Central sterile supply department surgical instrument ai management system

Abstract

The present application relates to the technical field of intelligent management of medical devices, in particular to an AI management system for surgical devices in a central sterile supply department, comprising: device identity management, visual tracking, automatic operation and updating, abnormality detection and disposal modules. The system establishes a device package and a unique device identity and binds it with a radio frequency identification tag, captures visual data at key nodes in the circulation through a distributed image acquisition terminal, compares and outputs device identification results, node positions and surface preliminary state information through a pre-trained model. The state information is analyzed for differences with the standard atlas to generate an abnormality report, and manual review, maintenance isolation or automatic loss reporting processes are triggered according to the severity level. The system realizes synchronous acquisition of device circulation information, standardized determination of state and automatic disposal of abnormality classification, reduces manual determination errors and information acquisition lag. The invention of the system is an innovation in AI management of surgical devices in a central sterile supply department and an important embodiment of intelligent management of medical institutions.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for medical devices, and in particular to an AI management system for surgical instruments in a sterilization supply center. Background Technology

[0002] Currently, the management of surgical instruments in the sterilization supply center mostly uses radio frequency identification (RFID) tags to bind instrument identities. The equipment department's requisition ledger records the instrument model, serial number, and activation time. The identification is completed by reading the tag information through an RFID reader / writer. The operation of instrument circulation nodes relies on manual execution. The surface condition, structural integrity, and contamination of instruments are judged by manual visual inspection. Information on changes in instrument status is entered manually or passively recorded by RFID devices. The verification and reporting of abnormal instruments are all initiated manually and reviewed at each level.

[0003] Radio frequency identification (RFID) technology can only identify the device's identity; it cannot simultaneously acquire the device's location at a working node and its surface condition. Manually collected status information is subject to subjective judgment bias, and information collection at key nodes in device circulation is delayed. Manual inspection struggles to accurately identify minute structural damage, hidden functional defects, and residual contamination. There is no unified standardized basis for judging abnormal conditions, and the device abnormality handling process lacks a tiered trigger mechanism. The initiation of manual review and damage reporting processes is highly arbitrary, and the automation and standardization of device lifecycle status recording and circulation operations are low.

[0004] To address the issues that single-identity recognition cannot simultaneously acquire individual instrument identification, node location, and preliminary surface condition, as well as the lack of standardized image comparison for instrument surface condition and the inability to automatically trigger corresponding treatment procedures for abnormal levels, an AI-based management system for surgical instruments adapted to the sterilization supply center is constructed. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI management system for surgical instruments in a sterilization supply center.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an AI management system for surgical instruments in a sterilization supply center, comprising:

[0007] The instrument identification management module, based on the surgical instrument requisition ledger of the Equipment Department, establishes a unique identification identifier for each instrument, which includes the instrument model code, factory serial number, and initial activation time. The unique identification identifier is then bound to the physical surgical instrument via an embedded radio frequency identification tag.

[0008] The visual tracking module uses a distributed image acquisition terminal to capture visual data of surgical instruments at key nodes in the flow of surgical instruments. The captured visual data of the instruments is compared with a pre-trained instrument feature recognition model. The comparison process outputs the individual instrument recognition result, the current working node position, and the preliminary state information of the instrument surface.

[0009] The automatic operation and update module calls the automated operation unit based on the robotic arm, and executes the preset physical operation actions at the corresponding key nodes according to the individual identification results of the instruments and the current working node position. It scans the embedded RFID tags in real time through the RFID reader and writer device, and records and updates the status change information of each instrument bound with an RFID tag during the circulation process.

[0010] The anomaly detection and handling module performs a difference analysis between the preliminary surface condition information of the instrument and the preset instrument integrity standard map, generates an anomaly detection report on instrument structural damage, functional defects or contamination residue, and triggers a manual review process, a maintenance isolation process or an automatic loss reporting process according to the severity level of the anomaly detection report.

[0011] The intelligent transport robot module is used for point-to-point transport of instrument packs, cleaning racks and sterile items between the operating room, pretreatment area, cleaning and decontamination area, inspection and packaging area, sterilization area and sterile storage area.

[0012] The equipment interface module is used for data exchange with pretreatment cleaning machines, automatic cleaning and disinfection machines, sterilizers, and automatic packaging machines;

[0013] The central control platform integrates radio frequency identification information, image recognition information, equipment status information, and logistics task information to uniformly schedule the instrument retrieval, cleaning, inspection, packaging, sterilization, storage, distribution, and damage reporting, thereby achieving closed-loop management of the entire surgical instrument process.

[0014] As a further aspect of the present invention, the method of capturing instrument visual data at key nodes in the flow of surgical instruments using a distributed image acquisition terminal includes:

[0015] The key nodes include at least the equipment department requisition node, the operating room usage node, the instrument recycling and receiving point, the cleaning and unloading point, the packaging operation table, and the sterilization loading point.

[0016] A first set of image acquisition terminals is deployed above the conveyor belt at the instrument recycling receiving point. When a recycling container containing used instruments passes by, the first set of image acquisition terminals is triggered to capture a global top-down image containing the stacked state of all instruments in the container.

[0017] A second set of image acquisition terminals is deployed at the outlet of the cleaning machine at the unloading point. When the cleaning rack is removed from the cleaning machine, the second set of image acquisition terminals is triggered to scan each instrument on each shelf of the cleaning rack to obtain a detailed image of the surface of each instrument after cleaning.

[0018] A third set of image acquisition terminals is deployed above the workstation of the packaging operation table. When the operator or the automated operation unit of the robotic arm performs the packaging operation, the operation area is continuously photographed, capturing multiple angle images of each instrument that is picked up, inspected and put into the packaging material.

[0019] A fourth set of image acquisition terminals is deployed at the sterilization vehicle parking position at the sterilization loading point. During the process of loading instruments into the sterilization vehicle, images of each basket or tray entering the sterilization vehicle are recorded to generate a loading process sequence image.

[0020] As a further aspect of the present invention, the comparison process between the captured instrument visual data and the pre-trained instrument feature recognition model includes:

[0021] The pre-trained instrument feature recognition model is obtained by deep learning training based on multi-angle instrument image samples;

[0022] The global top-view image is processed by image segmentation to separate the contour regions of individual instruments from the stacked state, and visual feature fragments are extracted for each contour region.

[0023] The surface detail image is subjected to feature enhancement and noise filtering to extract microscopic visual features that reflect the cleanliness, smoothness, and presence of water stains on the instrument.

[0024] The images from multiple angles are subjected to three-dimensional structural reconstruction processing to synthesize a three-dimensional appearance model of the instrument, and three-dimensional feature data reflecting the geometric shape and structural integrity of the instrument are extracted from the three-dimensional appearance model.

[0025] The visual feature fragments, the microscopic visual features, and the three-dimensional feature data are input in parallel into the pre-trained instrument feature recognition model;

[0026] The pre-trained instrument feature recognition model includes multiple parallel feature analysis sub-networks, which process the input visual feature fragments, microscopic visual features, and three-dimensional feature data respectively. The outputs of the feature analysis sub-networks are integrated in the model fusion layer, and finally output the individual instrument recognition result including the instrument model code, individual instrument serial number, and the standard instrument package type to which it belongs. At the same time, it outputs the current working node position of the instrument, whether it is being recycled, cleaned, packaged, or sterilized and loaded, as well as the preliminary surface state information of the instrument, including the overall cleanliness score and obvious stain markings.

[0027] As a further aspect of the present invention, the invocation of the robotic arm-based automated operation unit, based on the individual instrument identification result and the current working node position, executes a preset physical operation action at the corresponding key node, including:

[0028] The physical operations include sorting instruments from the recycling container, loading instruments onto the washing rack, arranging and combining instruments on the packaging table, and unloading instruments from the sterilization cart.

