Multi-modal sensing medical waste intelligent supervision platform and early warning method

Abstract

The invention relates to the field of artificial intelligence and intelligent medical management, and discloses a medical waste intelligent supervision platform based on multi-modal perception and an early warning method. According to the method, image, weight, RFID and position information of medical waste in the processes of generation, temporary storage, transfer and processing are collected, data fusion and standardized preprocessing are carried out, a medical waste transfer behavior map is constructed, a map neural network is adopted to carry out anomaly detection on behavior nodes and output risk scores, and a medical waste transfer behavior map is constructed. And multi-level early warning is generated in combination with a rule engine and is displayed on a supervision platform in real time. The platform comprises a multi-modal data acquisition module, a behavior map construction and anomaly detection module and an early warning generation and display module, realizes full-process, refined and intelligent risk supervision and early warning response of medical wastes, and remarkably improves compliance and safety of medical waste management.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and smart medical management, and in particular to a multimodal sensing medical waste smart supervision platform and early warning method, which are suitable for intelligent identification, graph modeling and risk monitoring of medical waste treatment processes. Background Art

[0002] Due to its high infectiousness, bio-contamination, and chemical hazards, the full-process supervision of medical waste is directly related to public health safety and ecological stability. Currently, the management of medical waste relies primarily on traditional methods such as manual registration, scheduled inspections, and passive reporting. Although some regions have begun to introduce RFID tags and electronic record systems, these systems suffer from incomplete information collection, lack of data integration, and weak anomaly recognition capabilities. This makes it difficult to achieve real-time supervision and intelligent early warning of the waste flow process, and there are still situations where hidden dangers such as illegal dumping and mixed loading and transportation cannot be identified and intervened in a timely manner.

[0003] In recent years, the rapid development of artificial intelligence (AI), image recognition, the Internet of Things (IoT), and graph neural networks has provided new insights for medical waste regulation. Previous studies have attempted to incorporate image recognition for waste classification, identity tracking through RFID, and behavioral modeling to identify anomalous behavior. However, most of these technologies operate independently, lacking support for integrating heterogeneous data from multiple sources, such as images, weight, and location information. Furthermore, they fail to construct graph-based flow behavior maps to enable node-level risk perception and anomaly detection. Therefore, there is an urgent need for an intelligent regulatory system capable of multimodal perception, fusion analysis, and graph anomaly recognition to achieve closed-loop intelligent management and dynamic risk warnings for medical waste from source to destination. Summary of the Invention

[0004] In order to achieve the above-mentioned purpose of the invention, the present invention provides the following technical solution: a multimodal sensing medical waste intelligent supervision and early warning method, comprising the following steps:

[0005] S1. Collect multimodal data of medical waste at each stage of generation, temporary storage, transportation and treatment, including image data, weight data, RFID identification information, location data and time data;

[0006] S2. performing standardized preprocessing and fusing the multimodal data to generate structured medical waste behavior data;

[0007] S3. Build a medical waste flow map, use a graph neural network model to detect anomalies in node behavior, and output a risk score.

[0008] S4. Based on risk scoring and rule engine, generate graded warning information and display it in real time on the supervision platform.

[0009] Preferably, in step S1:

[0010] Image data is captured by AI vision acquisition units deployed at medical waste storage points and recycling stations, and the YOLOv5 model is used to identify waste types and operating behaviors;

[0011] Weight data is automatically collected by the scale and uploaded in real time through the sensor network;

[0012] RFID identification information is obtained by reading the RFID chips installed on medical waste bags and boxes;

[0013] The location data is collected by the GPS module combined with the Bluetooth base station positioning, and the time data is managed by the system with a unified timestamp.

[0014] Preferably, in step S2:

[0015] Standardized preprocessing and fusion, mapping the image recognition results into standard waste category codes through the target mapping table, and associating and matching them with RFID identity information;

[0016] The weight data is checked for consistency with the identification category. If any inconsistency or data drift exceeds the threshold, it is marked as a suspected anomaly.

[0017] Using timing alignment technology and taking RFID reading time as the reference axis, the multimodal data is dynamically aggregated through sliding windows to construct a fused behavior data frame.

