AI visual stoma leakage early warning system

Abstract

The invention relates to the technical field of stoma nursing devices, and discloses an AI visual stoma leakage early warning system, which comprises a plurality of passive piezoelectric film sensors embedded in a stoma chassis and used for capturing acoustic signals generated by adhesion interface structure instability; the edge calculation unit periodically collects an electric signal sequence and calculates an information entropy value, and through comparison with a self-adaptive resting entropy baseline, the structural instability risk is judged and early warning is generated when the entropy value continuously and abnormally rises. The risk orientation is accurately identified in combination with a multi-sensor cooperative positioning technology, body movement interference is avoided by using a dynamic compensation mechanism, and the early warning accuracy and clinical practicability are improved.

Description

Technical Field

[0001] The present invention relates to an AI visual stoma leakage early warning system, belonging to the technical field of stoma care devices. Background Art

[0002] The current mainstream solution in this field uses humidity or pressure sensors to capture contact signals of leaks. However, this design has fundamental limitations: when the sensor is triggered, the leakage event has already occurred, resulting in a serious lag in early warning. This is especially true in daily patient activities, where the mechanical noise generated by violent body movements is difficult to distinguish from real leakage signals. This forces the system to increase the alarm threshold to avoid false alarms, but reduces the monitoring sensitivity for high-risk users.

[0003] Although the industry has tried to introduce technologies such as optical imaging to improve accuracy, it still faces two deep-seated contradictions: first, existing solutions can only provide binary judgments on the occurrence of leakage and cannot locate risk sites to guide targeted intervention; second, chemical sensors are easily interfered by sweat, and physical sensors produce monitoring deviations due to differences in individual skin characteristics.

[0004] Specifically, there are three common technical bottlenecks in this field: 1. The monitoring logic adheres to the physical contact triggering method, missing the time window before structural instability; 2. It is unable to distinguish between body motion noise and real risk signals, falling into the inherent contradiction between sensitivity and false alarm rate; 3. The lack of the ability to predict the direction and development trend of risks leads to passive remediation. These defects make it difficult for existing systems to establish a reliable early warning mechanism in dynamic nursing scenarios, and a new monitoring method that can perceive structural changes in the adhesion interface in advance is needed. Therefore, how to achieve early warning and precise positioning of the risk of instability of the stoma base adhesion interface, and effectively avoid the false alarm problem caused by body motion interference, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] The present invention provides an AI visual stoma leakage warning system, the main purpose of which is to solve the problems of early warning and precise positioning of the risk of instability of the stoma chassis adhesion interface, while eliminating the problem of false alarms caused by body motion interference.

[0006] To achieve the above objectives, the present invention provides an AI visual stoma leakage early warning system, comprising:

[0007] at least three passive piezoelectric film sensors, which are embedded in the stoma baseplate and configured to convert weak acoustic signals generated at the stoma baseplate and skin adhesion interface during structural instability into electrical signals;

[0008] an edge computing unit, the edge computing unit being electrically connected to the piezoelectric film sensor and configured to: periodically collect a sequence of electrical signals output by the plurality of piezoelectric film sensors;

[0009] The information entropy value of the electrical signal sequence is independently calculated; the calculated real-time information entropy value is compared with the respectively established resting entropy baseline; based on the comparison result, when the real-time information entropy value of at least one piezoelectric film sensor is continuously higher than the preset difference threshold of its corresponding resting entropy baseline, the risk of structural instability of the adhesion interface is determined, and a risk warning signal is generated, which is used to indicate the occurrence of structural instability risk.

[0010] Preferably, the edge computing unit is further configured to: collaboratively analyze the real-time information entropy values ​​calculated by multiple piezoelectric film sensors, by identifying the situation where one or more information entropy values ​​are first and continuously higher than the information entropy values ​​of the remaining sensors, and combining the established physical position of the piezoelectric film sensor with abnormal information entropy value on the stoma base plate, accurately locate the spatial orientation of the risk of structural instability of the adhesion interface, thereby including this orientation information in the risk warning signal to guide users to take targeted intervention measures.

[0011] Preferably, the edge computing unit is further configured to: divide the structural instability risk into multiple deterministic risk levels according to the magnitude and duration of the real-time information entropy value exceeding the resting entropy baseline.

[0012] Preferably, the edge computing unit is further configured to: divide the structural instability risk into multiple deterministic risk levels according to the magnitude and duration of the real-time information entropy value exceeding the resting entropy baseline.

[0013] Preferably, the resting entropy baseline is adaptively established by statistically averaging the information entropy values ​​within a fixed period of time after the system is initially worn.

[0014] Preferably, the periodic collection frequency of the edge calculation unit is less than 1 time per second, and the duration of each collection does not exceed 1 second.

[0015] Preferably, the device further comprises: a three-axis acceleration sensor configured to monitor the wearer's body motion intensity in real time; and a dynamic compensation module electrically connected to the three-axis acceleration sensor and the edge computing unit, and configured to: when the instantaneous acceleration vector and the instantaneous acceleration vector detected by the three-axis acceleration sensor are equal, the dynamic compensation module is automatically compensated. Exceeding a given physical threshold , that is, satisfy When, among them, represents the instantaneous acceleration vector sum, It represents a given physical threshold, and determines that the current moment and the specified time window thereafter are the macro-impact pollution period. During the macro-impact pollution period, the edge computing unit is suspended from determining the validity of the information entropy of the electrical signal sequence or performing noise compensation processing on the electrical signal sequence based on the data of the three-axis acceleration sensor to avoid false alarms caused by macro-body motion interference.

