts a multi-scale Retinex decomposition and adaptive gamma correction fusion algorithm, separates the brightness component through the HSV color space for nonlinear mapping, and establishes an illumination distribution probability map, and dynamically adjusts the compensation intensity of each area based on Bayesian estimation; the intelligent frequency domain filtering subsystem implements wavelet packet transform and guided filtering joint denoising, performs frequency band adaptive threshold shrinkage at the 3-5 level decomposition scale, and constructs a reflective feature dictionary library, and directionally suppresses the highlight component in the frequency domain through sparse representation.

[0008] Furthermore, the water level detection module includes an enhanced U-Net++ network architecture and a multi-source input fusion mechanism; among them, the enhanced U-Net++ network architecture embeds a deformable convolution layer in the encoding stage to dynamically adapt to the deformation of the water level line, and introduces a cross-scale attention gate in the decoding stage to prioritize the fusion of high-frequency edge features; the multi-source input fusion mechanism inputs the preprocessed RGB image, near-infrared depth map and pressure sensor data into the three-channel network, and uses a gated recurrent unit in the bottleneck layer to realize temporal feature modeling.

[0009] Furthermore, the abnormality discrimination module includes a multimodal perception layer and a dynamic correction engine; among them, the multimodal perception layer integrates visible light video stream, capacitive liquid level pulse signal and laser scattering data to construct a three-dimensional feature space; the dynamic correction engine adopts a volumetric Kalman filter algorithm to jointly estimate the real water level and the interference artifact state, and embeds physical constraints, and uses the water surface continuity equation and the law of conservation of mass to correct the prediction deviation.

[0010] Furthermore, the control module includes a dynamic threshold engine and a hierarchical execution unit; wherein, the dynamic threshold engine predicts the water level change trend in the next 5 minutes based on the LSTM-Transformer hybrid model, dynamically adjusts the alarm threshold, and the baseline value fluctuates by ±15%. A pressure-temperature compensation function is established, and when the cabin pressure is greater than 2.5ATA, the trigger threshold is automatically increased by 3%-5%; the hierarchical execution unit includes a four-level response mechanism, which includes early warning level, primary alarm, emergency response, and disaster disposal; when the water volume in the water collection box reaches 75% of the capacity, the early warning level is triggered, the operating parameters are recorded, and the backup drainage pipeline is activated; when the water volume in the water collection box reaches 85% of the capacity, the primary alarm is triggered, the sound and light warning is activated, and the sampling rate is increased to 15Hz; when the water volume in the water collection box reaches 92% of the capacity, the emergency response is triggered, the main drain valve is opened, and 50%-100% duty cycle PWM control is performed; when the water volume in the water collection box reaches 95% of the capacity, disaster disposal is triggered, the air conditioning system is linked to forced ventilation, and the redundant drainage pump group is started.

[0011] Furthermore, the dynamic illumination compensation subsystem includes a polarization characteristic analysis unit and a deep learning enhancement unit; the polarization characteristic analysis unit uses the Stokes vector model to analyze the image polarization state, separate the specular reflection and diffuse reflection components, and correct the artifacts introduced by the non-uniform polarized light source through the Mueller matrix; the deep learning enhancement unit deploys a lightweight Vision Transformer network to predict the optimal gamma curve parameters, and the training data includes a synthetic data set of 12 typical hyperbaric oxygen chamber lighting conditions.

[0012] Furthermore, the intelligent frequency domain filtering subsystem includes water mist physical modeling and frequency band attention mechanism; among them, the water mist physical modeling constructs a depth-guided transmittance map based on the atmospheric scattering model, jointly optimizes the dehazing process with the dark channel prior, and introduces a non-local mean constraint term to retain the edge texture of the water level line; the frequency band attention mechanism dynamically selects key frequency bands through a learnable weight matrix to suppress 80-120Hz high-frequency oscillation noise, and embeds an edge enhancement filter in the LL band to improve the characteristic response of the water level line.