[0029] When the current working node is a device recycling and receiving point, and the device identification result indicates that it is a specific type of device, the robotic arm automated operation unit located in the recycling area is controlled to grab the identified device from the recycling container according to the device model code in the device identification result, and place it in the pre-processing classification basket corresponding to the device model code.

[0030] When the current working node is a cleaning and unloading point, and the cleanliness score in the preliminary status information of the instrument surface is higher than a preset threshold, the robotic arm automated operation unit located at the cleaning and unloading point controls the instrument to be transferred from the cleaning rack to the packaging conveyor belt or temporary storage vehicle corresponding to the individual instrument identification result according to the individual instrument identification result.

[0031] When the current working node position is the packaging operation table and the system receives the packaging task instruction, the robotic arm automated operation unit located at the packaging operation table controls the corresponding instruments to be picked up from the conveyor belt or temporary storage vehicle in sequence according to the instrument package list required by the packaging task instruction, and arranges the instruments on the packaging material according to the preset instrument package placement template.

[0032] When the current working node position is the sterilization loading point, the robotic arm automated operation unit located at the sterilization loading point controls the packaged instrument packs to be picked up and stacked into the sterilization cart according to the sterilization batch plan and the preset loading mode.

[0033] As a further aspect of the present invention, the step of scanning the embedded RFID tag in real time using an RFID reader / writer to record and update the status change information of each instrument bound with an RFID tag during its circulation process includes:

[0034] The status change information includes the time of entry into the recycling area, the cleaning batch number, the sterilization cycle parameters, the storage shelf location, and the target department for distribution.

[0035] When the robotic arm automated operation unit places the device into the pre-processing sorting basket, it triggers the radio frequency identification reading and writing device set above the pre-processing sorting basket to scan the embedded radio frequency identification tags of all the devices in the basket, records the timestamp of the device entering the pre-processing area, and updates the status of the devices in the system to "recovered and waiting to be cleaned".

[0036] After the robotic arm automated operation unit transfers the equipment from the cleaning rack, it triggers the radio frequency identification (RFID) reader / writer located at the conveyor belt entrance to scan the embedded RFID tags of the passing equipment, record the corresponding cleaning machine number and cleaning cycle number, and update the status of the equipment in the system to "cleaned and awaiting inspection".

[0037] After the robotic arm automated operation unit arranges the instruments on the packaging material, it triggers the radio frequency identification (RFID) reader / writer set on the packaging platform to scan the embedded RFID tags of all instruments in the instrument package, binds the unique traceability barcode of the instrument package, records the packaging personnel and packaging time, and updates the status of the relevant instruments in the system to "packaged and awaiting sterilization".

[0038] After the robotic arm automated operation unit places the instrument pack code into the sterilization cart, it triggers the radio frequency identification (RFID) reader / writer device installed in the sterilization cart to scan the embedded RFID tags of all instrument packs in the sterilization cart, associates the sterilization batch number, sterilization program parameters and planned sterilization time, and updates the status of the relevant instruments in the system to "loaded and ready for sterilization".

[0039] As a further aspect of the present invention, the preliminary surface condition information of the instrument is compared with a preset instrument integrity standard map to generate an abnormal detection report on instrument structural damage, functional defects, or residual contamination, including:

[0040] From the preliminary surface state information of the instrument, extract edge feature data reflecting the overall contour of the instrument, key point angle data reflecting joint flexibility, and pixel grayscale distribution data reflecting surface cleanliness.

[0041] From the preset instrument integrity standard map, retrieve the standard edge features, standard joint angle range, and standard cleanliness grayscale threshold corresponding to the instrument model that matches the individual instrument identification result;

[0042] The edge feature data is compared pixel by pixel with the standard edge features to calculate the contour overlap. When the contour overlap is lower than the structural integrity threshold, it is determined that there is damage to the instrument structure, and the location and degree of deviation of the damage are recorded.

[0043] The key point angle data is compared with the standard joint angle range. If the angle data exceeds the standard range, it is determined that there is a functional defect in the device, and the defective joint and the angle of limited movement are recorded.

[0044] The pixel grayscale distribution data is compared with the standard cleanliness grayscale threshold. The area and distribution of abnormal pixel regions exceeding the threshold are counted. If the area of ​​abnormal regions exceeds the contamination threshold, it is determined that there is contamination residue, and the contamination area and severity level are recorded.

[0045] Based on the combined results of the assessment of device structural damage, device functional defects, and contamination residues, a structured anomaly detection report is generated. The anomaly detection report includes the anomaly type, anomaly location, a quantitative value of the anomaly severity, and the corresponding individual device identification result.

[0046] As a further aspect of the present invention, a manual review process or an automatic damage reporting process is triggered based on the severity level of the anomaly detection report, including:

[0047] The rules for determining the severity of anomalies are set. These rules are based on the comparison between the quantitative value of the anomaly degree in the anomaly detection report and a preset threshold, and divide the anomalies into three levels: mild, moderate and severe.

[0048] When the anomaly detection report is determined to be slightly abnormal, the system pushes the anomaly detection report and the device image to the review workstation interface, prompting the staff to conduct manual visual review and confirmation, and updates the device status in the system based on the manual review results;

[0049] When the anomaly detection report is determined to be a moderate anomaly, the system controls the robotic arm automated operation unit to sort the instruments into an independent isolation container for repair, and sends a repair request containing the anomaly detection report to the maintenance management system, while updating the instrument status to "pending repair".

[0050] When the abnormal detection report is determined to be a severe abnormality, or when the same device accumulates a predetermined number of moderate abnormalities, the system automatically generates a damage report in the device asset database. The damage report is associated with the complete circulation history of the device and all previous abnormal detection reports, and the device status is updated to "pending scrapping". At the same time, the system controls the robotic arm automated operation unit to transfer the device to the scrap temporary storage area.

[0051] As a further aspect of the present invention, it also includes:

[0052] The lifecycle prediction module periodically extracts the complete circulation record of each instrument from the system database. The complete circulation record includes the cumulative number of cleanings, the cumulative number of sterilizations, the contents of each abnormal detection report, and the maintenance history.

[0053] Data cleaning and feature extraction are performed on the complete circulation record to construct a life cycle feature vector that reflects the intensity of instrument use, maintenance frequency, and performance degradation trend;

[0054] The life cycle feature vector is input into a pre-trained device life prediction model, which is a regression model trained based on historical scrapped device data.

[0055] The pre-trained device life prediction model outputs the predicted remaining safe service life or the predicted time of next failure for each device.

[0056] By comparing the predicted remaining safe service life with the preset warning period, the system generates preventative replacement recommendations in advance for devices that are about to reach the end of their service life.

[0057] As a further aspect of the present invention, the complete circulation record is cleaned and its features are extracted to construct a lifecycle feature vector reflecting the intensity of instrument use, maintenance frequency, and performance degradation trend, including:

[0058] The average treatment frequency per unit time is calculated from the cumulative number of cleaning times and the cumulative number of sterilization times, and is used as a usage intensity characteristic.

[0059] From the various anomaly detection reports, the frequency of each anomaly, the changing trends of the time interval between occurrences, and the quantitative value of the anomaly severity of the most recent anomaly are statistically analyzed as performance degradation trend characteristics.

[0060] From the maintenance history, the number of maintenance operations, the parts repaired, and the usage time after maintenance are extracted as characteristics of maintenance frequency and repair effectiveness.

[0061] The usage intensity characteristics, the performance degradation trend characteristics, and the maintenance frequency and repair effect characteristics are arranged in chronological order and normalized to form a time-series feature sequence.