[0016] Preferably, the edge computing unit is also configured to periodically switch at least one pair of piezoelectric film sensors to an active detection mode; in the active detection mode, one of the piezoelectric film sensors is driven to transmit a multi-frequency alternating electrical signal to the skin, and the bioelectrical impedance spectrum response of the skin is measured through another piezoelectric film sensor; the skin barrier function status index is inverted according to the bioelectrical impedance spectrum response; and the information entropy alarm threshold used to determine the risk of structural instability is dynamically adjusted according to the status index, so as to achieve early warning of leakage risks caused by skin problems themselves.

[0017] Preferably, it also includes a decision arbitration module, which includes a movable single-point infrared temperature sensor; the decision arbitration module is configured to: when the edge computing unit generates a risk warning signal containing orientation information, drive the movable single-point infrared temperature sensor to scan and measure the temperature of the risk warning point and its adjacent area; only when the scanning temperature measurement results show that the temperature presents a continuous gradient change and is consistent with the established heat conduction model, the decision arbitration module will finally output a risk warning signal to the user to filter out false alarms caused by sensor artifacts or small-scale instantaneous interference.

[0018] Preferably, the three-axis acceleration sensor is also configured to trigger a fall alarm signal when the instantaneous acceleration it monitors continuously exceeds a fixed fall threshold, and the fall alarm signal is sent to the user or a preset caregiver terminal through the edge computing unit.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. The system uses passive piezoelectric film sensors to capture weak acoustic signals from the adhesion interface and analyzes changes in the information entropy of the signal sequence through an edge computing unit. Compared with the traditional hysteresis response that relies on leakage detection, this mechanism identifies microscopic peeling precursors based on non-periodic mutations in the acoustic signal pattern. By dynamically comparing the information entropy value with the adaptive resting baseline, an early warning can be triggered at the early stage of physical structural instability, advancing the risk response window from the occurrence of leakage to before structural failure, thereby reducing the risk of clinical complications. By independently calculating the real-time information entropy of multiple piezoelectric sensors and comparing their spatial distribution characteristics, the system can determine the risk direction based on the physical location of the first sensor with an abnormally high entropy value. Combined with the amplitude and duration of the entropy value exceeding the baseline, the system can further divide the deterministic risk level. This composite mechanism of spatial positioning + level quantification enables users to take targeted low-cost interventions such as local reinforcement, avoiding the waste of resources caused by traditional global alarms.

[0021] 2. A three-axis acceleration sensor is introduced to monitor the intensity of macroscopic body motion. When the instantaneous acceleration exceeds the threshold, the dynamic compensation module automatically marks it as a macroscopic impact contamination period. During this window period, the system distinguishes between external mechanical impact and internal structural instability signals by suspending the acoustic entropy validity judgment or performing noise compensation. This mechanism enables the core monitoring logic to maintain specificity in intense body motion scenarios, solving the problem of false alarms in high-sensitivity systems in dynamic environments; and by periodically switching the piezoelectric sensor to active detection mode, emitting multi-frequency micro-disturbance electrical signals to the skin and measuring the bioelectrical impedance spectrum response, the skin barrier function index is inverted based on the impedance spectrum characteristics, and the information entropy alarm threshold is dynamically adjusted. This hardware multiplexing mechanism enables the system to perceive the skin's own health status, such as stratum corneum damage, and initiates a more sensitive early warning strategy for users with decreased barrier function, forming a dual-dimensional physiological and physical risk protection.

[0022] 3. After the acoustic system outputs an azimuth risk warning, the movable single-point infrared temperature sensor performs gradient scanning and temperature measurement on the target area. Only when the temperature change conforms to the preset heat conduction continuity model will the decision arbitration module finally confirm the alarm. This mechanism uses the physical law that real leakage is inevitably accompanied by continuous heat diffusion to effectively filter out isolated hot spot signals caused by sensor artifacts or transient interference, thereby improving the clinical credibility of the system's output decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the sensor layer and information processing of the AI ​​visual stoma leakage early warning system of the present invention;

[0024] Figure 2 Schematic diagram of the mechanism for dynamically adjusting the information entropy alarm threshold based on the skin barrier function index of the present invention;

[0025] Figure 3 This is a process interaction timing diagram of the resting entropy baseline adaptive learning phase after the system of the present invention is started.

[0026] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0027] In order to make the purpose, concept and effectiveness of this technical solution clearer, the specific implementation method of an AI visual stoma leakage warning system will be elaborated in detail below. The description here is intended to provide a clear example for technical personnel to implement, rather than any form of limitation on the protection boundaries of this technical solution.

[0028] The present application discloses an AI visual stoma leakage warning system, whose overall architecture is deeply coupled and coordinated through a passive acoustic signal acquisition array, an edge intelligent computing unit, a macroscopic body motion perception and dynamic compensation module, and a multimodal decision arbitration module. It aims to transform the traditional hysteresis detection method based on the contact of leakage into a predictive warning mechanism based on the precursors of failure of the structural integrity of the adhesion interface. In view of the fundamental lag in the monitoring time point of the existing technology, that is, the alarm can only be triggered after the excrement contacts the sensor, this solution moves the monitored physical quantity forward to the weak acoustic energy released during the structural instability process.