[0013] Furthermore, the multimodal perception layer includes a foam feature extraction unit and an impurity identification unit; among them, the foam feature extraction unit analyzes the foam motion trajectory based on the spatiotemporal graph convolutional network, extracts the velocity field curl characteristics, and constructs a foam life cycle model, and predicts the dissipation time in combination with the surface tension coefficient; the impurity identification unit deploys a lightweight YOLOv7-Tiny model to detect suspended matter in real time, adapts to the reflective properties of 8 types of medical materials through transfer learning, calculates the proportion of impurity projection area, and dynamically adjusts the water level compensation coefficient.

[0014] Furthermore, the dynamic correction engine includes a dual-domain verification mechanism and an adaptive learning module; among them, the dual-domain verification mechanism includes time domain analysis and frequency domain analysis. The time domain analysis predicts the water level trend in the next 3 seconds through the LSTM network, and triggers correction when the deviation from the measured value is greater than 2σ. The frequency domain analysis performs wavelet packet decomposition on the water level signal to suppress 10-25Hz high-frequency oscillation interference; the adaptive learning module builds a federated learning framework, aggregates abnormal data from multiple devices to update the discrimination model, and designs a forgetting factor mechanism to dynamically adjust the weight of historical data.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The present invention collects images of the interior of the water collection box at a preset frequency, adopts dynamic illumination compensation and frequency domain filtering technology to eliminate water mist and reflection interference, integrates the U-Net deep learning model and segments the water level line through multi-scale feature fusion, combines time series analysis to identify foam and impurity artifacts and dynamically corrects the water level value, triggers a graded alarm according to the water level threshold and links the air conditioning drainage system, combines traditional image processing with deep learning models to improve the detection robustness under complex working conditions, effectively distinguishes the real water level from interference through time series analysis and physical feature modeling, adopts model compression and hardware acceleration technology to meet the real-time requirements under high-pressure environment, ensures the absolute reliability of equipment operation through the graded early warning mechanism and links with the drainage system, visually monitors the condensation water level and controls the discharge of condensation water through water level monitoring, not only improves the accuracy of water level monitoring while saving labor costs, but also avoids the overflow of condensation water in the water collection box due to negligence of staff. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flowchart of a visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1

[0020] like Figure 1 As shown, a visual monitoring system for the condensation water level in a micro-hyperbaric oxygen chamber includes: an image acquisition module, a preprocessing module, a water level detection module, an anomaly discrimination module, and a control module. The image acquisition module includes a pressure-resistant and fog-proof wide-angle camera that captures images of the interior of the water collection box at a preset frequency. The preprocessing module uses dynamic illumination compensation and frequency domain filtering technology to eliminate water mist and reflection interference. The water level detection module integrates a U-Net deep learning model and segments the water level line through multi-scale feature fusion. The anomaly discrimination module combines time series analysis to identify foam and impurity artifacts and dynamically correct the water level value. The control module triggers a graded alarm based on the water level threshold and links the air conditioning drainage system.

[0021] The working principle of the visual monitoring system for the condensation water level in a micro-hyperbaric oxygen chamber based on Example 1 is as follows:

[0022] The visual acquisition module uses an embedded pressure sensor to obtain real-time cabin pressure values ​​and dynamically adjusts the camera acquisition frequency. When the pressure is ≤1.3ATA, a 3Hz base sampling rate is used. For every 0.2ATA increase in pressure, the sampling rate increases by 1Hz, up to a maximum of 10Hz. The visual acquisition module triggers an emergency sampling mode based on the rate of change of the water level. If a water level rise rate of >2mm / s is detected for five consecutive frames, a burst of 20Hz high-frequency acquisition is initiated, and the base frequency is automatically restored after the water level stabilizes. The camera uses dual-channel heterogeneous imaging, alternating between the visible light channel and the 940nm near-infrared channel, and compensates for water mist interference by switching between bands. A differential exposure strategy is set, with the visible light channel fixed at 1 / 60s and the near-infrared channel adaptively adjusted (1 / 100s to 1 / 30s).