[0062] Sliding window feature statistics are performed on the time-series feature sequence to calculate the mean, variance, maximum value and slope of change within the window, and the statistics are combined into a fixed-dimensional life cycle feature vector.

[0063] As a further aspect of the present invention, it also includes:

[0064] The dynamic assembly module receives instrument pack requisition requests from the operating room in real time. The requisition requests include the required instrument pack type, quantity, and planned usage time.

[0065] Check the current inventory status of the sterile storage area, including the inventory quantity, storage location, and earliest sterilization date of each type of instrument pack;

[0066] If the inventory is sufficient, a picking list and optimal picking path for the medical kits requested will be generated based on the first-in, first-out principle and storage location, guiding staff or transport robots to carry out the distribution.

[0067] If there is insufficient stock of a certain type of instrument pack, a dynamic repacking task is triggered. The system queries the available stock of scattered sterile instruments and their locations. The stock of scattered sterile instruments comes from spare instruments or instruments temporarily transferred from other instrument packs.

[0068] Based on the individual device identification results and the standard list of device packages, the system automatically calculates the optimal combination of devices required to meet the application request, generates a packaging task instruction, guides staff or robotic arm automated operation units to perform the packaging operation, and updates the inventory status of relevant device packages and scattered devices after the packaging is completed.

[0069] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0070] By relying on distributed image acquisition terminals to capture visual data at key nodes in the flow of surgical instruments, and comparing the visual data with a pre-trained instrument feature recognition model, the system can directly output the individual instrument identification results, the current working node position, and the preliminary surface status information of the instrument. This overcomes the limitations of relying solely on radio frequency identification for single-identity recognition. The three types of information are generated synchronously without manual separation and collection. Individual instrument identification is not affected by manual verification. The current working node position is automatically matched with the corresponding flow link. The preliminary surface status of the instrument forms standardized data output. The status acquisition process does not require manual intervention. Information generation is synchronously matched with the flow node, eliminating the time delay and information deviation caused by manual collection.

[0071] By comparing the preliminary surface condition information of medical devices with a preset standard medical device integrity map, a difference analysis can be performed to generate anomaly detection reports covering structural damage, functional defects, or residual contamination. Based on the severity level of the anomaly detection report, a manual review process or an automatic loss reporting process is triggered. This eliminates the need for manual visual comparison to determine the condition of medical devices, and uses the standard map as a benchmark to achieve quantitative difference analysis. Anomaly types can be accurately classified and distinguished, and the severity of anomalies is automatically graded. Different severity levels correspond to specific handling processes. The manual review process is only initiated for low- to medium-level anomalies, while high-level anomalies directly trigger the automatic loss reporting process. The anomaly handling process does not require manual initiation. A unified standard for judging the condition of medical devices is established, and a precise correspondence is formed between handling actions and anomaly levels, reducing handling errors caused by human omissions and misjudgments.

[0072] By setting up intelligent transport robot modules, equipment interface modules, and a central control platform, the system can coordinate and connect all processing nodes of instruments from equipment department requisition, operating room use, CSSD recycling, cleaning, inspection, packaging, sterilization, cooling and storage to aseptic dispensing, realizing closed-loop management of information flow, logistics and quality control flow, reducing human contact and cross-contamination risks, and improving equipment utilization, traceability integrity and clinical supply continuity. Attached Figure Description

[0073] Figure 1 This is a timing diagram of the AI ​​management system for surgical instruments in a sterilization supply center as described in this invention;

[0074] Figure 2 A flowchart for the comparison and processing of instrument visual data and instrument feature recognition models;

[0075] Figure 3 Box plot of instrument joint angle detection distribution;

[0076] Figure 4 The relationship between repair effectiveness and usage time / frequency of SC-10-1208;

[0077] Figure 5 This is a trend analysis chart of the entire life cycle of surgical instruments. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0079] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0080] See Figure 1This invention provides an AI management system for surgical instruments in a sterilization supply center. The system includes: an instrument identification management module, a visual tracking module, an automatic operation and update module, an anomaly detection and handling module, an intelligent transport robot module, an equipment interface module, and a central control platform. The instrument identification management module, based on the equipment department's surgical instrument requisition ledger, creates a unique identifier for each surgical instrument or instrument package, including the instrument model code, factory serial number, and initial activation time. An RFID tag associated with this identifier is then affixed to the physical instrument or instrument package via pasting, embedding, or binding. The visual tracking module uses distributed image acquisition terminals deployed at key nodes in the instrument flow channel to capture visual data of the instruments. This data is transmitted to a pre-trained instrument feature recognition model for analysis and comparison, thereby obtaining individual instrument identification results, the current working node position of the instrument, and preliminary assessment information on its surface condition. The automated operation and update module dispatches robotic arm-based automated operation units. Based on the individual instrument identification results and location information provided by the vision tracking module, it executes preset physical operations such as sorting, loading, and arranging at corresponding key nodes. This module also scans instrument or instrument pack tags using RFID readers deployed at operation points, recording and updating the status changes of instruments in real time during recycling, cleaning, packaging, sterilization, and storage. The intelligent transport robot module is used to complete point-to-point transportation of contaminated instrument packs, cleaning racks, and sterile instrument packs between different functional areas. The equipment interface module is used to connect with pretreatment cleaning machines, automatic cleaning and disinfection machines, sterilizers, and automatic packaging machines to obtain equipment operating status and issue tasks. The central control platform is used to uniformly schedule robotic arms, transport robots, and various processing equipment based on RFID identification results, image recognition results, equipment status, and logistics task information.

[0081] In one embodiment of the present invention, within the sterilization supply center, key nodes for instrument circulation include at least an equipment department requisition node, an operating room usage node, an instrument recycling and receiving point, a cleaning and unloading point, a packaging workbench, and a sterilization loading point. At the equipment department requisition node, the system establishes instrument or instrument package identification information based on the instrument requisition ledger and records the issuance time, destination, and recipient. At the operating room usage node, the system monitors instrument inventory status and abnormal conditions through radio frequency identification, IoT sensors, and / or image recognition. When inventory falls below a preset safety threshold or damaged instruments are detected, a replenishment alert is automatically generated and pushed to the equipment department management platform. A first set of image acquisition terminals is deployed above the conveyor belt at the instrument recycling and receiving point. When a recycling container containing used surgical instruments passes this point, the first set of image acquisition terminals is triggered to capture a global top-down view image covering the stacked state of all instruments within the container. At the cleaning and unloading point, a second set of image acquisition terminals is deployed at the outlet of the cleaning machine. When the cleaning rack carrying the instruments is removed from the cleaning machine, the second set of image acquisition terminals is triggered to scan each instrument on each shelf of the cleaning rack. At the packaging operation table, a third set of image acquisition terminals is deployed above the workstation. At the sterilization loading point, a fourth set of image acquisition terminals is deployed at the parking position of the sterilization cart.