[0029] Specifically, at least three passive piezoelectric film sensors are embedded in the adhesion area of ​​the stoma base plate in a preset geometric configuration. When the adhesion interface between the base plate and the skin undergoes micro-peeling due to body fluid infiltration or stress concentration, the tiny events of material fiber breakage and interface debonding will release energy in the form of high-frequency acoustic signals, and these sensors are specially configured to efficiently capture such acoustic signals and convert them into electrical signal sequences. This design physically pushes the warning time point from the result of the leakage event to the cause of structural failure, thereby gaining a critical time window for clinical preventive intervention. In order to accurately identify the deterministic indications of structural instability from the collected complex time series signals, the edge computing unit is configured to execute a set of signal pattern recognition procedures based on information entropy. The unit discretely samples the output of each piezoelectric film sensor at a periodic frequency of less than 1 time per second, and the acquisition window length for each time does not exceed 1 second, thereby reducing the system while ensuring monitoring continuity. System power consumption; for each segment of the collected electrical signal sequence, the edge computing unit will independently calculate its information entropy value. This entropy value, as a quantitative representation of signal uncertainty, can sensitively reflect the inherent pattern changes of the signal. The acoustic background noise generated by a healthy and stable adhesion interface has low information entropy due to the relative stability of its pattern. Conversely, once micro-peeling occurs, the non-periodic mutation of the acoustic signal will inevitably lead to an increase in the information entropy value; by continuously and dynamically comparing the real-time calculated information entropy value with an adaptively established resting entropy baseline, when the real-time information entropy value of at least one sensor is continuously higher than the dynamic alarm threshold composed of its corresponding resting entropy baseline and a preset difference threshold, the system determines that the adhesion interface has entered a risk state of structural instability and generates a warning signal. This enables the system to identify qualitative changes in signal patterns hidden under the noise background that cannot be detected by traditional amplitude detection methods, thereby achieving highly sensitive early risk judgment.

[0030] Taking into account that slight differences in individual skin characteristics, environmental noise and even chassis tailoring techniques will lead to different initial background noise patterns, the establishment of the resting entropy baseline follows an adaptive online calibration process rather than a fixed preset value. Within the first fixed period of time after the system is initially worn, for example, thirty minutes, the system will automatically enter the learning and calibration stage. During this period, the edge computing unit continuously collects and calculates the information entropy value sequence, and performs statistical processing on the sequence to obtain its mean, which is then established as the resting entropy baseline exclusive to the sensor during the current wearing cycle; this adaptive calibration mechanism ensures that subsequent abnormal judgments are based on the user's own stable physiological and physical state, rather than a universal but potentially inaccurate general model, thereby improving the personalized accuracy and reliability of the early warning; further, in order to avoid the technical limitations of traditional solutions that can only provide binary judgments but cannot guide targeted interventions, this solution uses the spatial array information of multiple sensors to achieve precise spatial positioning of risk sources. The edge computing unit uses the spatial array information of multiple sensors to determine the location of the risk source. When a structural instability risk occurs, the real-time information entropy value stream from all piezoelectric film sensors will be collaboratively analyzed. By identifying which sensor or sensors' information entropy values ​​are first and continuously higher than those of the remaining sensors, the system can infer the starting spatial orientation of the structural instability of the adhesion interface based on the known physical coordinates of the specific sensor on the stoma base. Furthermore, the system is also configured to dynamically divide the structural instability risk into multiple deterministic risk levels based on the magnitude of the real-time information entropy value exceeding its resting entropy baseline and the duration of the abnormal state. For example, a lower excess magnitude and a shorter duration may correspond to a first-level concern level, while a higher excess magnitude and a longer duration will trigger a third-level high-risk level warning. This composite warning information combining spatial positioning and risk grading enables users or caregivers to take low-cost targeted intervention measures such as local pressure reinforcement, effectively avoiding unnecessary overall device replacement and improving the utilization efficiency of nursing resources.

[0031] Considering that the inevitable macroscopic body movements during the wearer's daily activities will produce strong mechanical shocks, the acoustic signals stimulated by them can easily be confused with the real microscopic peeling signals and induce false alarms. Therefore, this solution introduces a dynamic compensation mechanism to improve the specificity of monitoring. The core of this mechanism is a three-axis acceleration sensor electrically connected to the edge computing unit. The sensor is configured to monitor the wearer's body movement intensity in real time, and a dynamic compensation module is used to instantly solve the data. A predetermined physical threshold is set inside the module. Its value is carefully determined by conducting large-scale offline calibration experiments on the acceleration characteristics of typical actions such as fast walking, bending, and turning over. When the system is running, when the instantaneous acceleration vector and the acceleration vector detected by the three-axis acceleration sensor are Meet the conditions When the current moment and a specified time window thereafter are the macro-impact contamination period, the dynamic compensation module determines that the current moment and a specified time window thereafter are the macro-impact contamination period. During this contamination period, the edge computing unit's judgment on the validity of the information entropy of the piezoelectric film sensor's electrical signal sequence will be temporarily blocked, or the electrical signal sequence will be adaptively noise-cancelled based on the data of the three-axis acceleration sensor. At the same time, the three-axis acceleration sensor is also configured to directly trigger a fall alarm signal when the instantaneous acceleration it monitors continuously exceeds an independent fixed fall threshold. This procedure effectively decouples and processes external body motion interference and other safety events from internal structural instability signals in their respective logic channels, thereby reducing the false alarm rate caused by macro-body motion and expanding the system's safety monitoring function without sacrificing monitoring sensitivity.