[0023] The visual acquisition module also includes a multimodal trigger mechanism. Using an infrared temperature sensor, it monitors the risk of condensation on the water collection box surface. When the temperature difference exceeds 5°C, it activates the anti-fog heating ring and simultaneously increases the sampling rate to 8Hz. A microphone array captures the sound pattern of the drain valve's movement, forcing an image capture within 2 seconds before and after the valve opens and closes. The visual acquisition module also includes an intelligent sleep strategy. When the pressure sensor and water level detection module jointly determine that the cabin environment has been stable for more than 30 minutes, it initiates motion wake-up mode, activating the main camera only when the inter-frame difference method detects a water level movement of 1 pixel or more. The low-power Time of Flight sensor is maintained for millimeter-level displacement monitoring.

[0024] The preprocessing module includes a dynamic illumination compensation subsystem and an intelligent frequency domain filtering subsystem. The dynamic illumination compensation subsystem adopts a multi-scale Retinex decomposition and adaptive gamma correction fusion algorithm to separate the brightness component in the HSV color space for nonlinear mapping, establish an illumination distribution probability map, and dynamically adjust the compensation intensity of each area based on Bayesian estimation. The intelligent frequency domain filtering subsystem implements wavelet packet transform and guided filtering for joint denoising, performs frequency band adaptive threshold shrinkage at 3-5 decomposition scales, and constructs a reflective feature dictionary library, using sparse representation to directionally suppress highlight components in the frequency domain.

[0025] The multi-scale Retinex decomposition process includes image decomposition, reflection component extraction, and adaptive gamma correction. In image decomposition, the input image is converted to the HSV color space, the brightness component (V channel) is extracted, and the low, medium, and high scale illumination components are generated through the Gaussian filter kernel (σ=15 / 80 / 250). In reflection component extraction, Among them, F k is a Gaussian filter with weight w k Distributed by 0.3 / 0.4 / 0.3. In adaptive gamma correction, calculate local contrast σ is the local standard deviation, μ is the mean, and the gamma value is dynamically adjusted: γ = 1 + 0.5 (1 - tanh (10C)).

[0026] The implementation steps of the combined denoising of wavelet packet transform and guided filtering include wavelet packet decomposition, band-selective filtering, and sparse suppression of reflective components. In the wavelet packet decomposition, the sym4 wavelet basis is selected for three-layer decomposition to obtain 8 frequency band sub-graphs, and the threshold is calculated for each sub-band: σ j is the subband noise variance, and N is the number of pixels. In the band-selective filtering, hard threshold shrinkage is applied to the subband dominated by 80-120Hz high-frequency noise (usually LH3 / HL3), and guided filtering is embedded in the LL low-frequency subband to preserve the edge: p is the input image, q is the filtered output, W ij is the edge perception weight. In the sparse suppression of the reflective component, an overcomplete dictionary D = [D edge ,D specular ], solve the sparse coding problem: min α ||x-Dα|| 2 +λ||α||1, forced reflective atomic coefficient α specular Reconstruct the image after zeroing.

[0027] The dynamic illumination compensation subsystem includes a polarization characteristic analysis unit and a deep learning enhancement unit. The polarization characteristic analysis unit uses the Stokes vector model to analyze the image polarization state, separate the specular reflection and diffuse reflection components, and correct the artifacts introduced by the non-uniform polarized light source through the Mueller matrix. The deep learning enhancement unit deploys a lightweight VisionTransformer network to predict the optimal gamma curve parameters, and the training data includes a synthetic dataset of 12 typical hyperbaric oxygen chamber lighting conditions.

[0028] In Stokes vector analysis, the polarization images at 0°, 45°, and 90° are obtained by rotating the polarizer, and the Stokes vector is calculated: Separate the specular component:

[0029] The intelligent frequency domain filtering subsystem includes water mist physical modeling and frequency band attention mechanism; among them, the water mist physical modeling constructs a depth-guided transmittance map based on the atmospheric scattering model, jointly optimizes the dehazing process with the dark channel prior, and introduces a non-local mean constraint term to retain the edge texture of the water level line; the frequency band attention mechanism dynamically selects key frequency bands through a learnable weight matrix to suppress 80-120Hz high-frequency oscillation noise, and embeds an edge enhancement filter in the LL band to improve the characteristic response of the water level line.