[0082] In practical implementation, the visual tracking module of the surgical instrument AI management system in the sterilization supply center deploys distributed image acquisition terminals at the instrument recycling and receiving point, cleaning and unloading point, packaging workbench, and sterilization loading point. Above the conveyor belt at the instrument recycling and receiving point, the first set of image acquisition terminals is triggered when a recycling container containing used surgical instruments passes by, capturing a global top-down image showing the stacked state of all instruments within the container. At the cleaning machine exit at the cleaning and unloading point, the second set of image acquisition terminals is triggered when the cleaning rack is removed from the cleaning machine, scanning each instrument on each shelf of the cleaning rack to acquire high-resolution surface detail images of each instrument after cleaning. Above the workstation at the packaging workbench, the third set of image acquisition terminals continuously captures images from multiple angles of each instrument being handled during packaging operations. At the sterilization loading point, where the sterilization cart is parked, the fourth set of image acquisition terminals records images of each carrying unit entering the sterilization cart as the instrument packs are loaded, generating a loading process sequence image. In some embodiments, the captured instrument visual data is compared with a pre-trained instrument feature recognition model. The pre-trained instrument feature recognition model is obtained through deep learning training based on multi-angle instrument image samples. The function of this pre-trained model is to achieve end-to-end mapping from multi-source visual data to individual instrument identity, flow node location, and initial surface state. The model's input data is divided into four categories: the first category is visual feature fragments, derived from the outline region of a single instrument separated from the stacked state after image segmentation of the global top-view image; the second category is microscopic visual features, derived from pixel-level information reflecting the cleanliness, smoothness, and presence of water stains after feature enhancement and noise filtering of surface detail images; the third category is three-dimensional feature data, derived from the three-dimensional appearance model of the instrument synthesized after three-dimensional structural reconstruction of images from multiple angles, from which three-dimensional features reflecting the instrument's geometric shape and structural integrity are extracted; the fourth category is the overall semantic features of the global top-view image, directly extracted by the fourth sub-network from the original global top-view image through convolutional blocks of the same structure. The model's output data includes three parts: individual device identification results, specifically including the device model code, individual device serial number, and the type of standard device package to which it belongs; the current working node position, specifically one of the following: recycling, cleaning, packaging, or sterilization loading; and preliminary surface condition information of the device, specifically including the overall cleanliness score of the device and obvious stain identification. This pre-trained device feature recognition model consists of four parallel feature analysis sub-networks and one fusion classification network. Each feature analysis sub-network is a convolutional neural network structure, consisting of an input layer, three convolutional blocks, and a fully connected layer.Each convolutional block consists of a 2D convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. The convolutional kernel size of the convolutional layer is 3×3, with a stride of 1 and padding of 1, and the number of output channels is 64, 128, and 256 respectively. The pooling window size of the pooling layer is 2×2, with a stride of 2. The feature maps output from the three convolutional blocks are flattened and input into a fully connected layer containing 512 neurons, which outputs the feature vector of the corresponding sub-network. The four sub-networks process visual feature fragments, microscopic visual features, 3D feature data, and the overall semantic features of the global top-down image, respectively. The overall semantic features of the global top-down image are directly extracted from the original global top-down image by the fourth sub-network through convolutional blocks of the same structure. The feature vectors output by the four sub-networks are weighted and summed into a single comprehensive feature vector at the fusion layer. This fusion layer is followed by a fully connected hidden layer with 256 neurons and a fully connected output layer containing the total number of output categories. The output layer uses the Softmax activation function to generate predicted probabilities for various indicators in the instrument model code, individual instrument serial number, standard instrument package type, current working node position, and preliminary surface condition information of the instrument. Supervised learning is used for model training. The training dataset consists of multi-angle images of instruments collected during the historical flow of the sterilization supply center, along with corresponding manual annotations. The annotations include instrument model code, individual serial number, instrument package type, node position label, and surface condition score. During training, the cross-entropy loss function is used to measure the difference between the predicted probabilities and the true labels. The Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.9 times every 10 training epochs. The batch size is 32, and the training runs for 200 epochs. Early stopping is used during training to terminate the training when the validation set loss no longer decreases after 10 consecutive epochs, preventing overfitting. Image segmentation is performed on the global top-view image to separate the contour regions of individual instruments from the stacked state, and visual feature fragments are extracted for each contour region. Feature enhancement and noise filtering are performed on the surface detail image to extract microscopic visual features reflecting the cleanliness, smoothness, and presence of water stains on the instrument. Three-dimensional structure reconstruction is performed on images from multiple angles to synthesize a three-dimensional appearance model of the instrument, and three-dimensional feature data reflecting the geometric shape and structural integrity of the instrument are extracted from the three-dimensional appearance model. Visual feature fragments, microscopic visual features, and three-dimensional feature data are input in parallel into a pre-trained instrument feature recognition model.

[0083] In practice, the pre-trained device feature recognition model comprises multiple parallel feature analysis sub-networks. These sub-networks process the input visual feature fragments, microscopic visual features, and 3D feature data, respectively. The outputs of the feature analysis sub-networks are integrated in the model fusion layer. The fusion layer integrates the outputs of each sub-network using a weighted fusion formula, expressed as follows:

[0084]

[0085] in: The integrated characteristics after fusion are represented. The total number of subnetworks representing feature analysis. Representing the The feature vectors output by each sub-network Represents the corresponding number The learnable weight coefficients output by each sub-network are optimized and determined during model training. The final output of the model includes the device model code, device serial number, and the standard device package type to which it belongs. It also outputs the current working node position of the device (whether it is being recycled, cleaned, packaged, or being sterilized and loaded), and preliminary surface condition information including the overall cleanliness score and obvious stain indicators. Visual feature fragments are primarily derived from the device outline segmented from the global top-down image. Microscopic visual features are directly related to pixel-level texture and grayscale information in surface detail images. Three-dimensional feature data is derived from a stereoscopic appearance model reconstructed from images from multiple angles. Optionally, the training data for the pre-trained device feature recognition model covers image samples of the same device model under different wear levels and lighting conditions. In some embodiments, the identification of the device serial number relies on the representation of subtle, unique scratches or wear patterns on the device surface in the three-dimensional feature data. The overall cleanliness score is a quantitative value calculated based on the difference between microscopic visual features and preset standard features.

[0086] In one embodiment of the present invention, the automatic operation and update module invokes an automated operation unit based on a robotic arm. Based on the individual device identification results provided by the vision tracking module and the current working node position, the module executes preset physical operation actions at corresponding key nodes. These physical operation actions include sorting devices from the recycling container, loading devices onto the cleaning rack, arranging and combining devices on the packaging table, and unloading devices from the sterilization cart. When the system determines that the current working node position of a device is a device recycling receiving point, and the individual device identification result indicates that it is a specific type of device, the automated operation unit located in the recycling area controls the robotic arm to grab the device from the recycling container on the conveyor belt according to the device model code in the identification result, and accurately place it in the pre-treatment classification basket corresponding to this model code. When the system determines that the current working node position of a device is a cleaning and unloading point, and the cleanliness score in the preliminary surface status information of the device is higher than a preset threshold, the automated operation unit located at the cleaning and unloading point controls the robotic arm to transfer the device from the cleaning rack to the corresponding packaging conveyor belt or designated temporary storage cart according to the individual device identification result. When the system determines that the current working node of the instrument is the packaging operation table, and the system receives a packaging task instruction from the production plan, the automated operating unit of the robotic arm located at the packaging operation table, according to the instrument package list required by the packaging task instruction, sequentially picks up the corresponding individual instruments from the conveyor belt or temporary storage cart, and arranges the instruments on the prepared packaging material according to the preset instrument package placement template. When the system determines that the current working node of the instrument is the sterilization loading point, the automated operating unit of the robotic arm located at the sterilization loading point, according to the sterilization cycle plan, picks up the packaged instrument packages according to the preset loading mode and spatial layout, and stacks them into the sterilization cart.

[0087] In one embodiment of the present invention, after the pretreatment robotic arm completes the loading of the medical device package, the central control platform issues a transfer task to the intelligent transport robot module. The intelligent transport robot module transports the contaminated medical device package from the pretreatment area to the cleaning and decontamination area according to the target area and the current path status. After the cleaning rack is loaded, the intelligent transport robot module transports the cleaning rack to the interface of the automatic cleaning and disinfection machine. After sterilization is completed, the intelligent transport robot module transports the sterilization cart or sterilization rack to the cooling area or sterile storage area. When the system receives a distribution request from a clinical department, the intelligent transport robot module delivers the selected sterile medical device package to the target department, thereby realizing a closed-loop logistics system for medical devices from collection, cleaning, sterilization to distribution.