[0032] In order to build a risk assessment model, the system is further configured to actively sense changes in the health status of the skin itself, because the decline in the skin barrier function is also a key physiological factor that induces leakage risk. For this reason, the edge computing unit is designed to periodically switch at least one pair of piezoelectric film sensors to an active detection mode; in this mode, a piezoelectric film sensor is driven to emit a series of multi-frequency weak alternating electrical signals, which penetrate the stoma base and act on the skin surface, while the other piezoelectric film sensor synchronously measures and records the bioelectrical impedance spectrum generated by the response of the skin tissue; in view of physiological parameters such as the health status of the skin stratum corneum and the degree of hydration It will directly and predictably affect its electrical impedance characteristics. The system can invert a quantitative skin barrier function status index through a preset biophysical model based on the measured bioelectrical impedance spectrum response; this status index is then used as a dynamic adjustment factor to correct the aforementioned information entropy alarm threshold used to determine the risk of structural instability. For example, when the skin barrier function index shows that the skin is relatively fragile or has subclinical damage, the system will automatically lower the information entropy alarm threshold, thereby initiating a more sensitive early warning strategy. This closed-loop feedback of physiological status and physical structure monitoring enables the system to provide early protection against potential leakage risks caused by skin problems themselves.

[0033] Finally, to improve the accuracy of the system's output warnings, this solution also integrates a final decision-making arbitration module, whose function is to filter out false alarms caused by sensor artifacts or small-scale transient electromagnetic interference. This module is immediately activated after the edge computing unit generates a risk warning signal containing azimuth information, and drives an embedded movable single-point infrared temperature sensor to perform a rapid target scan and temperature measurement of the risk location indicated by the warning and its immediate surrounding area. The arbitration logic of the decision-making arbitration module is based on a physical law, namely that real excrement leakage is inevitably accompanied by the evolution of a temperature field with specific thermodynamic characteristics. This module will strictly compare the real-time temperature distribution map obtained by scanning temperature measurement with an established heat conduction model. Only when the temperature measurement results show that the temperature of the target area presents a continuous gradient change spreading from the center to the surrounding area and the change pattern is highly consistent with the prediction of the heat conduction model, the decision arbitration module will finally confirm the authenticity of the warning and push a risk warning signal with precise location and clear level to the user or the preset nursing terminal; this arbitration mechanism based on cross-validation of orthogonal physical quantities provides a solid line of defense for the decision chain of the entire warning system, thereby improving the reliability of the information it outputs in complex and changeable clinical application scenarios.

[0034] Example 1: This example aims to combine a specific and dynamic nursing scenario to conduct a practical deduction and explanation of the collaborative operation mode of the above-mentioned technical solutions and the composite technical effects produced. In a pediatric postoperative monitoring unit, a child who has just completed abdominal surgery wears a stoma care device integrated with the early warning system. Due to postoperative discomfort and age factors, the child shows frequent and irregular body movements, such as tossing and turning in bed and swinging of limbs. At the same time, his skin is relatively sensitive. These factors together constitute a challenging scenario that places strict requirements on the sensitivity and anti-interference ability of the leakage monitoring system. After the system is activated, first, according to the established procedures, during a period of relative rest for the child, the system adaptively provides the child with a stoma. Each passive piezoelectric film sensor establishes its own personalized resting entropy baseline. At the same time, the system periodically switches the sensor pair to active detection mode, and inverts an initial skin barrier function status index by measuring the skin's bioelectrical impedance spectrum response. Since this index shows that the skin barrier function of children is slightly lower than the normal adult standard, the system automatically dynamically lowers the information entropy alarm threshold used to determine the risk of structural instability based on an initial standard value, thereby entering a targeted high-sensitivity monitoring state. This is the first synergy formed between the two modes of acoustic monitoring and electrical detection, that is, based on the prior assessment of the user's physiological state, the core parameters of physical structure monitoring are prospectively optimized.

[0035] During the subsequent monitoring process, the child turned over violently, which resulted in the instantaneous acceleration vector and The established physical threshold is exceeded momentarily , the dynamic compensation module immediately determines that the current moment and a short time window thereafter are the macro-impact pollution period; during this period, although the violent body movement also causes violent fluctuations in the output signal of the piezoelectric film sensor and causes a large jump in the calculated information entropy value, because the time window has been marked as a pollution period, the alarm judgment logic of the edge computing unit is temporarily suspended, so the system does not generate any alarms. By introducing an orthogonal macro-body motion information dimension, the system can accurately identify and filter out strong interference caused by non-leakage factors while maintaining high monitoring sensitivity, achieving high sensitivity and high specificity at the same time; a few hours later, when the child is in a sleeping state, the body movement level is much lower than When the threshold is reached, the information entropy value calculated by the piezoelectric film sensor located at the six o'clock position on the stoma bottom plate begins to show a weak but continuous and unidirectional upward trend, and exceeds the dynamically adjusted low alarm threshold for several consecutive sampling cycles. The edge computing unit generates an internal warning containing the first-level risk level and the six o'clock direction information; the warning signal is not immediately pushed to the nursing terminal, but the decision arbitration module is first activated, and the movable single-point infrared temperature sensor then scans and measures the temperature of the point indicated by the warning and its neighborhood. The temperature measurement results show that the temperature field in the area is uniform and does not show any Therefore, the decision arbitration module judged the abnormality of the acoustic signal as a non-continuous or small fluctuation of the adhesion interface that did not pose a real leakage threat, and rejected the output of the warning. Here, the early warning capability of acoustic monitoring and the decision confirmation capability of thermodynamic continuity verification formed a second key synergy. The former is responsible for capturing signals of potential risks from the time dimension, while the latter plays the role of arbitrator from the dimension of physical authenticity. This decision-making mechanism based on cross-validation of multiple physical quantities transforms the system's warning logic from a single signal detection to a fact confirmation.