[0018] Preferably, in step S3:

[0019] The medical waste flow graph has waste treatment nodes as vertices and flow events as edges;

[0020] A graph convolutional neural network model is used to aggregate graph nodes to determine whether their behavior is abnormal. The specific calculation method is as follows:

[0021]

[0022] in, is the node v in the l+1th layer, N (v) represents the neighbor set of node v, W (l) is the weight matrix of the lth layer, d v and d u are the degrees of nodes v and u respectively, and σ is the activation function.

[0023] By comparing the output node score with historical violation data, the risk level is output: low, medium, or high.

[0024] Preferably, in step S4:

[0025] Based on the risk score results, the risk level is divided into levels by comparing with the set threshold. If the score is greater than 0.8, it is high risk; between 0.5 and 0.8, it is medium risk; and less than 0.5, it is low risk.

[0026] The early warning information includes the location of the violation, the type of behavior, identification images, and suggested handling methods, and is pushed to the responsible personnel through the web-based supervision platform to achieve real-time monitoring and emergency response.

[0027] The present invention also provides a multimodal sensing medical waste intelligent supervision platform, which includes the following modules:

[0028] A multimodal data acquisition module, used to collect images, weight, RFID, and location information, and provide raw data support. The module supports connection with third-party medical institution information systems;

[0029] The behavioral graph construction and anomaly detection module is used to fuse pre-processed multimodal data, construct a medical waste flow graph, and perform anomaly detection and risk scoring based on a graph neural network model;

[0030] The warning generation and display module is used to receive risk scoring results, execute rule engine matching to generate multi-level warnings, and support visual display, historical query and manual review intervention.

[0031] The multimodal data acquisition module includes:

[0032] Image acquisition submodule, used to identify waste types and human operation behaviors through AI vision units;

[0033] The weight collection submodule is used to obtain medical waste weighing data in real time and bind it with the timestamp;

[0034] RFID identification submodule, used to uniquely identify the medical waste bag and record the time and location when it is read;

[0035] The positioning submodule is used to combine GPS and Bluetooth base station technology to track the location of medical waste during transportation.

[0036] The behavior graph construction and anomaly detection module includes:

[0037] The data fusion submodule is used to standardize the format, align the time sequence, and aggregate the behavior frames of multi-source data;

[0038] The graph construction submodule is used to construct a medical waste behavior graph based on waste flow paths and operation events;

[0039] The graph neural network processing submodule is used to perform multi-layer information aggregation on graph nodes, identify and learn historical violation behavior patterns, and implement node-level risk scoring.

[0040] The warning generation and display module includes:

[0041] The early warning strategy submodule is used to generate different levels of early warnings based on the set risk score thresholds and rule engine strategies;

[0042] The push interface submodule is used to notify relevant personnel through various methods such as web pages and SMS push;

[0043] The early warning visualization sub-module is used to graphically display abnormal behavior trajectories, risk levels, processing suggestions and operation receipts on the supervision platform.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] Improve the real-time and accuracy of medical waste supervision: This invention integrates multimodal sensing methods such as image recognition, weight perception, RFID identity recognition, and GPS positioning. Through dynamic sliding window fusion and time alignment technology, it realizes real-time perception and consistency verification of the entire process status of medical waste, greatly improving the timeliness and accuracy of supervision data.

[0046] Enhance the intelligent identification and early warning capabilities of abnormal behaviors: By constructing a medical waste flow map and introducing a graph convolutional neural network (GCN) model, it is possible to explore potential violation paths in complex node behaviors. Combining historical violation characteristics and risk scoring mechanisms, it can achieve intelligent identification and graded early warning of potential abnormal behaviors, preventing human violations and management loopholes.

[0047] Realize automatic push and closed-loop processing of regulatory information: The automatic push mechanism of early warning information of the present invention can accurately transmit the image, location and processing suggestions of the violation incident to the supervisor responsible, realizing the transformation from manual discovery to system push. Combined with the Web-side regulatory platform, it effectively builds an intelligent supervision and emergency linkage closed-loop system, improving supervision efficiency and response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic flow chart of the method steps provided for this application;

[0049] Figure 2 Schematic diagram of the system modules provided for this application. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0051] refer to Figure 1 The embodiment of the present invention provides a multimodal sensing medical waste intelligent supervision and early warning method, comprising the following steps:

[0052] Step 1: Collect multimodal data of medical waste at each stage of generation, temporary storage, transportation and treatment, including image data, weight data, RFID identification information, location data and time data.