[0030] In the physical model optimization of depth-guided dehazing, the atmospheric scattering model is: I(x) = J(x)t(x) + A(1-t(x)), where J is the clear image, t is the transmittance, and A is the atmospheric light. For the dark channel prior, the depth sensor data is added to constrain the transmittance: d(x) is the depth map, ω = 0.95; non-local mean refinement is: t refined (x)=Σ y∈Ω w(x,y)t(y), the weight w(x,y) is determined by the color and depth similarity.

[0031] The preprocessing module also includes a real-time quality assessment module and a hardware acceleration architecture. The SSIM indicator of the images before and after processing is calculated. When the evaluation value is less than 0.85, parameter self-correction is triggered, and a GAN adversarial network is constructed to generate extreme interference samples for online model fine-tuning. The hardware acceleration architecture deploys a pipelined preprocessing pipeline on the FPGA, completing single-frame processing with a delay of less than 8ms, and designs a dedicated DSP core to implement parallel operations for wavelet transforms.

[0032] In the FPGA pipeline processing architecture, polarization separation is performed after the input image, followed by wavelet packet decomposition, guided filtering after threshold shrinkage, sparse coding and image reconstruction, and finally the image is output.

[0033] The water level detection module includes an enhanced U-Net++ network architecture and a multi-source input fusion mechanism. The enhanced U-Net++ network architecture embeds a deformable convolution layer in the encoding stage to dynamically adapt to the deformation of the water level line, and introduces a cross-scale attention gate in the decoding stage to prioritize the fusion of high-frequency edge features. The multi-source input fusion mechanism inputs the preprocessed RGB image, near-infrared depth map and pressure sensor data into the three-channel network, and uses a gated recurrent unit in the bottleneck layer to realize temporal feature modeling.

[0034] Spatial-channel attention is inserted into the attention module at each layer of the decoder, and the channel attention is: M C (F) = σ(MLP(AvgPool(F))+MLP(MaxPool(F))), spatial attention is: M S (F)=σ(f 7×7 ([AvgPool(F); MaxPool(F)])). Construct a feature pyramid to achieve cross-layer information interaction: Weight w i Dynamically generated through 1×1 convolution + Softmax.

[0035] The enhanced U-Net++ network includes a dynamic weight allocation module and a virtual-real fusion training strategy. The dynamic weight allocation module adaptively adjusts the fusion weights of features at different scales by calculating the curvature radius of the water level in real time. A geometrically constrained loss function is used to enforce consistency in horizontal feature responses. The virtual-real fusion training strategy utilizes StyleGAN to generate a synthetic water ripple dataset with physical parameter labels (covering fluctuation amplitudes of 0.5-5mm). Domain randomization is then implemented to simulate the surface tension effects of fluids under a high pressure of 2.8 ATA.

[0036] In the loss function design, the composite loss function is: L = 0.7L Dice +0.2L Edge +0.1L Geo , Dice Loss improves segmentation accuracy: Edge loss strengthens waterline continuity: Geometric constraints maintain horizontal consistency:

[0037] The water level detection module also includes a post-processing optimization engine and an embedded deployment solution. The post-processing optimization engine combines the probability map output by the network with a Radon transform to extract sub-pixel water level coordinates. A Kalman filter model is then built, integrating 10 frames of historical data to suppress transient interference fluctuations. The embedded deployment solution utilizes channel pruning and 8-bit fixed-point quantization to compress the model to less than 2MB. TensorRT acceleration is used to achieve 30fps real-time inference on a Jetson Nano.

[0038] The state equation of the adaptive Kalman filter is: The observation equation is: k =[1 0]x k +v k , the noise covariance is dynamically adjusted to: Q = 0.1 + 0.05·|v|.

