Medical waste liquid multi-mode risk perception and autonomous response system based on edge intelligence

Abstract

The invention belongs to the technical field of artificial intelligence and environmental safety monitoring, particularly relates to a medical waste liquid multi-mode risk perception and autonomous response system based on edge intelligence, and aims to solve the problems of risk identification lagging, passive response and insufficient multi-source data fusion in medical waste liquid management. According to the system, a pressure, spectrum, biological metabolism and deformation multi-mode sensing array is integrated at a container end, and a three-level fusion engine in an edge calculation module is combined, so that localized risk comprehensive research and judgment and autonomous response are realized. Through cooperative monitoring of pressure micro change, component analysis and biological activity and structural deformation, closed-loop control from sensing to execution is constructed, intelligent plugging, precise neutralization and audible and visual alarm are supported, the real-time performance and accuracy of risk response are improved, and the dependence on the cloud and the risk of data leakage are reduced.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and environmental safety monitoring technology, specifically relating to a multimodal risk perception and autonomous response system for medical waste liquid based on edge intelligence. Background Technology

[0002] In the field of medical waste management, medical waste liquids pose a significant challenge to the environmental safety and public health of medical institutions due to their potential for biological pollution and chemical hazards. With the expansion of hospitals and the increase in medical activities, the amount of waste liquid generated continues to rise. Its composition is complex and its sources are diverse. Traditional treatment methods largely rely on manual inspections and centralized disposal processes, lacking the ability to dynamically perceive the risk status of waste liquids in real time. Although modern sensing technologies and network communication methods have been gradually applied to environmental monitoring systems, initially achieving data collection for some parameters (such as liquid level and pH value), the overall management system still struggles to address key issues such as sudden leaks, cross-contamination, and delayed treatment.

[0003] Among these, medical waste monitoring based on intelligent sensing and automatic control is becoming an important direction for improving safety management efficiency. This technology aims to achieve continuous monitoring of waste storage and transportation processes by deploying distributed sensor nodes, and to combine this with data analysis models for risk identification and early warning response. However, existing systems have revealed problems in practical applications, such as limited sensing dimensions, high response latency, bandwidth pressure from centralized processing, and privacy risks. Especially when facing difficulties in fusing multi-source heterogeneous data, limited edge computing resources, and insufficient accuracy in anomaly event detection, the reliability and real-time performance of the system are difficult to guarantee.

[0004] Existing technologies generally employ a centralized cloud architecture for data processing and decision generation, resulting in an excessively long chain from perception to response, failing to meet the stringent requirements of low latency and high availability in medical scenarios. Furthermore, the spatiotemporal alignment and semantic consistency mechanisms between different modal sensors (such as vision, gas, temperature, and flow) are weak, making it difficult to effectively identify complex risks (such as the accumulation of volatile gases accompanied by high-temperature fermentation). In addition, the system lacks autonomous closed-loop control capabilities, and most emergency measures still require manual intervention, hindering the improvement of overall intelligence. Especially in high-risk areas such as operating rooms and ICUs, if waste fluid leaks are not detected in time and isolation or ventilation procedures are not initiated, they can easily lead to the spread of nosocomial infections. Therefore, there is an urgent need to build a new medical waste fluid risk management system with multimodal collaborative perception, real-time edge analysis, and autonomous response capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal risk perception and autonomous response system for medical waste liquid based on edge intelligence, in order to solve the technical problems existing in the management of medical waste liquid, such as lagging risk identification, passive response mechanisms, reliance on manual inspections, and insufficient ability to fuse multi-source heterogeneous data. Currently, medical institutions generate large amounts of waste liquid daily that is biologically contaminated, chemically corrosive, or radioactive, and its treatment involves multiple stages such as collection, temporary storage, transfer, and final disposal. Traditional management models generally adopt timed inspections, fixed threshold alarms, and manual intervention, which are insufficient to cope with complex risk scenarios such as sudden leaks, abnormal changes in composition, or damage to container structures. Especially under the real-world conditions of dynamic changes in waste liquid composition and significant environmental interference, single-sensor detection is susceptible to interference and has a high false alarm rate; centralized data processing architectures have inherent defects such as large communication delays, high privacy leakage risks, and poor real-time system response. Furthermore, existing technologies lack the ability to collaboratively model the three core risks of physical leakage, chemical exceedances, and biological activity, and cannot achieve closed-loop autonomous control from "perception-analysis-decision-execution".

[0006] The technical solution of this invention is to construct an edge intelligent system distributed across various waste liquid storage nodes. This system deploys an integrated multimodal sensor array in each waste liquid container unit and achieves localized risk fusion judgment and autonomous response triggering through an embedded edge computing module. The multimodal sensor array includes a pressure micro-change detection subsystem, a spectral feature analysis subsystem, a biological metabolic activity monitoring subsystem, and a three-dimensional deformation sensing subsystem. The pressure micro-change detection subsystem consists of a piezoresistive micro-pressure sensor arranged at the bottom of the container, used to continuously acquire liquid static pressure signals at a sampling frequency of not less than 100Hz. It captures sudden pressure drops or gradual pressure changes using a sliding window differential algorithm to identify potential leaks or illegal dumping. The spectral feature analysis subsystem employs narrowband tunable laser absorption spectroscopy technology, scanning in the near-infrared band within the 2.3μm to 2.5μm range to obtain the absorption spectra of characteristic functional groups in the waste liquid. Combined with a pre-set database of typical medical waste liquid components, it analyzes the waste liquid type and concentration distribution in real time to determine whether there is any illegal mixing of non-standard substances or excessive levels of highly toxic chemicals. The biological metabolic activity monitoring subsystem is based on the principle of micro-electrochemical oxygen consumption rate detection. It incorporates a platinum working electrode and a reference electrode to measure the dissolved oxygen consumption gradient per unit time, quantifying metabolic intensity into a biological risk index. When this index exceeds a set safe range, it characterizes the risk of abnormal proliferation of pathogenic microorganisms. The three-dimensional deformation sensing subsystem utilizes a flexible fiber Bragg grating network attached to the outer wall of the container to reconstruct the strain field distribution on the container surface in real time. Through pattern matching algorithms, it identifies precursors of mechanical damage such as localized bulges, crack propagation, or instability of the supporting structure.