[0088] While the robotic arm performs physical operations, an RFID reader scans the embedded RFID tags on the instruments in real time, recording and updating the status changes of each tagged instrument during its transfer. This status change information includes the time of entry into the recycling area, the cleaning batch number, sterilization cycle parameters, storage shelf location, and the target department. When the robotic arm's automated operation unit places the recycled instruments into the pre-processing sorting basket, the RFID reader located above the basket scans the embedded RFID tags of all instruments within the basket, records the timestamp of the instrument entering the pre-processing area, and updates the instrument's status in the management system to "recycled and awaiting cleaning." After the robotic arm's automated operation unit transfers the instruments from the cleaning rack to the conveyor belt, the RFID reader located at the conveyor belt entrance scans the tags of the passing instruments, records the corresponding cleaning machine number and cleaning cycle number, and updates the instrument's status to "cleaned and awaiting inspection." After the robotic arm automated operation unit arranges and places the instruments on the packaging material and completes the packaging, it triggers an RFID reader / writer located on the packaging platform to scan the labels of all instruments inside the instrument pack. These labels are then linked to the unique traceability barcode affixed to the outside of the instrument pack, recording the packaging personnel and the packaging completion time. The status of all instruments inside the pack is updated to "packaged and awaiting sterilization." After the robotic arm automated operation unit places the instrument packs into the sterilization cart, it triggers an RFID reader / writer located inside the sterilization cart to scan the labels of all instrument packs within the cart. This identifies the sterilization batch number, sterilization program parameters, and planned sterilization time, and the status of the relevant instrument packs and their internal instruments is updated to "loaded and awaiting sterilization."

[0089] In one embodiment of the present invention, the central control platform obtains the operating status, idle status, process parameters, and task completion signals of the pretreatment cleaning machine, automatic cleaning and disinfection machine, sterilizer, and automatic packaging machine through the device interface module. When the central control platform determines that the corresponding device is in an available state, it issues loading or transfer instructions to the robotic arm automated operation unit and the intelligent transport robot module. After the device completes the corresponding processing procedure, the device interface module sends the completion signal and process parameters back to the central control platform. The central control platform associates and stores the process parameters with the instrument or instrument package circulation record for quality traceability and management analysis.

[0090] In one embodiment of the present invention, after the sterilization process is completed and verified to be qualified, the intelligent transport robot module transports the sterilization cart or sterilization rack to the cooling area; after cooling is completed, the robotic arm automated operation unit places the sterile instrument pack into the sterile storage rack or automated warehousing system, and the radio frequency identification reading and writing device records its storage location and updates the inventory status; when the system receives a request from a clinical department, the central control platform generates a picking task according to the first-in-first-out principle and storage location information, and the robotic arm automatically grabs the target instrument pack and loads it into the intelligent transport robot module, which then transports it to the target department to complete the distribution.

[0091] In practice, the automated operation and update module of the surgical instrument AI management system in the sterilization supply center calls upon the robotic arm-based automated operation unit. Based on the individual instrument identification results provided by the visual tracking module and the current working node position, it performs physical operations. For example, if a group of instruments, including surgical scissors with model code "FH-301" and hemostats with model code "NH-455," is identified as being at the instrument retrieval receiving point, the robotic arm automated operation unit sorts the two types of instruments from the retrieval container according to their model codes "FH-301" and "NH-455," placing them separately in pre-processing sorting baskets marked with their corresponding model codes. When a group of instruments is identified as being at the receiving point... When the cleanliness score of the instrument surface in the initial condition information is higher than the preset threshold of 95 points at the cleaning unloading point, the robotic arm automated operation unit transfers the instrument from the cleaning rack to the packaging conveyor belt. When the system receives the packaging task instruction to assemble the "laparoscopic basic package", the robotic arm automated operation unit sequentially grabs the instruments on the list from the packaging conveyor belt and arranges the instruments on the packaging material according to the preset placement template inside the laparoscopic basic package. When the instrument package is at the sterilization loading point, the robotic arm automated operation unit places the instrument package into the sterilization cart numbered "SC-07" according to the sterilization batch plan "BL-20250327-03" in the loading mode "Mode C".

[0092] In some embodiments, the execution of physical manipulation actions relies on the precise matching of instrument individual identification results and position information. For example, if the instrument individual identification result includes the instrument model code "ST-210" and the individual serial number "88453", and the current work node position is "packaging operation table - station 5", then the robotic arm automated operation unit performs grasping and arranging actions for the "ST-210" model instrument at station 5. When the instrument individual identification result indicates a fine microscopic instrument, the robotic arm automated operation unit switches to an end effector equipped with soft grippers and reduces the grasping force. In contrast, for heavy orthopedic instruments, the robotic arm automated operation unit switches to high-strength grippers and adopts a stable grasping strategy. In the packaging operation, the robotic arm automated operation unit calculates the grasping order and placement coordinates according to the instrument package list "List A" required by the packaging task instruction. The placement coordinates are defined by the placement template inside the instrument package. The objective function for calculating the grasping order is to minimize the total motion path of the robotic arm. , The calculation formula is:

[0093]

[0094] in: This represents the total number of equipment items that need to be grabbed in the current packaging task. Representing the The estimated pick-up coordinates of the equipment on the conveyor belt or temporary storage vehicle to be packaged. The coordinates of the next device to be picked up are represented by the square root operation, which calculates the straight-line distance between two consecutive picks. By optimizing the device picking sequence, the following can be achieved: Minimum.

[0095] In practice, the RFID reader scans embedded RFID tags in real time to record status change information. For example, when the robotic arm's automated operating unit places the equipment in the pre-processing sorting basket, the RFID reader above the basket scans the equipment tag with serial number "88453", records the timestamp "2026-03-25 10:23:15", and updates the equipment's status to "recovered and awaiting cleaning". When the equipment is transferred from the cleaning rack, the RFID reader at the conveyor belt entrance scans the same equipment tag, records the cleaning batch number "WQ-20250325-08", and updates the status to "cleaned". After packaging is completed, the RFID reader on the packaging table scans all instrument labels inside the instrument pack and binds them to the instrument pack traceability barcode "ZSB-20250325-10087". It records the packaging employee number "EMP1002" and the time "14:30:00", and updates the instrument status to "packaged and awaiting sterilization". After sterilization loading, the RFID reader in the sterilization cart scans the instrument pack label, associates it with the sterilizer number "BL-20250325-12", and the sterilization program parameters "temperature 134℃, pressure 2.1 bar, time 7 minutes", and updates the status to "loaded and awaiting sterilization". It is understandable that status transition information constitutes a complete digital log of device circulation. For example, the log fragment for device serial number "77124" shows that the time of entry into the recycling area was "08:15", the cleaning batch was "WQ-20250325-05", the packaging barcode was "ZSB-20250325-10035", and the distribution target department was "Operating Room 3". Optionally, if the RFID reader fails to read the tag, it will trigger an audio-visual prompt and pause the current action of the robotic arm automated operation unit, awaiting manual intervention. In some embodiments, the RFID reader scanning data at different key nodes is correlated with the action log of the robotic arm automated operation unit to ensure the atomicity and consistency of status updates. For example, the completion of a placement action by the robotic arm automated operation unit and the successful recording of a scan by the RFID reader are considered as an inseparable transaction. It is understood that the status information recorded by the RFID reader is a key data foundation for subsequent lifecycle prediction and dynamic distribution. Optionally, for devices without embedded RFID tags, the system generates a temporary virtual tag and associates it with their visual features. The status information is updated by capturing and recording images at key nodes through an image acquisition terminal.