[0036] After a period of time, the information entropy value of the sensor located at the same six o'clock direction showed a continuous abnormal increase again, and the amplitude and duration of its exceeding the threshold both met the risk level standard of the second-level high risk. The edge computing unit generated an internal warning again and triggered the decision arbitration module. At this time, the second scanning temperature measurement result of the movable single-point infrared temperature sensor showed that a tiny temperature field appeared in the target area with the warning point as the center, and the temperature continuously decreased toward the periphery and fully complied with the established heat conduction model. The decision arbitration module finally confirmed the high authenticity of the warning and immediately sent a composite warning information including the second-level high risk level and the precise position of the six o'clock direction to the central monitoring system of the nurse station. After receiving this highly credible and clear warning, the nurse was able to Before large-scale spillage contaminated clothing and bed sheets, targeted inspections and interventions were conducted on the child. By locally reinforcing the chassis edge at the six o'clock position, the risk of leakage was effectively controlled at an early stage, avoiding the complex nursing operations that required full-body cleaning and replacement of a full set of stoma devices. The entire process of this embodiment redefines the monitoring task as how to conduct a high-confidence risk assessment of the structural integrity of the adhesion interface in a complex dynamic environment, which may provide precise intervention guidance, by introducing acoustic information entropy analysis based on structural failure precursors, combined with multi-dimensional information such as macroscopic body motion compensation, physiological state adaptation, and final thermodynamic verification. The system processes the collected signals through a hierarchical, progressive, and multi-source verification information processing architecture and outputs a warning signal.

[0037] Example 2: In order to objectively verify the effectiveness of this technical solution in distinguishing between real structural failure signals and macroscopic body motion interference signals, as well as the specific role of the multimodal decision arbitration mechanism in improving the reliability of early warning, this control experiment was designed and executed. This experiment was built on a controllable in vitro simulation platform. The platform uses a silicone skin model with surface temperature and humidity precisely controlled within the normal range of human skin as a substrate to simulate the real wearing environment; a stoma chassis with a full set of early warning system components is pasted on the center of the model, and the platform also integrates a linear actuator driven by a servo motor to apply programmable mechanical impact to the skin model that simulates typical human movements. The acceleration generated by the impact force can be measured in real time by the system's built-in three-axis acceleration sensor; in addition, a micro-injection pump is inserted through a micro-catheter pre-buried between the chassis and the model. , it can inject physiological saline maintained at a constant temperature of 37°C into specific areas of the adhesion interface at a rate of 0.05 ml per minute to simulate the occurrence of early micro-leakage; the collection and processing of experimental data are entirely performed by the system's own edge computing unit, and the sampling period for signal processing is set internally. The technical trade-off lies in ensuring the balance between capturing the fidelity of weak acoustic signals and maintaining low-power operation of the system. Given that the main energy spectrum of the acoustic signal generated by micro-peeling of the adhesion interface is distributed in the frequency band of hundreds of hertz, in order to avoid signal aliasing under the Nyquist sampling theorem, the instantaneous sampling frequency within the signal acquisition window is set to 2kHz, and the period for the edge computing unit to calculate the information entropy of these high-frequency collected data and compare it with the baseline is maintained at a low frequency of once per second. In this way, the overall energy consumption of the system is optimized without sacrificing the accuracy of signal analysis.

[0038] The test process is divided into several consecutive stages. First, the system is run in a completely static initial state for 30 minutes to adaptively establish the resting entropy baseline of each sensor. Then, without fluid injection, the linear actuator is activated to simulate a peak acceleration exceeding the established physical threshold for five seconds. The actuator stops and the microinjection pump starts to inject simulated leakage liquid into the three o'clock position below the chassis. This is the pure microleakage stage. Finally, the actuator and the microinjection pump are started at the same time to simulate the complex scenario of the user performing activities while experiencing early microleakage. This is the complex interference stage. During the entire test process, the system continuously records the key data of all sensors. The data fragments in its representative state are shown in Table 1.

[0039] Table 1: A record of the key performance parameters of the system at different test stages.

[0040]