[0039] The abnormality discrimination module includes a multimodal perception layer and a dynamic correction engine; the multimodal perception layer integrates visible light video stream, capacitive liquid level pulse signal and laser scattering data to construct a three-dimensional feature space; the dynamic correction engine adopts a volumetric Kalman filter algorithm to jointly estimate the real water level and the interference artifact state, and embeds physical constraints, and uses the water surface continuity equation and the law of conservation of mass to correct the prediction deviation.

[0040] The multimodal perception layer includes a foam feature extraction unit and an impurity identification unit. The foam feature extraction unit analyzes the foam motion trajectory based on the spatiotemporal graph convolutional network, extracts the velocity field curl characteristics, and constructs a foam life cycle model, and predicts the dissipation time in combination with the surface tension coefficient. The impurity identification unit deploys a lightweight YOLOv7-Tiny model to detect suspended matter in real time, adapts to the reflective properties of 8 types of medical materials through transfer learning, calculates the proportion of impurity projection area, and dynamically adjusts the water level compensation coefficient.

[0041] The foam feature modeling adopts a spatiotemporal graph convolutional network. The input is a video clip of 10 consecutive frames with a 256×256 ROI area. Each foam area is extracted as a graph node through OpenCV Blob detection. The edge weight is the Euclidean distance and velocity correlation based on the foam motion trajectory. The output feature is the curl. Lifecycle prediction value. In the impurity detection process, the transfer learning strategy includes pre-training datasets and data enhancement.

[0042] The dynamic correction engine includes a dual-domain verification mechanism and an adaptive learning module. The dual-domain verification mechanism includes time domain analysis and frequency domain analysis. The time domain analysis predicts the water level trend in the next 3 seconds through the LSTM network, and triggers correction when the deviation from the measured value is greater than 2σ. The frequency domain analysis performs wavelet packet decomposition on the water level signal to suppress 10-25Hz high-frequency oscillation interference. The adaptive learning module builds a federated learning framework, aggregates abnormal data from multiple devices to update the discrimination model, and designs a forgetting factor mechanism to dynamically adjust the weight of historical data.

[0043] In wavelet packet energy analysis, the water level signal is decomposed into 8 sub-bands, and the energy proportion of the characteristic frequency band (20-50Hz) is calculated: When E abnormal When it is greater than 0.4, it is determined that there is high-frequency interference. The parameter aggregation of the federated learning framework is: A global update is performed every 24 hours, and differential privacy protection adds Gaussian noise of σ = 0.01.

[0044] The control module includes a dynamic threshold engine and a hierarchical execution unit. The dynamic threshold engine predicts the water level change trend in the next 5 minutes based on the LSTM-Transformer hybrid model, dynamically adjusts the alarm threshold, and the baseline value fluctuates by ±15%. A pressure-temperature compensation function is established, and the trigger threshold is automatically increased by 3%-5% when the cabin pressure is greater than 2.5ATA. The hierarchical execution unit includes a four-level response mechanism, which includes early warning level, primary alarm, emergency response, and disaster disposal. When the water level in the water collection box reaches 75% of its capacity, the early warning level is triggered, the operating parameters are recorded, and the backup drainage pipeline is activated. When the water level in the water collection box reaches 85% of its capacity, the primary alarm is triggered, the sound and light warning is activated, and the sampling rate is increased to 15Hz. When the water level in the water collection box reaches 92% of its capacity, the emergency response is triggered, the main drain valve is opened, and 50%-100% duty cycle PWM control is performed. When the water level in the water collection box reaches 95% of its capacity, disaster disposal is triggered, the air conditioning system is linked to forced ventilation, and the redundant drainage pump group is started.