[0053] In step one, the system was deployed within a tertiary hospital's medical waste management scenario, implementing multimodal data collection at key nodes such as waste generation, temporary storage, and transfer. First, an AI visual acquisition unit was installed in the temporary storage area and recycling site. Embedded with a YOLOv5 model, this unit captures and recognizes images of each waste placement operation in real time. It outputs the operator's action type (e.g., placement, handling, recycling) and waste type (e.g., infectious waste, pathological waste), and labels the corresponding confidence level and image recognition result. Second, a smart floor scale with data upload capability was installed at the temporary storage site. This scale automatically captures the current weight of each waste placement or removal, and uploads it to the monitoring platform in real time via a ZigBee wireless sensor network. The system automatically records and matches the weight with the recognized image. Regarding RFID collection, each medical waste bin or packaging bag is pre-installed with a unique RFID chip. Readers are located at the entrance of the temporary storage site and above the transfer cart. These readers automatically read the tag content within the recognition range and record the read time and tag ID. This information is then combined with the image recognition timestamp to form a waste tracking record. In addition, medical waste transfer vehicles are equipped with GPS modules, and low-power Bluetooth base stations are deployed at key locations throughout the hospital. By integrating signal strength and distance, the system enables real-time collection and dynamic tracking of waste location information, resolving positioning issues between floors or under building obstructions. All collected data is centrally transmitted to the platform for timestamp alignment, and the system server generates a structured data stream for subsequent behavioral mapping and risk identification analysis.

[0054] Step 2: Standardize and preprocess the multimodal data and fuse them to generate structured medical waste behavior data.

[0055] In step 2, after the system receives the raw multimodal data uploaded from the image recognition unit, the weighing scale, the RFID reader and the positioning module, it first performs standardized preprocessing on each type of data. For the image recognition results, the system converts the image output results into a unified waste category code based on the built-in target mapping table. For example, infectious waste is mapped to the code YW-01, and is matched one by one with the RFID identification code read at the same time to ensure that the waste category identified in the image is consistent with the information bound to the RFID. Subsequently, the system performs a consistency check on the weight data collected by the weighing scale, calculates the standard weight range under the identification category (for example, the standard weight of an infectious waste bag is 1.0kg-3.5kg). If the actual weight data exceeds the upper limit of the range, or deviates significantly from other data under the same label (the drift exceeds the set threshold such as ±30%), the system automatically marks the data as a suspected anomaly point and records the anomaly type and drift amplitude. At the same time, in order to solve the problem of time asynchrony of multi-source data, the system uses RFID reading time as the main axis of alignment, and through a sliding time window mechanism (such as setting a 5-second dynamic window), the collected image information, weight data, GPS, and Bluetooth location information are integrated and aggregated to form a "behavioral data frame" with continuous time and complete data. Each frame contains a unified timestamp, image recognition label, standard category code, RFID identification, weight value, location information and other fields, which are used for subsequent anomaly detection and intelligent early warning processes.

[0056] Step 3: Build a medical waste flow map, use a graph neural network model to detect anomalies in node behavior, and output a risk score.

[0057] In step three, the system first constructs a medical waste flow map based on the complete flow path of waste treatment. The map uses each medical waste treatment node (such as temporary storage point A, transfer station B, incineration plant C) as the vertex of the graph, and each flow event recorded by the waste carrying RFID tag (such as transportation from temporary storage point A to transfer station B) as an edge. By performing time series combing and geographical tracking on multimodal data, graph structure data is automatically generated. Based on this map, the system introduces a graph convolutional neural network model to learn the contextual behavior characteristics of the nodes. During the training process of the graph neural network, the system aggregates and updates the representation of each node v at the l+1 layer. The calculation formula is as follows:

[0058]

[0059] in, is the node v in the l+1th layer, N (v) represents the neighbor set of node v, W (l) is the weight matrix of the lth layer, d v and d uwhere ∠ is the degree of nodes v and u, respectively, and σ is the activation function. The system calculates a final embedding vector for each node through multi-layer graph convolution. This embedding representation incorporates the behavioral characteristics of the node itself and its upstream and downstream neighbors. To further determine whether abnormal behavior exists, the system introduces a risk comparison mechanism trained on historical violation samples. This mechanism matches the feature representation of each node's current output with historically known high-risk, medium-risk, and low-risk behavior vectors. Based on the matching results, the risk level of the node is output. Ultimately, the system identifies and categorizes the risk of each waste treatment node's behavior within the graph structure, providing a highly accurate basis for subsequent early warning decisions.