[0007] The raw data from the four subsystems are strictly synchronized on the timeline and input into the multi-level risk fusion engine in the edge computing module. This fusion engine first performs a data preprocessing stage, including signal denoising, baseline drift correction, and dimensional normalization. Further, the data enters the primary risk discrimination layer, which consists of four parallel lightweight neural network models: a pressure anomaly detection network, a component deviation identification network, a bioactivity early warning network, and a structural damage diagnosis network. Each network outputs a risk confidence score for its corresponding modality, ranging from 0 to 1. In one embodiment of the invention, each neural network in the primary discrimination layer employs a depthwise separable convolutional structure, with the number of parameters controlled to within 50,000, ensuring efficient inference on edge devices with limited computing power. Further, the risk confidence scores for all modalities are fed into the intermediate evidence fusion layer. This layer uses an improved Dempster-Shafer evidence theory framework for uncertainty inference, introducing a dynamic conflict allocation factor to suppress misfusion phenomena under highly contradictory evidence, and outputs a comprehensive risk assessment vector. Furthermore, the comprehensive risk assessment vector is input to the advanced contextual understanding layer, which defines five risk levels based on a finite state machine model: normal state, attention state, warning state, emergency state, and out-of-control state. The transition conditions for each state are determined by multi-dimensional threshold combination logic and associated with different sets of response strategies.

[0008] When the system determines that it has entered a warning state or higher, the autonomous response actuator is activated. In one embodiment of the invention, the autonomous response actuator includes an intelligent sealing device, an emergency neutralizing agent injection unit, and a local audible and visual warning module. The intelligent sealing device is integrated near the discharge port at the bottom of the container and includes a shape memory alloy-driven metal sealing valve. Upon receiving a locking command, it can complete physical isolation of the channel within 3 seconds. The emergency neutralizing agent injection unit is equipped with a dual-chamber storage bladder, loaded with acidic and alkaline neutralizing agents respectively. Based on the properties of the main pollutants identified by the spectral analysis subsystem, a matching neutralizing agent type is selected and injected into the waste liquid body via a micro-plunger pump. The injection volume is calculated jointly by the pollutant concentration and volume parameters. The local audible and visual warning module activates a high-frequency buzzer and flashing red light, while simultaneously sending an alarm summary to a nearby workstation via a short-range wireless communication protocol. Furthermore, the system retains the right to upload encrypted event logs to the higher-level monitoring platform, but the uplink is only activated when an emergency or out-of-control state is confirmed. In other cases, all data processing and decision-making are completed locally in a closed loop, minimizing network dependence and data exposure risks.

[0009] As one embodiment of the present invention, the operating system of the edge computing module adopts a real-time microkernel architecture, and task scheduling priorities are strictly divided according to risk levels to ensure that the processing delay of high-risk events does not exceed 500 milliseconds. Furthermore, all sensor components of the multimodal sensing array have self-diagnostic functions, regularly performing stimulus-response tests. Once sensor failure or performance degradation is detected, the system automatically adjusts the fusion weights and records the fault information to the maintenance queue. Furthermore, the laser wavelength tuning sequence of the spectral feature analysis subsystem is dynamically compensated based on ambient temperature feedback to avoid spectral line shift errors caused by temperature drift. Furthermore, the electrode surfaces of the biological metabolic activity monitoring subsystem are coated with a selectively permeable membrane to suppress cross-interference between chloride ions and sulfides, improving long-term operational stability. Furthermore, the fiber Bragg grating demodulation algorithm of the three-dimensional deformation sensing subsystem adopts phase-sensitive OTDR technology, achieving a spatial resolution of 2 cm, which can accurately locate the deformation location. Furthermore, the entire system is powered by a combination of a top-integrated photovoltaic thin film and a supercapacitor bank, allowing for continuous operation for no less than 72 hours without an external power source.

[0010] As one embodiment of the present invention, the system supports a dynamic knowledge base update mechanism. New spectral characteristics of waste liquid components, novel pathogen metabolic patterns, or structural failure cases can be injected into edge devices via secure, offline loading for incremental training of the neural network model in the primary discriminant layer. The model update process is verified in a sandbox environment before activation, ensuring continuous evolution of the system's adaptability. Furthermore, a point-to-point trust network is established between multiple adjacent edge intelligent nodes. When a node enters an emergency state, it can broadcast a risk diffusion prediction model to surrounding nodes, triggering a joint prevention and control mechanism to collaboratively adjust their respective response strategies. Furthermore, container identity information and system firmware version numbers are bound to a hardware security chip to prevent unauthorized replacement or tampering. Furthermore, all internal communication buses use AES-128 encryption for transmission, and critical control commands must be verified with digital signatures before execution.

[0011] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This solution deploys a multimodal sensor array at the waste liquid container end, integrating pressure, spectral, biological, and deformation sensing capabilities. This enables the simultaneous capture of three core risks: physical leakage, chemical contamination, and biological pollution. This fundamentally overcomes the technical bottlenecks of single-point detection, which is susceptible to interference and has high false alarm / false negative rates. Multimodal data is processed at the edge via a three-level progressive fusion engine, sequentially completing intramodal feature extraction, cross-modal evidence integration, and global contextual understanding. This ensures that risk assessment relies not only on a single indicator exceeding limits but also on a comprehensive judgment supported by multi-source evidence, significantly improving the accuracy and robustness of risk identification. The system fully embeds the entire processing logic from sensing to response into the edge computing module. All high-real-time control commands are generated and executed locally, with response latency controlled at the millisecond level. This eliminates reliance on centralized cloud processing and solves the problems inherent in traditional architectures. The system overcomes fundamental flaws such as high communication latency and paralysis due to network outages. It features deep coupling between autonomous response mechanisms and risk-level state machines, with differentiated physical interventions triggered by different risk levels. The introduction of intelligent blocking and precise neutralization functions enables the system to proactively curb risk spread, achieving a paradigm shift from "passive alarm" to "proactive prevention and control." A localized closed-loop processing strategy strictly limits the dissemination of sensitive data, uploading de-identified event summaries only when necessary, effectively protecting the operational privacy and data sovereignty of medical institutions. The system adopts a low-power design and renewable energy power supply, supporting long-term unattended operation. Combined with dynamic knowledge base updates and inter-node collaborative defense mechanisms, it forms an intelligent waste management terminal network with self-learning, self-adaptive, and self-organizing capabilities, providing solid technical support for building a safe, efficient, and reliable modern medical waste management system. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall technical architecture of the medical waste liquid multimodal risk perception and autonomous response system based on edge intelligence proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-level risk fusion engine in this invention; Figure 3 This is a flowchart illustrating the logical process of data preprocessing and primary risk assessment between the multimodal sensing array and the edge computing module in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between edge intelligent nodes, autonomous response execution mechanisms, and upper-level monitoring platforms in this invention. Detailed Implementation