[0096] In one embodiment of the present invention, the anomaly detection and handling module extracts edge feature data reflecting the overall contour of the device, key point angle data reflecting the flexibility of the device's joints or hinges, and pixel grayscale distribution data reflecting surface cleanliness from the preliminary surface state information of the device provided by the visual tracking module. The module retrieves the standard edge feature map, standard joint angle range, and standard cleanliness grayscale threshold corresponding to the device model that matches the individual device identification result from a preset device integrity standard map. The module compares the extracted edge feature data pixel-by-pixel with the standard edge features, calculates the contour overlap between the two, and determines that the device has structural damage, such as deformation or breakage, when the contour overlap is lower than the preset structural integrity threshold, and records the specific location of the damage and the degree of contour deviation. The module compares the extracted key point angle data with the standard joint angle range; if the angle data exceeds the standard range, it determines that there is a functional defect in the device, such as joints that are too tight or too loose, and records the location of the defective joint and the angle at which its movement is restricted. The module compares the extracted pixel grayscale distribution data with the standard cleanliness grayscale threshold, and calculates the area and distribution of abnormal pixel regions whose grayscale values ​​exceed the threshold. If the total area of ​​abnormal regions exceeds a preset contamination threshold, contamination residue is determined to exist, and the location and severity level of the contamination are recorded. The module integrates the judgment results of device structural damage, device functional defects, and contamination residue to generate a structured anomaly detection report. This report includes the anomaly type, anomaly location, anomaly severity quantification value, and the corresponding device identification result. The system has anomaly severity level judgment rules, which are based on the comparison relationship between the anomaly severity quantification value in the anomaly detection report and a series of preset thresholds, classifying anomalies into three levels: mild, moderate, and severe. When the anomaly detection report is judged as mild by the system, the system pushes the anomaly detection report and related images of the device to the quality control personnel's review workstation interface, prompting the staff to conduct manual visual review and confirmation. The system updates the device's status in the database based on the final result of the manual review. When an anomaly detection report is determined to be a moderate anomaly, the system controls the robotic arm automated operating unit to sort the equipment to an independent isolation container or area awaiting repair. Simultaneously, it sends a repair request containing a detailed anomaly detection report to the equipment maintenance management system and updates the equipment's status in the system to "awaiting repair." When an anomaly detection report is determined to be a severe anomaly, or when the system records show that the same equipment has accumulated a predetermined number of moderate anomalies, the system automatically generates a damage report in the equipment asset database. The damage report is linked to the equipment's complete circulation history and all previous anomaly detection reports, and the equipment's status is updated to "awaiting scrapping." Simultaneously, the system controls the robotic arm automated operating unit to transfer the equipment to the designated scrap storage area.

[0097] In practical implementation, the anomaly detection and handling module of the surgical instrument AI management system in the sterilization supply center performs difference analysis. For example, for an instrument whose individual identification result is "surgical forceps, model GP-220, serial number 7705", the module extracts edge feature data, key point angle data, and pixel grayscale distribution data from the preliminary surface condition information of the instrument. The module retrieves standard edge features, standard joint angle range, and standard cleanliness grayscale threshold that match the "GP-220" model from the preset instrument integrity standard map. In practical implementation, the module compares the extracted edge feature data with the standard edge features pixel by pixel, calculates the contour overlap, and determines that there is structural damage to the instrument when the contour overlap is 92% and the structural integrity threshold is 95%. The damage location is recorded as "2 mm from the tip of the forceps jaw" and the deviation degree is "contour deviation 0.8 mm". The module compares the extracted key point angle data with the standard joint angle range of "10 degrees to 45 degrees". If the key point angle data is 52 degrees, it is determined that there is a functional defect in the device, and the defective joint is recorded as "main axis hinge" and the angle of limited movement is "out of standard range by 7 degrees". The module compares the extracted pixel grayscale distribution data with the standard cleanliness grayscale threshold and counts the area and distribution of abnormal pixel regions that exceed the threshold. In some embodiments, the difference analysis involves specific quantitative calculations. For the determination of contamination residue, the module calculates the contamination area ratio. The calculation formula is:

[0098]

[0099] in: This represents the total area of ​​abnormal pixels within the effective detection area on the instrument surface whose pixel grayscale values ​​exceed the standard cleanliness grayscale threshold. This represents the total pixel area of ​​the effective detection area on the instrument surface. The calculated contamination area ratio... Exceeding the preset pollution threshold If contamination residue is detected, the contamination area and severity level are recorded. The module integrates the results of the assessment of device structural damage, device functional defects, and contamination residue to generate a structured anomaly detection report. The anomaly detection report includes the anomaly type, anomaly location, a quantitative value of the anomaly severity, and the corresponding individual device identification result. See Table 1 for an example excerpt of an anomaly detection report.

[0100] Table 1: Anomaly Detection Report Data Table

[0101]

[0102] In implementation, the system triggers subsequent processes based on the severity level of the anomaly detection report. The system sets anomaly severity level determination rules, which classify levels based on the comparison between the quantified anomaly degree in the anomaly detection report and a preset threshold. Taking the contour overlap of structural damage as an example, a mild anomaly is defined as a contour overlap of 90% to 95%, a moderate anomaly as a contour overlap of 85% to 90%, and a severe anomaly as a contour overlap of less than 85%. The system also uses the contaminated area ratio as a criterion. For example, the range of mild abnormalities is set as follows: The range of moderate abnormality is between 0.010 and 0.020. A score between 0.020 and 0.050 indicates a severe abnormality. Greater than 0.050. When an anomaly detection report is determined to be a minor anomaly, the system pushes the report, including the device identification result "GP-220,7705" and related images, to the quality control personnel's review workstation interface, prompting staff to conduct manual visual review and confirmation. The system updates the device status in the database based on the final result of the manual review. When an anomaly detection report is determined to be a moderate anomaly, the system controls the robotic arm automated operation unit to sort the device into an independent isolation container awaiting repair, and sends a repair request containing a complete anomaly detection report to the maintenance management system, while simultaneously updating the device status to "awaiting repair."

[0103] It is understandable that severity level determination rules can be configured differently for different anomaly types and instrument models. Optionally, for delicate and expensive microscopic instruments, their structural integrity threshold and contamination threshold may be set more strictly than those for ordinary instruments, resulting in the same quantification value of anomaly severity potentially triggering different severity levels on different instruments. In some embodiments, the cumulative number of moderate anomaly records for the same instrument is considered as one of the conditions for triggering severe anomaly handling. When an anomaly detection report is determined to be a severe anomaly, or when system records show that the same instrument has accumulated a predetermined number of moderate anomalies, the system automatically generates a damage report in the instrument asset database. It is understandable that the damage report is associated with the complete circulation history of the instrument and all previous anomaly detection reports. The system updates the instrument status to "pending scrapping" and simultaneously controls the robotic arm automated operation unit to move it to the scrap temporary storage area. Optionally, the generation of the damage report will simultaneously notify the equipment management department and the purchasing department.

[0104] In one embodiment of the present invention, when the system determines that an instrument is severely abnormal or has reached the preset damage reporting conditions, in addition to generating a damage reporting application form, it also pushes the name, model, specifications, abnormality type, abnormality image evidence, circulation history and service life information of the instrument to the equipment department management platform; the equipment department management platform then makes maintenance, scrapping or supplementary procurement decisions accordingly; when the system simultaneously detects that the inventory of the target instrument is lower than the safety threshold, the central control platform further generates a supplementary procurement or replenishment reminder to ensure the continuity of surgical instrument supply.