[0060] Step 4: Based on the risk score and rule engine, generate graded warning information and display it in real time on the supervision platform.

[0061] In step four, after completing the graph neural network calculation and obtaining the risk score for each node, the system determines the risk level based on the set risk classification threshold. Specifically, if a node's risk score is higher than 0.8, the system automatically classifies it as high risk and immediately triggers the first-level alert process; if the score is between 0.5 and 0.8, it is marked as medium risk and enters the second-level warning state; if the score is lower than 0.5, it is considered low risk and only routine records and backend archiving are performed. For nodes determined to be medium or high risk, the system will automatically generate a structured warning information package, which includes: the location of the violation (geographic coordinates and site number determined by combining GPS and Bluetooth base station information), the type of violation (such as unauthorized transfer, weighing discrepancy, abnormal detention, etc.), a screenshot of the image recognition results (an overlay of the original image provided by the AI visual acquisition unit and the YOLOv5 recognition results), and recommended handling methods (such as on-site verification, suspension of transfer, initiation of accountability procedures, etc.). This information is pushed to relevant responsible personnel, including site administrators, transfer supervisors, and environmental protection inspection units, in a graphic and text format via a dedicated medical waste supervision web platform. The system also records the response time and processing status to ensure that each warning information has traceability and closed-loop disposal capabilities, truly realizing real-time monitoring, intelligent classification, timely response and closed-loop control of abnormal behaviors in the entire medical waste process.

[0062] refer to Figure 2 The embodiment of the present invention also provides a multimodal perception medical waste intelligent supervision platform, including the following modules:

[0063] A multimodal data acquisition module, used to collect images, weight, RFID, and location information, and provide raw data support. The module supports connection with third-party medical institution information systems;

[0064] The behavioral graph construction and anomaly detection module is used to fuse pre-processed multimodal data, construct a medical waste flow graph, and perform anomaly detection and risk scoring based on a graph neural network model;

[0065] The warning generation and display module is used to receive risk scoring results, execute rule engine matching to generate multi-level warnings, and support visual display, historical query and manual review intervention.

[0066] The multimodal data acquisition module includes:

[0067] Image acquisition submodule, used to identify waste types and human operation behaviors through AI vision units;

[0068] The weight collection submodule is used to obtain medical waste weighing data in real time and bind it with the timestamp;

[0069] RFID identification submodule, used to uniquely identify the medical waste bag and record the time and location when it is read;

[0070] The positioning submodule is used to combine GPS and Bluetooth base station technology to track the location of medical waste during transportation.

[0071] The behavior graph construction and anomaly detection module includes:

[0072] The data fusion submodule is used to standardize the format, align the time sequence, and aggregate the behavior frames of multi-source data;

[0073] The graph construction submodule is used to construct a medical waste behavior graph based on waste flow paths and operation events;

[0074] The graph neural network processing submodule is used to perform multi-layer information aggregation on graph nodes, identify and learn historical violation behavior patterns, and implement node-level risk scoring.

[0075] The warning generation and display module includes:

[0076] The early warning strategy submodule is used to generate different levels of early warnings based on the set risk score thresholds and rule engine strategies;

[0077] The push interface submodule is used to notify relevant personnel through various methods such as web pages and SMS push;

[0078] The early warning visualization sub-module is used to graphically display abnormal behavior trajectories, risk levels, processing suggestions and operation receipts on the supervision platform.

[0079] Obviously, the embodiments described above are only some embodiments of the present invention, rather than all embodiments. The preferred embodiments of the present invention are given in the accompanying drawings, but they do not limit the patent scope of the present invention. The present invention can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive. Although the present invention has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present invention specification and drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present invention.

Claims

1. A multimodal sensing medical waste intelligent supervision and early warning method, characterized by: The following steps are involved: S1. Collect multimodal data of medical waste at each stage of generation, temporary storage, transportation and treatment, including image data, weight data, RFID identification information, location data and time data; S2. performing standardized preprocessing and fusing the multimodal data to generate structured medical waste behavior data; S3. Build a medical waste flow map, use a graph neural network model to detect anomalies in node behavior, and output a risk score. S4. Based on risk scoring and rule engine, generate graded warning information and display it in real time on the supervision platform.