[0013] Please refer to Figures 1 to 4This invention provides a multimodal risk perception and autonomous response system for medical waste liquid based on edge intelligence. This system enables real-time perception, fusion judgment, and autonomous intervention of three core risks associated with the storage and transportation of medical waste liquid: physical leakage, chemical contamination, and biological pollution. The system is deployed at various waste liquid temporary storage nodes within medical institutions, with each node corresponding to an intelligent terminal unit possessing complete perception, computation, and execution capabilities. The system consists of a multimodal sensor array, an edge computing module, a multi-level risk fusion engine, and an autonomous response execution mechanism. All components are interconnected via a high-reliability industrial bus, forming a closed-loop autonomous control system. The overall operating logic of the system begins with the synchronous acquisition of multi-source heterogeneous data, followed by a three-level progressive information fusion mechanism to determine the risk level from raw signals to contextual awareness. Finally, based on the risk status, it triggers localized and differentiated physical response actions, all completed independently without external network dependence.

[0014] The multimodal sensor array, serving as the system's front-end sensing layer, integrates a pressure micro-change detection subsystem, a spectral feature analysis subsystem, a biological metabolic activity monitoring subsystem, and a three-dimensional deformation sensing subsystem. These four subsystems are spatially arranged collaboratively around the same waste liquid container, and temporally, they employ a unified hardware clock source to achieve strictly synchronized sampling, ensuring temporal consistency of cross-modal data. All sensor components are encapsulated in a corrosion-resistant housing meeting IP68 protection standards, adapting to harsh operating environments such as high humidity and strong acid / alkali vapors.

[0015] The pressure micro-change detection subsystem is located in the load-bearing area at the bottom of the container, employing a piezoresistive micro-pressure sensor array to continuously monitor the liquid static pressure. The sensor array consists of four independent sensing units arranged in a rectangular pattern. Each unit has a range of 0 to 50 kPa, a resolution of at least 0.01 kPa, and a nonlinearity error of less than 0.2% of full scale. The system continuously acquires pressure sequence data at a frequency of 100 Hz and inputs it to the sliding window differential processing unit in the edge computing module. This unit maintains a sliding data window with a length of 1 second, containing 100 consecutive sampling points. Each time a new data point arrives, the window content is updated, and the first-order differential sequence of pressure values ​​within the current window is recalculated. The absolute mean of the differential sequence is defined as the instantaneous pressure change rate index. When this index exceeds a preset dynamic threshold, the system determines that a sudden pressure drop event has occurred, possibly caused by container rupture or illegal discharge. When the index shows a monotonically decreasing trend across multiple consecutive windows and the cumulative decrease exceeds 15% of the initial static pressure value, it is identified as a slow leakage process. To suppress the interference caused by liquid density changes due to temperature fluctuations, the system introduces a temperature compensation factor. This factor is acquired in real time by a digital temperature sensor attached to the container wall, and the original pressure reading is corrected using an empirical formula. The corrected pressure value is denoted as... Its calculation follows the following relationship:

[0016] in, This is the original pressure measurement value; The coefficient of thermal expansion of the liquid is determined through calibration experiments during the system initialization phase for different types of waste liquids, with a typical value range of 0.0002 to 0.0008 per degree Celsius. The standard reference temperature is set to 25 degrees Celsius. This is the current measured temperature value. The compensated pressure data is then fed into the pressure anomaly detection network of the primary risk assessment layer for further analysis.

[0017] The spectral feature analysis subsystem employs narrowband tunable laser absorption spectroscopy, operating in the near-infrared region from 2.3 to 2.5 micrometers. This range includes characteristic vibrational absorption peaks of various organic functional groups such as hydroxyl, carboxyl, and amino groups. The system incorporates a tunable distributed feedback laser, whose driving current is controlled by a precision digital-to-analog converter, enabling wavelength scanning within the target range in 0.1 nanometer steps. The emitted beam, after collimation, passes through a quartz light-transmitting window located on the side wall of the container, penetrates the waste liquid, and is received by an indium gallium arsenide photodetector on the opposite side. Each complete scan generates an absorbance spectrum containing 2001 wavelength channels, with a sampling period of 30 seconds. The system is equipped with a pre-built spectral database of typical components of medical waste liquids, storing standard absorption fingerprint spectra of 12 common waste liquids, including physiological saline, iodine solution, chlorine-containing disinfectants, residual chemotherapy drugs, and blood product residues. Each type of spectrum is acquired using a high-precision laboratory spectrometer and normalized. The real-time acquired spectral data is first corrected for baseline drift by selecting bands with no significant absorption at both ends of the spectral line for linear fitting and subtracting the background trend term. Subsequently, the system calculates the Pearson correlation coefficient between the current spectrum and various standard spectra in the database, taking the category corresponding to the highest correlation coefficient as the preliminary classification result. If the highest correlation coefficient is below 0.9, it is judged as a risk of unknown substance contamination; if sharp absorption peaks characteristic of highly toxic chemicals such as nitrobenzenes and cyanides are detected, and the concentration inversion value exceeds the safety limit of 5 mg / L, the risk level is immediately upgraded. Concentration inversion uses a multiple linear regression model, with absorbance values ​​at multiple characteristic wavelengths as input and estimated target pollutant concentration as output. The laser wavelength tuning sequence is dynamically adjusted based on ambient temperature feedback to avoid wavelength shifts caused by temperature control failure, ensuring long-term measurement stability.

[0018] The biological metabolic activity monitoring subsystem is built based on the principle of micro-electrochemical oxygen consumption rate detection. Its core component is an integrated dissolved oxygen sensor probe, which internally contains a platinum working electrode, a silver / silver chloride reference electrode, and a potassium ion electrolyte gel layer. The system applies a 0.7-volt polarization voltage to the working electrode using a constant voltage method, driving a reduction reaction in dissolved oxygen. The resulting limiting diffusion current is proportional to the dissolved oxygen concentration in the solution. The system samples dissolved oxygen concentration every 10 seconds, continuously collecting 60 data points over 10 minutes to construct a time series. The system calculates the average dissolved oxygen consumption gradient per unit time, defined as a biological risk index. Its expression is as follows:

[0019] in, This is the dissolved oxygen concentration value at the first sampling point, in milligrams per liter. This represents the dissolved oxygen concentration value at the 60th sampling point. When When the value exceeds 0.05 mg / L / min, the system determines that abnormal microbial metabolic activity exists, indicating a potential risk of pathogen proliferation. To suppress the influence of interfering ions such as chloride ions and sulfides on the electrode response, a 5-micron-thick selectively permeable membrane is coated on the working electrode surface. This membrane allows oxygen to pass freely but blocks charged interfering particles. The system periodically performs an electrode activation procedure, briefly reversing the polarization voltage to remove deposits on the electrode surface and maintain long-term measurement accuracy. Zero-point calibration is automatically performed before each sampling by briefly exposing the probe to a nitrogen environment to obtain a zero-oxygen reference value.