[0045] The input features of the LSTM-Transformer hybrid prediction model include the historical water level series (60-second window), the real-time values ​​of cabin pressure, temperature and humidity, and the status of the air conditioning drain valve (opening, flow rate); the dynamic threshold is calculated as T alert =μ base +0.15σ pred +0.03P cabin , μ base is the reference threshold, σ pred is the predicted volatility standard deviation, P cabin is the current cabin pressure. In the pressure compensation mechanism, when the pressure is greater than 2.5ATA, T adjusted =T alert ×(1+0.05(P cabin -2.5)); Temperature compensation coefficient α T =1+0.002(T water -25).

[0046] The dynamic threshold engine includes a multi-source sensing interface and an energy optimization strategy. The multi-source sensing interface integrates data from three channels: laser ranging, capacitive sensing, and visual detection to construct a Bayesian confidence model. A sensor self-check is initiated when the data divergence exceeds 0.3. The energy optimization strategy uses a fuzzy PID control algorithm to dynamically adjust the drain valve opening, improving energy efficiency by 40%. During the drain phase, the air conditioning dehumidification power is simultaneously adjusted to maintain cabin pressure fluctuations below 0.02 ATA.

[0047] For the input variable of the fuzzy PID control algorithm, the water level deviation e(t) = h target -h current , deviation change rate The output is the drain valve PWM duty cycle with a resolution of 0.1%.

[0048] The hierarchical execution unit integrates a fault-tolerant control module and a human-machine collaboration mechanism. In the fault-tolerant control module, a dual CAN bus architecture ensures redundant transmission of control commands, with response latency less than 50ms. A hardware watchdog circuit triggers safety mode if no heartbeat signal is received within 300ms. The human-machine collaboration mechanism provides a three-dimensional water level situation map and a list of recommended actions through the HMI interface. Voice commands can be used to interrupt the automatic control process, prioritizing manual operations.

[0049] The above specific embodiments are only several preferred embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber, characterized in that: include: Image acquisition module, preprocessing module, water level detection module, abnormality discrimination module, control module; wherein the image acquisition module includes a pressure-resistant and fog-proof wide-angle camera, which captures images inside the water collection box at a preset frequency; The pre-processing module uses dynamic illumination compensation and frequency domain filtering technology to eliminate water mist and reflection interference; the water level detection module integrates the U-Net deep learning model and segments the water level line through multi-scale feature fusion; the anomaly discrimination module combines time series analysis to identify foam and impurity artifacts and dynamically correct the water level value; the control module triggers graded alarms based on the water level threshold and links the air conditioning drainage system.

2. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 1, characterized in that: The visual acquisition module obtains the cabin pressure value in real time through the embedded pressure sensor, dynamically adjusts the camera acquisition frequency, and triggers the emergency sampling mode based on the water level change rate.

3. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 1, characterized in that: The preprocessing module includes a dynamic illumination compensation subsystem and an intelligent frequency domain filtering subsystem. The dynamic illumination compensation subsystem adopts a multi-scale Retinex decomposition and adaptive gamma correction fusion algorithm to separate the brightness component in the HSV color space for nonlinear mapping, establish an illumination distribution probability map, and dynamically adjust the compensation intensity of each area based on Bayesian estimation. The intelligent frequency domain filtering subsystem implements wavelet packet transform and guided filtering for joint denoising, performs frequency band adaptive threshold shrinkage at 3-5 decomposition scales, and constructs a reflective feature dictionary library, using sparse representation to directionally suppress highlight components in the frequency domain.

4. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 1, characterized in that: The water level detection module includes an enhanced U-Net++ network architecture and a multi-source input fusion mechanism. The enhanced U-Net++ network architecture embeds a deformable convolution layer in the encoding stage to dynamically adapt to the deformation of the water level line, and introduces a cross-scale attention gate in the decoding stage to prioritize the fusion of high-frequency edge features. The multi-source input fusion mechanism inputs the preprocessed RGB image, near-infrared depth map and pressure sensor data into the three-channel network, and uses a gated recurrent unit in the bottleneck layer to realize temporal feature modeling.

5. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 1, characterized in that: The abnormality discrimination module includes a multimodal perception layer and a dynamic correction engine; the multimodal perception layer integrates visible light video stream, capacitive liquid level pulse signal and laser scattering data to construct a three-dimensional feature space; the dynamic correction engine adopts a volumetric Kalman filter algorithm to jointly estimate the real water level and the interference artifact state, and embeds physical constraints, and uses the water surface continuity equation and the law of conservation of mass to correct the prediction deviation.

6. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 1, characterized in that: The control module includes a dynamic threshold engine and a hierarchical execution unit. The dynamic threshold engine predicts the water level change trend in the next 5 minutes based on the LSTM-Transformer hybrid model, dynamically adjusts the alarm threshold, and the baseline value fluctuates by ±15%. A pressure-temperature compensation function is established, and the trigger threshold is automatically increased by 3%-5% when the cabin pressure is greater than 2.5ATA. The hierarchical execution unit includes a four-level response mechanism, which includes early warning level, primary alarm, emergency response, and disaster disposal. When the water level in the water collection box reaches 75% of its capacity, the early warning level is triggered, the operating parameters are recorded, and the backup drainage pipeline is activated. When the water level in the water collection box reaches 85% of its capacity, the primary alarm is triggered, the sound and light warning is activated, and the sampling rate is increased to 15Hz. When the water level in the water collection box reaches 92% of its capacity, the emergency response is triggered, the main drain valve is opened, and 50%-100% duty cycle PWM control is performed. When the water level in the water collection box reaches 95% of its capacity, disaster disposal is triggered, the air conditioning system is linked to forced ventilation, and the redundant drainage pump group is started.

7. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 3, characterized in that: The dynamic illumination compensation subsystem includes a polarization characteristic analysis unit and a deep learning enhancement unit. The polarization characteristic analysis unit uses the Stokes vector model to analyze the image polarization state, separate the specular reflection and diffuse reflection components, and correct the artifacts introduced by the non-uniform polarized light source through the Mueller matrix. The deep learning enhancement unit deploys a lightweight Vision Transformer network to predict the optimal gamma curve parameters, and the training data includes a synthetic dataset of 12 typical hyperbaric oxygen chamber lighting conditions.

8. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 3, characterized in that: The intelligent frequency domain filtering subsystem includes water mist physical modeling and frequency band attention mechanism; among them, the water mist physical modeling constructs a depth-guided transmittance map based on the atmospheric scattering model, jointly optimizes the dehazing process with the dark channel prior, and introduces a non-local mean constraint term to retain the edge texture of the water level line; the frequency band attention mechanism dynamically selects key frequency bands through a learnable weight matrix to suppress 80-120Hz high-frequency oscillation noise, and embeds an edge enhancement filter in the LL band to improve the characteristic response of the water level line.

9. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 5, characterized in that: The multimodal perception layer includes a foam feature extraction unit and an impurity identification unit. The foam feature extraction unit analyzes the foam motion trajectory based on the spatiotemporal graph convolutional network, extracts the velocity field curl characteristics, and constructs a foam life cycle model, and predicts the dissipation time in combination with the surface tension coefficient. The impurity identification unit deploys a lightweight YOLOv7-Tiny model to detect suspended matter in real time, adapts to the reflective properties of 8 types of medical materials through transfer learning, calculates the proportion of impurity projection area, and dynamically adjusts the water level compensation coefficient.

10. The visual monitoring system for condensation water level in a micro-hyperbaric oxygen chamber according to claim 5, characterized in that: The dynamic correction engine includes a dual-domain verification mechanism and an adaptive learning module. The dual-domain verification mechanism includes time domain analysis and frequency domain analysis. The time domain analysis predicts the water level trend in the next 3 seconds through the LSTM network, and triggers correction when the deviation from the measured value is greater than 2σ. The frequency domain analysis performs wavelet packet decomposition on the water level signal to suppress 10-25Hz high-frequency oscillation interference. The adaptive learning module builds a federated learning framework, aggregates abnormal data from multiple devices to update the discrimination model, and designs a forgetting factor mechanism to dynamically adjust the weight of historical data.

Images