[0105] See Figure 3 In the visualization analysis of the distribution of joint angles detected by the instrument, the box plot intuitively presents the statistical distribution characteristics of the target instrument joint angles and compares them with the preset standard angle range (lower limit 10°, upper limit 45°). From the structure of the box plot, the box covers the interquartile range (IQR) of the data. The orange vertical line inside the box represents the median of the angle data, and the upper and lower edges of the box correspond to the 75th percentile and the 25th percentile, respectively. The whisker lines extending from both sides of the box show the overall distribution range of the data, reflecting the fluctuation range of the joint angles. As can be seen from the preset standard angle boundaries (red dashed lines), the overall distribution range of the box plot partially overlaps with the standard range: the core area of ​​the box (25th to 75th percentiles) is basically within the standard range of 10° to 45°, indicating that the joint angles of most tested samples meet the specifications; however, the right edge of the box and the upper beard line extend beyond the 45° standard upper limit, indicating that the joint angles of some samples exceed the standard upper limit, suggesting that the device has a functional defect risk of excessive joint mobility. This is consistent with the logic in the anomaly detection and handling module that "key point angle data exceeding the standard range is judged as a functional defect." The visualization results can directly provide quantitative basis for the anomaly detection module: the extreme values ​​and quantiles of the angle distribution obtained through box plot statistics can be automatically compared with the standard angle range, thereby generating an anomaly detection report that includes the location of the defective joint and the angle of limited movement, providing data support for subsequent manual review, sorting for repair, or handling of damage.

[0106] In one embodiment of the present invention, the system's lifecycle prediction module periodically extracts complete circulation records for each instrument from the system database. These complete circulation records include the instrument's cumulative cleaning counts, cumulative sterilization counts, contents of all anomaly detection reports, and all maintenance history records. The lifecycle prediction module performs data cleaning and feature extraction on these complete circulation records to construct a lifecycle feature vector reflecting the instrument's usage intensity, maintenance frequency, and performance degradation trend. The module calculates the average processing frequency per unit time from the cumulative cleaning and sterilization counts, using this as the usage intensity feature. From the anomaly detection reports, the module statistically analyzes the frequency of various anomalies, the changing trends of the time intervals between anomaly occurrences, and extracts the quantitative value of the most recent anomaly's severity; these data are integrated into a performance degradation trend feature. From the maintenance history, the module extracts the number of maintenance visits, the maintenance location, and the instrument's continued use duration after each maintenance, using this as the maintenance frequency and repair effectiveness features. The module arranges the usage intensity feature, performance degradation trend feature, and maintenance frequency and repair effectiveness feature in chronological order and performs normalization processing to form a time-series feature sequence. The module performs sliding window feature statistics on the time-series feature sequence, calculating the mean, variance, maximum value, and slope of change for each feature within the sliding window. These statistics are combined into a fixed-dimensional lifecycle feature vector. This lifecycle feature vector is input into a pre-trained device lifespan prediction model. The input data for the pre-trained model is the lifecycle feature vector, which is composed of usage intensity features, performance degradation trend features, and maintenance frequency and repair effectiveness features, all statistically analyzed using a sliding window. Usage intensity features represent the average processing frequency per unit time; performance degradation trend features include the number of occurrences of various anomalies, the trend of anomaly time intervals, and the quantified severity of the most recent anomaly; and maintenance frequency and repair effectiveness features include the number of repairs, the repair location, and the post-repair usage time. The model's output data is the predicted remaining safe service life or the predicted next failure time for each device. To demonstrate the intrinsic relationship between the input and output data, the model employs a multilayer perceptron regression network structure, consisting of an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the lifetime feature vector. The number of neurons in the three hidden layers are 256, 128, and 64, respectively. Each hidden layer is followed by a batch normalization layer and a ReLU activation function layer. Dropout layers are added after the first and second hidden layers, with a dropout rate of 0.2 to prevent overfitting. The output layer contains one neuron without an activation function, directly outputting the predicted remaining safe usage period in days.The model's function is to use a regression model trained on historical scrapped equipment data. By learning the nonlinear mapping relationship between features such as usage intensity, abnormal evolution trends, and maintenance effectiveness and the actual remaining lifespan, it achieves scientific prediction of the future safe service life of in-service equipment, providing a quantitative basis for preventative replacement. This pre-trained equipment lifespan prediction model is a regression model trained on a large amount of historical scrapped equipment data. The pre-trained equipment lifespan prediction model adopts a multilayer perceptron regression network structure, which includes an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the lifespan feature vector, i.e., the fixed-dimensional value after combining sliding window statistics. The number of neurons in the three hidden layers are 256, 128, and 64, respectively. Each hidden layer is followed by a batch normalization layer and a ReLU activation function layer. Dropout layers are added after the first and second hidden layers, with a dropout rate of 0.2 to prevent overfitting. The output layer contains one neuron, has no activation function, and directly outputs the predicted remaining safe service life in days. The model training dataset is derived from the complete circulation records of historically scrapped medical devices in the sterilization supply center. Each record includes the total number of cleanings, total number of sterilizations, contents of all abnormal detection reports, maintenance history, and the actual scrapping date of the device from its activation to its scrapping. For each record, a feature vector is extracted using the lifecycle feature vector construction method, and the remaining days from the time point corresponding to the feature vector to the actual scrapping date are used as the regression label. The mean squared error loss function is used during training, with Adam as the optimizer. The initial learning rate is set to 0.0005, decreasing by 0.5 times every 50 epochs. The batch size is 16, and a total of 500 epochs are trained. During training, an early stopping method is used: training stops when the validation set loss no longer decreases after 20 consecutive epochs, and the model parameters with the lowest validation set loss are saved. There is a clear causal relationship between the lifecycle feature vector input to the model and the predicted remaining safe service life output: the usage intensity feature in the feature vector reflects the rate of physical wear and tear of the equipment, the performance degradation trend feature characterizes the frequency and severity evolution of anomalies, and the maintenance frequency and repair effectiveness feature reflects the role of maintenance in extending the equipment's lifespan. The model learns the nonlinear mapping relationship between these features and the actual remaining lifespan from historical data to predict the future safe service life of in-service equipment. After receiving the vector, the model outputs the predicted remaining safe service life or the predicted time of the next failure for the corresponding equipment. The module compares the predicted remaining safe service life with a preset warning period. For equipment with a predicted remaining lifespan shorter than the warning period, the system generates a preventative replacement recommendation in advance and notifies management personnel.

[0107] In practical implementation, the lifecycle prediction module of the surgical instrument AI management system in the sterilization supply center periodically extracts the complete circulation record of each instrument from the system database. For example, for an instrument with the individual identification result of "surgical scissors, model SC-10, serial number 1208", the extracted complete circulation record includes a cumulative cleaning count of 1532 times, a cumulative sterilization count of 1489 times, the contents of each abnormal detection report, and maintenance history. The lifecycle prediction module performs data cleaning and feature extraction on the complete circulation record to construct a lifecycle feature vector reflecting the instrument's usage intensity, maintenance frequency, and performance degradation trend. The module calculates the average processing frequency per unit time from the cumulative cleaning count and cumulative sterilization count, using this as the usage intensity feature. Assuming the instrument was put into use on January 1, 2022, and the analysis cutoff date is March 25, 2026, with a total usage period of 1545 days, the average daily processing frequency is... The calculation formula is:

[0108]

[0109] in: This represents a cumulative total of 1532 cleaning sessions. This represents a cumulative total of 1489 sterilizations. The total number of days used is 1545, calculated as follows: Approximately 1.96 times / day. The module analyzes past anomaly detection reports to statistically analyze the frequency of various anomalies, the changing trends in the time intervals between occurrences, and the quantitative value of the most recent anomaly's severity, using these as performance degradation trend characteristics. For example, the module analyzes that "structural damage" anomalies occurred 5 times in the past year, the average time interval between two consecutive anomalies decreased from 120 days to 60 days recently, and the contour overlap of the most recent anomaly was 93%. The module extracts the number of repairs, the repaired parts, and the post-repair usage time from the maintenance history as characteristics of maintenance frequency and repair effectiveness. For example, if there were 2 repairs, both on "joint axes," the continuous usage time after the most recent repair was 180 days.