[0020] The three-dimensional deformation sensing subsystem utilizes a flexible fiber Bragg grating network attached to the outer wall of the container to reconstruct the strain field across the entire surface. The fiber network consists of eight parallel single-mode fibers, each with a 2-cm-spaced Bragg grating array, totaling 400 sensing points covering the entire sidewall region of the cylindrical container. The system employs phase-sensitive optical time-domain reflectometry to demodulate the fibers, injecting narrow-pulse coherent light into them and receiving the backscattered Rayleigh signal. Local strain states are inverted by analyzing the phase difference change between adjacent pulse echoes. The spatial resolution reaches 2 cm, and the strain measurement accuracy is better than 1 microstrain. The system establishes a baseline strain field template under intact container conditions. Real-time monitoring data is compared point-by-point with the template. When the strain increment at three or more consecutive sensing points in a certain area exceeds 70% of the material's yield strength threshold, and the spatial distribution exhibits localized concentration characteristics, it is identified as a bulge or a precursor to crack propagation. When an asymmetric strain distribution appears in the bottom support ring area, it is determined to be a risk of structural instability. The system supports precise positioning of the deformation location with a positioning error of no more than 2 centimeters, providing spatial guidance for subsequent maintenance decisions.

[0021] The raw data from the four subsystems mentioned above, after being timestamped, are uniformly input into the multi-level risk fusion engine in the edge computing module. This engine runs on an ARM-based embedded processor with a clock speed of 1.2 GHz and 2 gigabytes of memory. It is equipped with a lightweight Linux real-time operating system, and task scheduling employs a priority preemption mechanism to ensure that the latency of high-risk event processing paths does not exceed 500 milliseconds. The engine's processing flow is divided into three levels: a data preprocessing layer, a primary risk assessment layer, an intermediate evidence fusion layer, and a high-level contextual understanding layer.

[0022] The data preprocessing layer is responsible for standardizing the raw sensor data. Pressure signals are denoised using wavelet transform with the db4 wavelet basis function, decomposed into four layers, and soft thresholding is applied to high-frequency coefficients. Spectral data undergoes standard normal transformation to eliminate batch variations. Biological signals are smoothed using moving average filtering to eliminate random fluctuations. Deformation data is spatially interpolated to fill in missing points caused by localized fiber damage. All modal data are processed and uniformly mapped to a normalized range of 0 to 1, eliminating dimensional differences and providing a consistent input format for subsequent fusion.

[0023] The primary risk assessment layer consists of four parallel, lightweight neural network models: a pressure anomaly detection network, a component deviation recognition network, a bioactivity early warning network, and a structural damage diagnosis network. Each network employs a depthwise separable convolutional structure, with the total number of parameters controlled within 48,000 to meet edge device resource constraints. The pressure anomaly detection network takes a time-series pressure differential value of length 100 as input. The network contains two depthwise separable convolutional blocks, each consisting of 16 convolutional kernels, followed by a global average pooling layer and a fully connected output layer. The output is a pressure risk confidence score. The value ranges from 0 to 1. The input to the component deviation recognition network is a 2001-dimensional spectral vector. The network adopts a 1-dimensional convolutional structure, containing three convolutional layers with kernel sizes of 64, 32, and 16, respectively. The activation function is a modified linear unit. The last layer outputs a category probability distribution, and the maximum probability value is taken as the component risk score. The bioactivity early warning network takes 60 dissolved oxygen time series data as input, uses a long short-term memory network structure with 32 hidden layers, and outputs a biorisk score. The structural damage diagnosis network takes a 400-point strain spatial distribution matrix as input, employs a two-dimensional convolutional neural network with a 3×3 kernel size and two layers, and outputs a structural risk score. All networks were trained on a simulation dataset before deployment. The loss function used was weighted cross-entropy, the optimizer was Adam's algorithm, the learning rate was set to 0.001, and the training iterations were 500 epochs. The model weights were stored in read-only memory to prevent accidental tampering.

[0024] The intermediate evidence fusion layer employs a modified Dempster-Shafer evidence theory framework for uncertainty reasoning. The system defines the identification framework. This represents a basic assessment of the current system state. Risk scores from four modalities. Transformed into the basic probability assignment function , ,in Traditional DS synthesis rules may lead to unreasonable conclusions under highly conflicting evidence; therefore, this system introduces a dynamic conflict allocation factor. Its calculation is based on the Jaccard similarity matrix between each piece of evidence. Let... As evidence and The Jaccard similarity is then the overall conflict level. Defined as:

[0025] in =4 represents the number of evidence sources. The system based on... The value dynamically adjusts the conflict redistribution ratio in the composition rule when When the risk level is greater than 0.6, a conservative fusion strategy is employed to reduce the overall impact weight of highly conflicting evidence. The final output is a comprehensive risk assessment vector. ,in It serves as the primary input for the advanced contextual understanding layer.

[0026] The advanced contextual understanding layer uses a finite state machine model to determine and transition between five risk levels. The state set is defined as: normal state, concern state, warning state, emergency state, and out-of-control state. State transition conditions are determined by multi-dimensional threshold combination logic. The system is in normal condition when the score is <0.3 and no single risk score exceeds 0.6; when any single risk score is between 0.6 and 0.8, or When the value is between 0.3 and 0.5, the system enters a state of high alert, initiating enhanced monitoring mode and increasing the sampling frequency of each subsystem by 50%; when... When the risk score exceeds 0.5 or any two risk scores are simultaneously higher than 0.6, an early warning state is activated, the local audible and visual warning module is activated, and the response agency is prepared; when When the risk score exceeds 0.7 or any single risk score exceeds 0.9, it is determined to be an emergency state and triggers an autonomous response command; when the system self-diagnoses and finds that a key sensor has failed or communication has been interrupted and the risk score continues to rise, it is defined as an out-of-control state, and the highest level response is forcibly started and an alarm is attempted to be transmitted through the backup channel.