2. The multimodal sensing medical waste intelligent supervision and early warning method according to claim 1 is characterized in that: In the step S1: Image data is captured by AI vision acquisition units deployed at medical waste storage points and recycling stations, and the YOLOv5 model is used to identify waste types and operating behaviors; Weight data is automatically collected by the scale and uploaded in real time through the sensor network; RFID identification information is obtained by reading the RFID chips installed on medical waste bags and boxes; The location data is collected by the GPS module combined with the Bluetooth base station positioning, and the time data is managed by the system with a unified timestamp.

3. The multimodal sensing medical waste intelligent supervision and early warning method according to claim 1 is characterized in that: In the step S2: Standardized preprocessing and fusion, mapping the image recognition results into standard waste category codes through the target mapping table, and associating and matching them with RFID identity information; The weight data is checked for consistency with the identification category. If any inconsistency or data drift exceeds the threshold, it is marked as a suspected anomaly. Using timing alignment technology and taking RFID reading time as the reference axis, the multimodal data is dynamically aggregated through sliding windows to construct a fused behavior data frame.

4. The multimodal sensing medical waste intelligent supervision and early warning method according to claim 1 is characterized in that: In the step S3: The medical waste flow graph has waste treatment nodes as vertices and flow events as edges; A graph convolutional neural network model is used to aggregate graph nodes to determine whether their behavior is abnormal. The specific calculation method is as follows: in, is the node v in the l+1th layer, N (v) represents the neighbor set of node v, W (l) is the weight matrix of the lth layer, d v and d u are the degrees of nodes v and u respectively, and σ is the activation function. By comparing the output node score with historical violation data, the risk level is output: low, medium, or high.

5. The multimodal sensing medical waste intelligent supervision and early warning method according to claim 1 is characterized in that: In the step S4: Based on the risk score results, the risk level is divided into levels by comparing with the set threshold. If the score is greater than 0.8, it is high risk; between 0.5 and 0.8, it is medium risk; and less than 0.5, it is low risk. The warning information includes the location of the violation, the type of behavior, identification images, and suggested handling methods, and is pushed to the responsible personnel through the web-based supervision platform to achieve real-time monitoring and emergency response.

6. A multimodal sensing medical waste intelligent supervision platform, characterized by: Includes the following modules: A multimodal data acquisition module, used to collect images, weight, RFID, and location information, and provide raw data support. The module supports connection with third-party medical institution information systems; The behavioral graph construction and anomaly detection module is used to fuse pre-processed multimodal data, construct a medical waste flow graph, and perform anomaly detection and risk scoring based on a graph neural network model; The warning generation and display module is used to receive risk scoring results, execute rule engine matching to generate multi-level warnings, and support visual display, historical query and manual review intervention.

7. The multimodal sensing medical waste intelligent supervision platform according to claim 6 is characterized in that: The multimodal data acquisition module includes: Image acquisition submodule, used to identify waste types and human operation behaviors through AI vision units; The weight collection submodule is used to obtain medical waste weighing data in real time and bind it with the timestamp; RFID identification submodule, used to uniquely identify the medical waste bag and record the time and location when it is read; The positioning submodule is used to combine GPS and Bluetooth base station technology to track the location of medical waste during transportation.

8. The multimodal sensing medical waste intelligent supervision platform according to claim 6 is characterized in that: The behavior graph construction and anomaly detection module includes: The data fusion submodule is used to standardize the format, align the time sequence, and aggregate the behavior frames of multi-source data; The graph construction submodule is used to construct a medical waste behavior graph based on waste flow paths and operation events; The graph neural network processing submodule is used to perform multi-layer information aggregation on graph nodes, identify and learn historical violation behavior patterns, and implement node-level risk scoring.

9. The multimodal sensing medical waste intelligent supervision platform according to claim 6 is characterized in that: The warning generation and display module includes: The early warning strategy submodule is used to generate different levels of early warnings based on the set risk score thresholds and rule engine strategies; The push interface submodule is used to notify relevant personnel through various methods such as web pages and SMS push; The early warning visualization sub-module is used to graphically display abnormal behavior trajectories, risk levels, processing suggestions and operation receipts on the supervision platform.

Images