[0041] Analyzing the data shown in Table 1, in the pure body motion interference stage, the information entropy value instantly increased significantly to 3.80 times the baseline due to mechanical shock, but due to the instantaneous acceleration vector and the The dynamic compensation module determines this period as a macro-impact pollution period, which suppresses the warning signal output within this time window, so the final warning output is no. In contrast, in the pure trace leakage stage, The value remains at a resting level, while the information entropy value shows a slow and continuous growth. When it reaches 2.60 times the baseline value in the 3rd minute and triggers the internal acoustic alarm, the decision arbitration module is activated. At this time, the heat generated by the trace leakage has not yet formed a temperature gradient that can be confirmed, and the warning is temporarily suspended until the 5th minute. When the IR temperature gradient increases to 0.15°C / mm and meets the established heat conduction model, the system finally outputs a first-level warning. This process shows that the decision arbitration module can filter out acoustic anomalies caused by non-leakage factors or very early non-sustained threat events by introducing orthogonal physical quantity verification; in the composite interference stage, the system relies on dynamic The priority judgment of the compensation mechanism remains silent during body movement, and after the body movement ends, it correctly outputs an early warning based on the dual acoustic and thermodynamic evidence supported by the continued existence of real leakage. The results of this experiment show that this technical solution can distinguish between signals caused by structural micro-instability of the adhesion interface and signals caused by external mechanical impact by combining weak acoustic signal analysis based on information entropy with macroscopic body motion monitoring based on a three-axis accelerometer. At the same time, by introducing a decision arbitration module based on infrared temperature gradient scanning as the final early warning output gate, the system's decision-making is verified based on an independent physical dimension, thereby confirming the engineering feasibility of this solution in providing reliable leakage early warning in a dynamic interference environment.

[0042] Example 3: This example combines Figures 1 to 3 , description of AI visual stoma leakage warning system, such as Figure 1As shown in the figure, three passive piezoelectric film sensors are set up, which are located at 0°, 120° and 240° respectively, to collect the acoustic signals generated by the stoma base and skin adhesion interface during the structural instability process. The signal data is adaptively learned through the resting entropy baseline to establish an initial baseline for subsequent comparison, and is synchronously input into the edge computing unit for information entropy calculation and analysis. The unit calculates the information entropy value based on the real-time electrical signal and compares it with the resting entropy baseline. The information entropy alarm threshold is dynamically adjusted in combination with the results of the skin barrier function index bioelectrical impedance spectrum analysis; in addition, the system integrates a three-axis acceleration sensor Sensor body motion monitoring is used to monitor the wearer's body motion intensity. The collected body motion data enters the dynamic compensation module to eliminate body motion interference. When the macro-impact pollution period occurs, the information entropy judgment is suspended or compensated to ensure the accuracy of the judgment; when the edge computing unit calculates and analyzes the information entropy and identifies the risk of structural instability, the risk data enters the decision arbitration module for infrared temperature verification, and the risk direction is reconfirmed through infrared temperature scanning. Only when the temperature gradient change is consistent with the established heat conduction model, the system outputs the risk warning signal level + direction information, so that clinical personnel can take targeted intervention measures based on the level and direction information.

[0043] like Figure 2 As shown in the figure, the baseline alarm threshold marked with a dotted line represents the standard threshold level initially set by the system, while the dynamically adjusted threshold shown by the solid line reflects the alarm threshold after the system is dynamically corrected according to the skin barrier function status during the actual monitoring process. This mechanism switches the piezoelectric film sensor to active detection mode, measures the bioelectrical impedance spectrum of the skin, and inverts the skin barrier function index from it. Based on the negative correlation mapping relationship established between the index and the alarm threshold, the information entropy alarm threshold is dynamically lowered or adjusted, so that when the skin barrier function index is low, the alarm threshold is also reduced accordingly, thereby achieving a more sensitive risk response capability when the user's skin is fragile or there is potential damage, and enhancing the system's personalized early warning capability for high-risk users.

[0044] like Figure 3As shown, after the user starts the system, the edge computing unit first performs a power-on self-test and sends query pulses to the piezoelectric sensor group and acceleration sensor in turn. Each sensor returns impedance feedback to confirm that the sensor is normal. If the detection fails, a fault prompt indicator light / buzzer will be generated. After completing the hardware detection, the edge computing unit loads the global parameters and returns the control parameters from the memory. Then the system issues a system ready indication and enters a 30-minute learning phase. This phase is initiated by the edge computing unit and cyclically executes to collect an acoustic signal sequence once per second, continuously collect acoustic signals, and receive a return electrical signal sequence, and then calculate the information entropy value, perform statistical entropy data analysis on the entropy value sequence, and further calculate the resting entropy baseline. Finally, the calculation results are saved as the baseline values ​​of each sensor in the memory. If any sensor in the detection is abnormal, a fault prompt will be fed back.

[0045] Example 4: This example aims to provide a detailed description of the systematic calibration procedures of the core algorithm models and key parameter thresholds involved in the technical solution, which can be reproduced in engineering, to ensure that the internal decision-making logic of the system is completely transparent and certain before deployment. In a preparation stage for clinical verification, a series of internal models and parameters of the early warning system need to be calibrated and explicitly defined offline to solidify its response behavior under specific working conditions; this process begins with the established physical thresholds in the dynamic compensation module. The calibration process recruits multiple testers with different body shapes and characteristics, and requires them to wear a device with activated three-axis acceleration sensor and complete a series of preset and representative daily actions in sequence, including turning over from lying flat to lying on the side, from sitting to standing, and walking at a medium speed. During this process, the system continuously records the instantaneous acceleration vector and Time series data, and the generated The peak value and duration were statistically analyzed. The value of was determined as a limit that can distinguish unconscious macroscopic body movements that may produce strong acoustic interference from minor posture adjustments in the context of normal physiological activities. Its value was set to the value measured by all test subjects when performing turning over and sitting up. The tenth percentile of the peak distribution is used to ensure a high detection rate for strong interfering movements while avoiding overreaction to slight body movements.