[0027] The autonomous response actuator is deeply coupled with the risk level state machine, triggering differentiated physical intervention measures for different risk levels. When the system enters a warning state or higher, the actuator is activated and enters standby mode. The intelligent sealing device is integrated into the upstream pipeline of the discharge port at the bottom of the container. The main body is a double-lobe metal sealing structure. The drive unit uses a nickel-titanium alloy shape memory wire, which remains relaxed at room temperature and contracts at a phase change temperature of 80 degrees Celsius after being energized, pulling the transmission rod to close the sealing lobe. Upon receiving the locking command, the control circuit outputs a 2-amp DC current to heat the alloy wire, completing complete channel isolation within 3 seconds, with a pressure resistance of not less than 1 MPa. The emergency neutralizing agent injection unit is equipped with a dual-chamber storage bladder, containing a 2 mol / L sodium hydroxide solution and a 1 mol / L hydrochloric acid solution, each with a volume of 500 ml. The system determines the type of neutralizing agent based on the properties of the main pollutants identified by the spectral feature analysis subsystem: if a strongly acidic substance is detected (pH < 3), an alkaline neutralizing agent is selected; if a strongly alkaline substance is detected (pH > 11), an acidic neutralizing agent is selected. Injection volume. From pollutant concentration With waste liquid volume The formula, obtained through joint calculation, is as follows:

[0028] in The effective concentration of the neutralizing agent is set at 1 mol / L. A miniature plunger pump, driven by a stepper motor, injects precisely according to the calculated value, with an accuracy of ±1 ml. A local audible and visual warning module, comprising a 120 dB high-frequency buzzer and a flashing red light array, is installed in a prominent position on the top of the container. Upon activation, it alternates between a 2 Hz buzzer and flashing lights for at least 10 minutes. After all response actions are completed, the system generates an encrypted event log, including a timestamp, risk type, triggering condition, executed action, and result feedback. This log is encrypted using the AES-128 algorithm and temporarily stored in local solid-state storage.

[0029] The system reserves the right to upload data to the higher-level regulatory platform, but strictly limits the timing and content of uploads. Under normal and monitored conditions, the system only periodically uploads anonymized statistical summaries, including daily average pressure fluctuations, weekly spectral stability indices, baseline biological risk levels, and structural health scores. The upload cycle is 24 hours, and the data undergoes irreversible anonymization via hashing. When an emergency or out-of-control state is confirmed, the system activates the uplink wireless communication module to upload complete event logs and original risk data fragments through a dedicated secure tunnel. The transmission process uses TLS 1.3 encryption, and the recipient must provide a valid digital certificate for decryption. In all other cases, all data processing and decision-making are completed locally in a closed loop, without relying on external network connections, fundamentally eliminating response failures caused by network latency, congestion, or attacks.

[0030] The edge computing module's operating system adopts a real-time microkernel architecture, with core service processes divided into independent protection domains and their memory spaces isolated. The task scheduler prioritizes tasks based on risk level: emergency response tasks have priority 0 and the highest preemption right; evidence fusion and context assessment tasks have priority 2; primary risk assessment tasks have priority 4; routine data acquisition and preprocessing tasks have priority 6; and system maintenance and self-test tasks have priority 8. The scheduling cycle is 1 millisecond, ensuring that the end-to-end latency from detection to response for high-risk events does not exceed 500 milliseconds. The system supports dynamic voltage and frequency adjustment technology, automatically adjusting the processor's operating frequency according to load conditions, reducing it to 400 MHz during idle periods to lower power consumption.

[0031] All sensor components in the multimodal sensing array possess self-diagnostic capabilities. Pressure sensors periodically apply known pressure excitation to detect output deviations exceeding 5%; the spectral system performs dark-field and white-field calibration before each scan to verify that the signal-to-noise ratio deteriorates by more than 10 dB; the bioelectrode undergoes a polarization cleaning cycle weekly, comparing the change in response slope before and after cleaning; the fiber optic system performs a full-link optical power test daily to locate attenuation anomalies. Once any component's performance degradation or failure is detected, the system automatically resets its corresponding risk score weight to zero, masks it during the fusion process, and encodes the fault information as a specific error code, storing it in the maintenance queue for retrieval via the near-field communication interface when technicians arrive.

[0032] The system supports a dynamic knowledge base update mechanism. New spectral characteristics of waste liquid components, novel pathogen metabolic patterns, or structural failure cases can be injected into edge devices via secure offline loading. Update files must be verified for digital signature legitimacy by a hardware security chip before importation. The model update process is completed in a sandbox environment. The new version of the neural network is loaded in the background and infers in parallel with the currently running model for 1 hour. Only after the output consistency reaches over 99% can it be switched to the primary model. Old models are retained as copies until the next update, and one-click rollback is supported.

[0033] A peer-to-peer trust network is established among multiple adjacent edge smart nodes, employing a lightweight blockchain consensus mechanism to maintain a shared risk ledger. When a node enters an emergency state, its risk diffusion prediction model is encapsulated as a smart contract and broadcast to nodes within a 3-hop radius. The prediction model is constructed based on the pollutant diffusion dynamics equation, with inputs including leakage rate, ventilation conditions, and spatial topology, and outputting the probability of contamination in the vicinity. Nodes receiving the broadcast assess the associated risks based on their location and environmental parameters, and, if necessary, enter a state of alert and strengthen monitoring, forming a joint prevention and control mechanism. The trust network communication uses an elliptic curve digital signature algorithm for identity authentication to prevent forged alarms.

[0034] The container's identity information and system firmware version number are bound to a hardware security chip, which uses the national cryptographic algorithm SM2 to generate a unique device fingerprint. Integrity verification is performed at each startup. If firmware tampering or unauthorized device replacement is detected, the system enters a locked mode, allowing only basic sensing operations, disabling all response functions, and reporting the anomaly. All internal communication buses use AES-128 encryption for transmission. Critical control commands such as blocking and injection must carry a timestamp and digital signature; they can only be executed after verification by the receiving end, preventing replay attacks and unauthorized manipulation.

[0035] The entire system is powered by a combination of a top-integrated amorphous silicon photovoltaic film and a supercapacitor bank. The photovoltaic film has an area of ​​0.3 square meters and a peak output power of 15 watts, capable of generating electricity continuously under indoor lighting conditions. Electrical energy is rectified and regulated by the charging management circuit before being fed into the supercapacitor bank with a total capacity of 50 kJ. The system can maintain full operation for at least 72 hours without an external power source, and can operate continuously for 30 days in a timed inspection mode. The power management system dynamically adjusts the operating mode based on the remaining energy: when the capacitor voltage is below 20% of the rated value, unnecessary sensors are shut down, maintaining only pressure and spectral core monitoring; when it is below 10%, it enters a sleep state, waking up every 30 minutes to perform rapid testing.

[0036] This embodiment achieves simultaneous capture of three core risks—physical leakage, chemical exceedance, and biological contamination—by deploying a multimodal sensor array integrating pressure, spectral, biological, and deformation sensing capabilities at the waste liquid container end. This fundamentally overcomes the technical bottlenecks of single-point detection, which is susceptible to interference and has high false alarm and false negative rates. Multimodal data is processed at the edge via a three-level progressive fusion engine, sequentially completing intramodal feature extraction, cross-modal evidence integration, and global contextual understanding. This ensures that risk assessment relies not only on a single indicator exceeding its limits but also on a comprehensive judgment based on multi-source evidence, significantly improving the accuracy and robustness of risk identification. The system fully embeds the entire chain of processing logic from sensing to response within the edge computing module. All high-real-time control commands are generated and executed locally, with response latency controlled at the millisecond level. This eliminates reliance on centralized cloud processing and solves the fundamental defects of traditional architectures, such as large communication latency and paralysis upon network outages. The autonomous response mechanism is deeply coupled with the risk level state machine, triggering differentiated physical interventions for different risk levels. In particular, the introduction of intelligent blocking and precise neutralization functions enables the system to proactively curb the spread of risks, achieving a paradigm shift from "passive alarm" to "proactive prevention and control." The localized closed-loop processing strategy strictly limits the dissemination of sensitive data, uploading anonymized event summaries only when necessary, effectively protecting the operational privacy and data sovereignty of medical institutions. The system adopts a low-power design and renewable energy power supply, supporting long-term unattended operation. Combined with dynamic knowledge base updates and inter-node collaborative defense mechanisms, it forms an intelligent waste management terminal network with self-learning, self-adaptive, and self-organizing capabilities, providing solid technical support for building a safe, efficient, and reliable modern medical waste management system.

[0037] Existing technologies for medical waste management generally employ a centralized monitoring architecture, relying on regular manual inspections and fixed threshold alarm mechanisms. Such systems typically only utilize a single type of sensor, such as a level gauge or pH meter, failing to comprehensively capture complex risk scenarios. When slow leaks or trace amounts of toxic substances are introduced, the lack of multimodal cross-validation capabilities easily leads to missed detections; furthermore, environmental interference frequently triggers false alarms, resulting in a "boy who cried wolf" effect and diminishing the willingness of management to respond. Moreover, all data must be uploaded to a central server for processing, and communication delays cause response lags, making it difficult to meet the millisecond-level handling requirements of emergencies. More seriously, a network outage paralyzes the entire monitoring system, posing a significant security risk. The core difference of this solution lies in decentralizing the complete "perception-analysis-decision-execution" chain to each waste container terminal, building an autonomous intelligent agent at the edge. By integrating four-dimensional heterogeneous sensing modalities, the system can perform a three-dimensional scan of risks from physical, chemical, biological, and structural dimensions. Any anomaly in a single dimension can be cross-verified or ruled out by data from other dimensions, greatly enhancing the reliability of the judgment. The three-tiered, progressive fusion engine design enables the system not only to detect "what happened" but also to understand "the context," thus allowing for tiered responses that align with the actual threat level. A local closed-loop control mechanism ensures that even in a completely offline environment, the system can independently complete the entire process from risk identification to physical containment, truly achieving proactive defense capabilities without human intervention. This edge-to-edge collaborative, autonomously controllable technical architecture represents a fundamental shift in medical waste management from "digital monitoring" to "intelligent governance."

[0038] In the multi-level risk fusion engine, the improved Dempster-Shafer evidence theory framework adopted in the intermediate evidence fusion layer aims to solve the problem of reasonable fusion of multi-source heterogeneous evidence when conflicts exist. Traditional Dempster-Shafer theory may produce counterintuitive synthesis results when faced with highly contradictory evidence; for example, when two experts give completely opposite judgments, the synthesis may actually reinforce a certain erroneous conclusion. To overcome this deficiency, this system introduces a dynamic conflict allocation mechanism. Specifically, the Jaccard similarity between each evidence source is first calculated, defined as the ratio of the intersection to the union of the probabilities that commonly support the proposition in the basic probability allocation functions of the two pieces of evidence. Let the evidence... and The basic probability distributions are respectively and Then their Jaccard similarity The calculation is as follows:

[0039] Based on the average similarity of all evidence pairs, an overall consistency index is obtained, from which the dynamic conflict degree is derived. .when When the value is relatively high, it indicates that most of the evidence supports each other, and the standard D-S synthesis rule is adopted; when the value is relatively low, it indicates that there are significant contradictions among the evidence. At this time, the modified synthesis rule is enabled to reduce the comprehensive weight of the conflicting evidence and direct the unallocated trust degree to the uncertain set. This mechanism effectively suppresses the misfusion phenomenon caused by isolated abnormal sensors and improves the robustness of the system under sensor failures or extreme interference.

[0040] In the finite state machine model of the advanced situation understanding layer, the state transition logic not only depends on the risk score threshold but also introduces a time persistence constraint. For example, entering the warning state not only requires > 0.5, but also requires this condition to persist for more than 15 seconds to prevent false actions triggered by instantaneous noise; when upgrading from the warning state to the emergency state, if shows a continuous upward trend (the increase in three consecutive sampling periods is greater than 0.1), the decision delay can be shortened to 5 seconds, reflecting the rapid response to the accelerating deterioration situation. The state machine also includes a self-recovery path: when the system is in the attention or warning state, if all risk indicators fall back to the normal range within 2 minutes, it will automatically downgrade and resume the normal sampling frequency. This design not only ensures a timely response to real threats but also avoids over-reactions caused by short-term disturbances.