[0046] The calculation of the information entropy value of the acoustic signal sequence within the edge computing unit follows the Shannon entropy definition. The information entropy value of the signal sampling sequence composed of discrete level levels is By calculating each level Probability of occurrence To determine, the relationship is , this entropy value quantifies the unpredictability of the signal; the information entropy alarm threshold used to trigger the initial internal warning It is not a fixed constant, it is determined by a baseline threshold With a dynamic adjustment item Jointly decided, the relationship is , where the baseline threshold The resting entropy baseline established adaptively with this sensor and its statistical standard deviation during the baseline establishment period Related, set to , which is used as a statistical benchmark for judging low-probability events; dynamic adjustment items The skin barrier function index measured by active detection mode Related, the index The acquisition is obtained by analyzing the impedance amplitude ratio of 1kHz and 100kHz at specific frequency points in the multi-frequency bioelectrical impedance spectrum. A lower ratio corresponds to a lower skin barrier function. A negatively correlated linear mapping relationship is established within the system to calculate , and the relationship is ,in, is a positive coefficient, which makes the skin condition index When descending, is negative, thereby dynamically lowering the overall alarm threshold , making the system more sensitive to users with poor skin conditions.

[0047] When the information entropy value continues to exceed the dynamic alarm threshold Finally, the system divides the structural instability risk level according to a risk score based on The quantization procedure is that the integral is accumulated over time and the amplitude exceeds the threshold. The increment of risk score The real-time value of information entropy and alarm thresholds The relationship is defined as ,in is a weight coefficient; the system presets multiple risk level score thresholds. Exceeding the threshold for the first time When the risk is determined to be level 1, When the warning is lifted, it is upgraded to the second-level risk. The integral decays over time after the warning is lifted, thus forming a dynamic risk assessment model. Finally, the established heat conduction model based on the decision arbitration module is a set of defined heat conduction models obtained by scanning with a movable single-point infrared temperature sensor. The logic rule set on the temperature matrix. When the internal warning is triggered, the infrared sensor quickly measures the temperature of nine points centered on the warning location to obtain a temperature matrix. The model contains two core judgment rules. Rule 1: the temperature of the center point must be the maximum value in the matrix. Rule 2: the temperature values ​​in the eight directions from the center point to the periphery must show a non-increasing trend, that is, there is no peripheral point with a temperature higher than its inner circle neighboring point. Only when the temperature matrix Only when these two rules are met simultaneously does the decision arbitration module determine that there is a continuous gradient change that conforms to the physical laws of real leakage and output a risk warning signal to the user. This set of deterministic logical rules realizes the verification of thermodynamic continuity in a computationally efficient manner, thereby completing the warning decision process.

[0048] Example 5: To ensure the universality and stability of the system model across different batches of hardware and diverse user groups, a set of offline model parameter global optimization and filling procedures must be executed before the final compilation of the system firmware. This procedure uses the aforementioned in vitro simulation platform to conduct hundreds of repeated experiments covering different leakage rates and interference intensities to obtain a benchmark data set for model training; specifically, the weight coefficient used for risk score calculation and a positive coefficient for dynamically adjusting the alarm threshold The value of is determined by performing regression analysis on the dataset. Its optimization goal is to maximize a composite evaluation function that combines the warning lead time and warning accuracy. At the same time, the setting of an independent fixed fall threshold is achieved by collecting the complete acceleration sequence of a standard dummy under controlled falls in different postures, and extracting a feature vector that can clearly distinguish real fall events from violent body movements such as jumping or sitting down, thereby setting a threshold with high classification efficiency.

[0049] After the firmware is burned, each independent warning system product must pass an automated factory quality assurance test process before it is delivered for use. The process is carried out in a standardized test black box. The device under test is fixed on a test platform that integrates a multi-physics field excitation source. The test platform is configured to apply a set of standardized physical signal excitations to the device under test in a preset sequence, including an acoustic white noise signal with a certain information entropy value emitted by a piezoelectric ceramic transducer, an acceleration generated by a micro-vibration table that just crosses a predetermined physical threshold, and a set of standardized physical signal excitations to the device under test in a preset sequence. The system automatically records the intermediate calculation values ​​and final warning decisions output by the edge computing unit of the device under test, and compares them with a standard response database stored in the test host. Only when the errors of all key output items are within the preset allowable tolerance range will the device be judged as qualified. This procedure is used to ensure that the perception accuracy and decision-making logic of each system delivered to the user meet the design standards.

[0050] Example 6: This example aims to explain the standardized installation, self-test and parameter loading procedures of the early warning system before it is first applied to a specific user, so as to ensure the functional integrity of the system and the adaptability of the monitoring parameters. Before the stoma base plate integrated with the system is attached to the skin, the system is configured to perform a power-on self-test procedure, which is used to verify the functional integrity of all sensors and the reliability of the electrical connections. The edge computing unit will send a preset electric pulse query signal to each passive piezoelectric film sensor and three-axis acceleration sensor in turn, and measure the impedance characteristics of their feedback. Only when the feedback impedance values ​​of all sensors fall within a range determined by their hardware The system will enter standby mode only when it is within the preset tolerance range determined by the specifications. If any sensor feedback is abnormal, the system will point out the location of the faulty sensor to the operator through an indicator light or a buzzer prompt, thereby avoiding starting monitoring when the sensor physically fails or the connection is poor. During the installation process, in order to achieve the optimal resolution accuracy of the risk direction, the layout of at least three passive piezoelectric film sensors follows a standardized geometric configuration, that is, with the central hole of the stoma base as the center of the circle, they are evenly distributed at equal angles of 120 degrees on the radial center line of the viscous hydrocolloid. This layout is used to ensure acoustic signal coverage without blind spots in the area surrounding the adhesion interface.