[0041] In the emergency neutralizing agent injection unit of the autonomous response execution mechanism, its dose calculation logic fully considers the reaction efficiency and safety margin. After the system calculates the theoretical injection volume and adds a 10% safety margin, the final instruction is 1.1×Q to cope with the situation of uneven mixing or incomplete reaction. After the injection is completed, the micro-plunger pump starts the stirring program, and the electromagnetic oscillator makes the container generate low-frequency vibration to promote the uniform dispersion of the neutralizing agent for 30 seconds. After the stirring ends, the spectral analysis subsystem immediately performs a quick scan to verify whether the characteristic peak of the target pollutant decays below the safety threshold. If not up to the standard, it triggers a secondary injection process, and at most 3 cycles are executed.

[0042] In addition to warning the on-site personnel, the local acoustic-optical warning module also has an environmental perception linkage function. The module is built with a microphone array to detect the surrounding sound pressure level. When the environmental noise exceeds 85 decibels, the buzzer volume is automatically increased to a maximum of 120 decibels; the built-in light sensor detects the environmental brightness. When the illuminance is lower than 50 lux, the LED flash brightness is automatically enhanced and the lighting cycle is extended. This design ensures that warning information can still be effectively conveyed in noisy or dim environments.

[0043] The system maintains a rolling health monitoring log in the edge computing module, recording the operating temperature, signal strength, self-test results, and cumulative operating time of each sensor. This log is used for predictive maintenance analysis. When a sensor's self-test deviation shows an upward trend for seven consecutive days, or its cumulative operating time approaches 90% of its design life, the system generates a preventative replacement recommendation and sends it to the maintenance terminal of a nearby workstation via short-range wireless communication, reminding technicians to arrange maintenance plans and avoid sudden failures affecting system availability.

[0044] The entire system's software architecture adopts a modular design, with each functional component running as an independent daemon process and communicating loosely through a message queue middleware. The message bus uses a publish / subscribe pattern, and topic naming follows a hierarchical structure such as "sensor / pressure / raw" and "fusion / mid_level / output". Inter-process communication data packets include message type, timestamp, source address, destination address, and payload fields. The payload is serialized using Protocol Buffers to ensure efficient transmission and cross-platform compatibility. Each process has an independent crash recovery mechanism, with a watchdog daemon monitoring its heartbeat signal. Once a process crash is detected, the instance is immediately restarted and restored from the most recent checkpoint.

[0045] Before leaving the factory, the system undergoes a rigorous multi-scenario calibration process. The pressure subsystem performs 100 leakage experiments under different pore sizes and pressure conditions on a simulated leakage platform to establish a mapping model between leakage rate and pressure change rate. The spectral subsystem uses a test set covering 20 types of real medical waste samples for cross-validation to ensure a classification accuracy of no less than 98%. The biological subsystem cultivates Escherichia coli and Staphylococcus aureus in a controlled environment and records their oxygen consumption characteristic curves at different growth stages to optimize the early warning threshold. The deformation subsystem conducts accelerated aging tests on the container prototype on a pressure fatigue testing machine, collecting strain evolution data from the initial microcrack to macroscopic rupture to train the structural damage diagnosis network. All calibration data is anonymized to form the initial training set, which is burned along with the system firmware.

[0046] In actual deployment, the system supports two installation modes: for new facilities, an integrated smart container is used, integrating all components into a standard 200-liter medical waste liquid tank; for retrofitting existing facilities, modular kits are provided, allowing the sensor array and edge computing box to be externally mounted on the existing container and fixed by magnetic attraction and snap-fit, with installation time not exceeding 30 minutes. Both modes ensure effective contact and reliable sealing between the sensor and the waste liquid body.

[0047] The system's time synchronization mechanism employs a lightweight implementation of the IEEE 1588 precision time protocol. The master clock is deployed at the regional aggregation node, distributing timestamps to each edge node via wired Ethernet. When there is no network connection, each node uses a local high-precision temperature-compensated crystal oscillator with a monthly drift rate of less than 1 second, ensuring time consistency during long-term offline operation. All event records use Coordinated Universal Time (UTC) timestamps to avoid time zone confusion.

[0048] Data storage employs a tiered management strategy. Raw sensor data is cached locally for no more than 24 hours, automatically overwritten upon expiration; preprocessed feature data is retained for 7 days; risk assessment results and event logs are permanently stored until manually deleted. Industrial-grade wide-temperature solid-state drives are used for storage, supporting operating temperatures from -40°C to 85°C and shock resistance up to 50g acceleration. Data writing utilizes a log-structured approach, first writing to the cache and then batch flushing to disk to reduce mechanical wear.

[0049] In terms of network security, the system implements a defense-in-depth strategy. At the hardware level, all external interfaces are equipped with electromagnetic interference filters to prevent physical layer injection attacks; at the firmware level, the bootloader verifies the kernel signature to prevent unauthorized code execution; at the system level, all unnecessary network ports are closed, and only encrypted communication channels are open; at the application level, critical operations require multi-factor authentication, such as remote firmware updates requiring both a hardware token and a biometric password. Firewall rules are uniformly issued by the security management center and support dynamic adjustments based on risk profiles.

[0050] The human-machine interface is implemented using near-field communication (NFC) technology. Maintenance personnel carrying a dedicated handheld terminal approach the device, automatically establishing a secure Bluetooth connection and accessing the device status panel to view real-time data, historical data curves, fault records, and maintenance suggestions. The interface uses a high-contrast black-and-white display, suitable for reading in bright light. All operation commands require secondary confirmation before being sent to prevent accidental touches. The connection is automatically disconnected and the temporary key is cleared after the interaction session ends.

[0051] The system supports swarm intelligence evolution under a federated learning framework. With institutional authorization, each node can complete incremental model training locally and then upload only the model gradient parameters to the central aggregation server. The server uses a secure aggregation algorithm to merge gradients, generate a globally updated model, and distribute it to all nodes. The entire process does not require sharing original data, protecting the data privacy of each medical institution. Updates occur monthly or are triggered as an emergency update when a new risk pattern is detected.

[0052] The environmental adaptability design takes extreme operating conditions into account. When the system starts up in a low-temperature environment of -20 degrees Celsius, the heating circuit is activated first to preheat the bioelectrode and drug reservoir, and sampling only begins after the temperature rises above 5 degrees Celsius. In high-humidity environments, the fiber optic demodulation unit activates a dehumidification mode to reduce the risk of internal condensation through Peltier elements. The outer shell material is made of UV-resistant engineering plastic with a service life of no less than 10 years.

[0053] A quality assurance system is implemented throughout the entire R&D and production process. Key components such as shape memory alloys, fiber optic gratings, and electrochemical electrodes are sourced from ISO 13485 certified suppliers. Each device undergoes a 168-hour continuous aging test before leaving the factory, simulating various risk scenarios to ensure functional integrity and stability. Test data generates electronic certificates of conformity, which are linked to the device's unique serial number for traceability.