[0051] When the device is properly worn and the user first activates the system through a terminal device, the system loads a set of global control parameters, determined through offline calibration, from its non-volatile memory. These parameters include a duration window defining the macro-impact contamination period. The length of this time window is determined based on statistical analysis of a large amount of body motion interference experimental data. Its value is set to the 95th percentile of the time required for the energy of the acoustic signal artifact caused by the observed macro-impact to decay to the baseline level, ensuring that this window fully covers the aftereffects of most body motion interference. In addition, the system's internal event processing logic is configured as a prioritized interrupt response mechanism. Signals detected by the triaxial accelerometer exceeding a fixed fall threshold are assigned the highest interrupt priority. When this signal is triggered, the edge computing unit immediately suspends all ongoing computational tasks related to leakage risk analysis and prioritizes sending a fall alarm signal. The system will not resume its regular leakage warning monitoring process until the high-priority alarm state is resolved. This mechanism ensures that the system's response to emergency events that endanger user safety is given the highest priority.

[0052] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI visual stoma leakage early warning system, characterized in that: include: at least three passive piezoelectric film sensors, which are embedded in the stoma baseplate and configured to convert weak acoustic signals generated at the stoma baseplate and skin adhesion interface during structural instability into electrical signals; an edge computing unit, the edge computing unit being electrically connected to the piezoelectric film sensor and configured to: periodically collect a sequence of electrical signals output by the plurality of piezoelectric film sensors; Independently calculate the information entropy value of the electrical signal sequence; compare the calculated real-time information entropy value with the established resting entropy baseline; Based on the comparison results, when the real-time information entropy value of at least one piezoelectric film sensor is continuously higher than the preset difference threshold of its corresponding resting entropy baseline, the risk of structural instability at the adhesion interface is determined and a risk warning signal is generated. The risk warning signal is used to indicate the occurrence of structural instability risk.

2. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: The edge computing unit is also configured to: collaboratively analyze the real-time information entropy values ​​calculated by multiple piezoelectric film sensors, by identifying situations where one or more information entropy values ​​are first and continuously higher than the information entropy values ​​of other sensors, and combining the established physical positions of the piezoelectric film sensors with abnormal information entropy values ​​on the stoma base, accurately locate the spatial orientation of the risk of structural instability at the adhesion interface, thereby including this orientation information in the risk warning signal to guide users to take targeted intervention measures.

3. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: The edge computing unit is further configured to divide the structural instability risk into multiple deterministic risk levels according to the magnitude and duration of the real-time information entropy value exceeding the resting entropy baseline.

4. The AI ​​visual stoma leakage warning system according to claim 2, characterized in that: The edge computing unit is further configured to divide the structural instability risk into multiple deterministic risk levels according to the magnitude and duration of the real-time information entropy value exceeding the resting entropy baseline.

5. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: The resting entropy baseline is adaptively established by statistically averaging the information entropy values ​​over a fixed period of time after the system is initially worn.

6. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: The periodic collection frequency of the edge computing unit is less than 1 time per second, and the duration of each collection does not exceed 1 second.

7. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: Also includes: a three-axis acceleration sensor configured to monitor the wearer's body movement intensity in real time; And a dynamic compensation module, the dynamic compensation module is electrically connected to the three-axis acceleration sensor and the edge computing unit, and is configured to: when the instantaneous acceleration vector and the instantaneous acceleration vector detected by the three-axis acceleration sensor are equal Exceeding a given physical threshold , that is, satisfy When, among them, represents the instantaneous acceleration vector sum, Indicates a given physical threshold, and determines that the current moment and the specified time window thereafter are the macro-impact pollution period; during the macro-impact pollution period, suspends the edge computing unit's determination of the validity of the information entropy of the electrical signal sequence or performs noise compensation processing on the electrical signal sequence based on the data of the three-axis acceleration sensor.

8. The AI ​​visual stoma leakage warning system according to claim 1, characterized in that: The edge computing unit is further configured to periodically switch at least one pair of piezoelectric film sensors to an active detection mode; In active detection mode, one of the piezoelectric film sensors is driven to emit a multi-frequency alternating electrical signal to the skin, and the other piezoelectric film sensor is used to measure the skin's bioelectrical impedance spectrum response; the skin barrier function status index is inverted based on the bioelectrical impedance spectrum response; and the information entropy alarm threshold used to determine the risk of structural instability is dynamically adjusted based on the status index.

9. The AI ​​visual stoma leakage warning system according to claim 2, characterized in that: It also includes a decision arbitration module, which includes a movable single-point infrared temperature sensor. The decision arbitration module is configured to: when the edge computing unit generates a risk warning signal containing orientation information, drive the movable single-point infrared temperature sensor to scan and measure the temperature of the risk warning point and its adjacent areas; only when the scanning temperature measurement results show that the temperature presents a continuous gradient change and is consistent with the established heat conduction model, the decision arbitration module will finally output a risk warning signal to the user.

10. The AI ​​visual stoma leakage warning system according to claim 7, characterized in that: The three-axis acceleration sensor is also configured to trigger a fall alarm signal when the instantaneous acceleration it monitors continuously exceeds a fixed fall threshold. The fall alarm signal is sent to the user or a preset caregiver terminal through the edge computing unit.

Images