[0054] The system described in this embodiment has completed a 6-month field pilot operation in the waste storage room of the infectious disease department of a tertiary-level Class A hospital. During this period, a total of 3 real leakage incidents, 2 instances of illegal dumping, and 1 case of abnormal microbial proliferation were recorded. The system successfully identified and triggered corresponding responses in all cases, with an average response time of 217 milliseconds and a false alarm rate of 0.3 times per thousand hours, significantly better than traditional monitoring systems. The pilot results verify the feasibility and superiority of this invention in a real medical environment.

Claims

1. An edge intelligence based medical waste liquid multi-modal risk perception and autonomous response system, characterized in that, include: A multimodal sensor array is used to simultaneously collect multi-source heterogeneous sensor data around a single medical waste container to characterize three core risks: physical leakage, chemical exceedance, and biological contamination. The edge computing module is used to receive the multi-source heterogeneous sensing data and run a multi-level risk fusion engine to generate risk level determination results. An autonomous response execution mechanism is used to implement differentiated physical intervention measures locally based on the risk level assessment results; The multi-level risk fusion engine includes: a primary risk discrimination layer, used to extract features from sensor data of each modality and output corresponding risk confidence scores; an intermediate evidence fusion layer, used to perform uncertainty reasoning on all risk confidence scores based on the improved Dempster-Shafer evidence theory framework, introducing a dynamic conflict allocation factor to suppress misfusion under highly contradictory evidence, and outputting a comprehensive risk assessment vector; and a high-level contextual understanding layer, used to determine the risk level of the system based on a finite state machine model, according to the comprehensive risk assessment vector and the multi-dimensional threshold combination logic of individual risk confidence scores.

2. The edge-intelligence based medical waste fluid multi-modal risk perception and autonomous response system as claimed in claim 1, wherein, The multimodal sensing array includes: The pressure micro-change detection subsystem is used to continuously monitor the hydrostatic pressure signal of the liquid at the bottom of the container to identify potential leaks or illegal dumping. The spectral feature analysis subsystem is used to scan the waste liquid to obtain the absorption lines of its characteristic functional groups, and combine it with a pre-set spectral database of typical components of medical waste liquid to analyze the waste liquid type and concentration distribution. A biological metabolic activity monitoring subsystem is used to measure the dissolved oxygen consumption gradient in waste liquid per unit time to quantify the risk of abnormal proliferation of pathogenic microorganisms; A three-dimensional deformation sensing subsystem is used to reconstruct the strain field distribution on the outer surface of the container to identify early signs of mechanical damage such as local bulging, crack propagation, or structural instability.

3. The edge-intelligence based medical waste fluid multi-modal risk perception and autonomous response system as claimed in claim 2, wherein, The primary risk assessment layer includes: A pressure anomaly detection network is used to process the time-series data of the pressure micro-change detection subsystem to output a pressure risk confidence score. A component deviation identification network is used to process the spectral vectors of the spectral feature analysis subsystem to output a component risk confidence score. A bioactivity early warning network is used to process the dissolved oxygen time series of the biological metabolic activity monitoring subsystem to output a biological risk confidence score. A structural damage diagnosis network is used to process the strain space distribution matrix of the three-dimensional deformation sensing subsystem to output a structural risk confidence score. Each network is a lightweight neural network model with limited parameters to adapt to the computing power constraints of the edge computing module.

4. The edge-intelligence based medical waste fluid multi-modal risk perception and autonomous response system as claimed in claim 3, wherein, The intermediate evidence fusion layer is used to convert each risk confidence score into a basic probability allocation function, calculate the Jaccard similarity between each evidence source to obtain the overall conflict degree, and then dynamically adjust the conflict redistribution ratio in the synthesis rules according to the overall conflict degree to complete the robust fusion of multi-source evidence.

5. The edge-intelligence based medical waste fluid multi-modal risk perception and autonomous response system as claimed in claim 4, wherein, The risk levels defined by the advanced contextual understanding layer include normal state, attention state, warning state, emergency state, and out-of-control state. The state transition conditions introduce time persistence constraints to prevent instantaneous noise from triggering erroneous actions, and include a self-recovery path to automatically downgrade when risk indicators fall back.

6. The edge intelligence-based multimodal risk perception and autonomous response system for medical waste liquid according to claim 5, characterized in that, The autonomous response execution mechanism includes: Intelligent sealing device, used to physically isolate the container discharge port after receiving a locking command; The emergency neutralizing agent injection unit is used to select the matching neutralizing agent type based on the spectral analysis results, and to perform precise injection after calculating the injection volume based on the pollutant concentration and waste liquid volume. The local audible and visual alarm module is used to issue audible and visual alarm signals and send alarm summaries to nearby workstations when activated.

7. The edge intelligence-based multimodal risk perception and autonomous response system for medical waste liquid according to claim 6, characterized in that, The emergency neutralizing agent injection unit is equipped with a dual-chamber storage bladder, which is loaded with acidic and alkaline neutralizing agents respectively. After injection, a stirring program is started to promote uniform dispersion of the neutralizing agent. After stirring, a spectral analysis subsystem is triggered to perform a rapid verification scan. If the standard is not met, a second injection process is executed.

8. The medical waste liquid multimodal risk perception and autonomous response system based on edge intelligence according to claim 1, characterized in that, The operating system of the edge computing module adopts a real-time microkernel architecture, and its task scheduling priority is strictly divided according to risk level to ensure that the end-to-end latency of the high-risk event handling path meets the millisecond-level response requirement.

9. The edge intelligence-based multimodal risk perception and autonomous response system for medical waste liquid according to claim 2, characterized in that, All sensor components of the multimodal sensing array have self-diagnostic capabilities, which are used to periodically perform stimulus-response tests and automatically adjust their weights in the intermediate evidence fusion layer and record fault information to the maintenance queue when performance degradation or failure is detected.

10. The edge-intelligence based medical waste fluid multi-modal risk perception and autonomous response system as claimed in claim 1, wherein, A point-to-point trust network is established between multiple adjacent systems. When any system enters an emergency state, it broadcasts a risk diffusion prediction model to surrounding nodes to trigger a joint defense mechanism. Furthermore, the system supports updating the dynamic knowledge base in the edge device through offline loading with secure authentication, which is used to incrementally train the neural network model of the primary risk discrimination layer.

Images