[0110] In some embodiments, constructing a lifecycle feature vector involves processing a time-series feature sequence. The module uses intensity features, performance degradation trend features, and maintenance frequency and repair effect features, arranged chronologically and normalized to form a time-series feature sequence. A sliding window feature statistical analysis is performed on the time-series feature sequence, calculating the mean, variance, maximum value, and slope of change within the window. For example, using a 90-day sliding window, the feature of "average daily processing frequency" within the window is statistically analyzed, and the mean frequency within the window is calculated. ,variance Maximum value and the slope of the frequency change at adjacent time points The module combines these statistics into a fixed-dimensional lifecycle feature vector. A simplified complete flow record segment and extracted features are shown in Table 2.

[0111] Table 2: Complete Transfer Record Fragment and Feature Extraction Table of Instrument SC-10(1208)

[0112]

[0113] In practical implementation, the lifecycle prediction module inputs the constructed lifecycle feature vector into a pre-trained device lifecycle prediction model. This pre-trained model is a regression model trained based on historical scrapped device data. After receiving the lifecycle feature vector, the model outputs the predicted remaining safe usage period or the predicted next failure time for each device. For example, for device "SC-10, serial number 1208", the model outputs a predicted remaining safe usage period of "65 days" and a predicted next failure time of "2026-05-20". The module compares the predicted remaining safe usage period with a preset warning period, for example, "90 days". Because the predicted remaining safe usage period of "65 days" is shorter than the warning period of "90 days", the system generates a preventative replacement recommendation for device "SC-10, serial number 1208" in advance.

[0114] It is understood that the training data for the pre-trained medical device lifespan prediction model comes from the complete circulation records and final disposal dates of historically scrapped medical devices of the same or similar models. Optionally, the operating cycle of the lifespan prediction module is configurable, for example, set to perform a full-scale predictive analysis of all medical devices weekly. In some embodiments, the calculation of performance degradation trend characteristics considers not only the number of anomalies but also introduces a weighting coefficient for the severity level of the anomalies; for example, severe anomalies have a higher weight than mild anomalies. It is understood that the size of the sliding window and the selection of statistics can be adjusted according to the type of medical device and management needs. Optionally, preventative replacement recommendations not only include medical device information and prediction results but can also be associated with the historical departments and frequencies of use of the medical device, providing a reference for procurement decisions.

[0115] See Figure 4The relationship between the maintenance effect and usage time / abnormality frequency of SC-10-1208 is visualized using a dual-axis approach, quantifying the performance and fault evolution patterns of device SC-10-1208 before and after maintenance intervention. From the left vertical axis (usage time after maintenance, unit: days) and the bar chart, it can be seen that: before maintenance, the usage time between single faults was approximately 60 days; after the first maintenance, the usage time increased to approximately 90 days, with maintenance intervention extending the fault-free operating cycle by 50%; after the second maintenance, the usage time further increased to approximately 180 days, a 200% increase compared to before maintenance and a 100% increase compared to after the first maintenance, demonstrating the significant benefit of continuous maintenance intervention in extending the safe operating cycle of the device. The trend of the broken line along the right vertical axis (frequency of anomalies, unit: times) shows that before maintenance, the frequency of anomalies was approximately 4.0 times, indicating a high-frequency failure state. After the first maintenance, the frequency of anomalies dropped to approximately 2.0 times, a 50% decrease in failure rate. After the second maintenance, the frequency of anomalies further decreased to approximately 1.0 times, a 75% decrease compared to before maintenance and a 50% decrease compared to after the first maintenance. This indicates that targeted maintenance can effectively suppress anomaly recurrence and significantly improve device stability. The figure visually verifies the repair effect of maintenance intervention on the performance degradation of the device's life cycle from two dimensions: usage duration and failure frequency. As the number of maintenance increases, the fault-free operating period of the device continues to extend, and the frequency of anomalies continues to decrease. This provides empirical evidence for the extraction of "maintenance frequency and repair effect characteristics" in the life cycle prediction module and also provides data support for the development of preventive maintenance strategies.

[0116] In one embodiment of the invention, the system's dynamic allocation module receives instrument pack requisition requests from the operating room in real time. These requests include the type, quantity, and planned usage time of the required instrument packs. The dynamic allocation module queries the current inventory status of the sterile storage area. Inventory status information includes the real-time inventory quantity, specific storage location, and earliest sterilization date for each type of instrument pack. If the query results show sufficient inventory, the module generates a detailed instrument pack picking list and optimal picking path for the requisition request based on the first-in, first-out (FIFO) principle and the storage location of the instrument packs. This list and path guide staff or transport robots to efficiently perform the allocation operation. If the query finds that the inventory quantity of a certain type of instrument pack is insufficient to meet the requisition request, the module triggers a dynamic allocation task. The system first queries the available inventory of scattered sterile instruments and their locations. These scattered sterile instruments originate from spare instruments or other instrument packs that allow for temporary allocation. Then, based on the individual instrument identification results and the standard instrument pack list, the system automatically calculates the optimal combination of instrument types and quantities required to meet the requisition request and generates an allocation task instruction. The assembly task instruction guides workers or robotic arm automated operation units to perform assembly operations, select suitable equipment from scattered inventory to form packages, and update the inventory status of relevant equipment packages and scattered equipment after the assembly operation is completed.

[0117] In practice, the dynamic allocation module of the surgical instrument AI management system in the sterilization supply center receives instrument pack requisition requests from the operating room in real time. For example, the received requisition request includes the required instrument pack type "laparoscopic basic pack", the quantity "3", and the planned usage time "2026-03-25 14:00". The dynamic allocation module then queries the current inventory status of the sterile storage area. The inventory status information shows that the inventory quantity of "laparoscopic basic pack" is 2, stored in "sterile shelf A-05" and "sterile shelf B-12" respectively, and the earliest sterilization date of both is "2026-03-20". The inventory quantity of another instrument pack type, "orthopedic debridement pack", is 15, stored in "sterile shelf C-03", and the earliest sterilization date is "2026-03-22". The dynamic allocation module compares the requested quantity of "3" with the inventory quantity of "2". If the inventory of "Laparoscopic Basic Kit" is insufficient, the module triggers a dynamic allocation task. The system queries the inventory of available spare sterile instruments and their locations. The query results show that the spare sterile instrument inventory includes "Laparoscopic trocars" located in "Spare Instrument Cabinet D-07" and "Laparoscopic grasping forceps" also located in "Spare Instrument Cabinet D-07". These spare sterile instruments originate from spare instruments. In some embodiments, the dynamic allocation module executes a standard allocation process when inventory is sufficient. If the requested quantity is "5 Orthopedic Debridement Kits" and the inventory quantity meets the requirement, the module selects the batch with the earliest sterilization date according to the first-in, first-out (FIFO) principle and generates an instrument pack picking list and optimal picking path based on the storage location. The optimal picking path guides staff or transport robots to perform the allocation. In scenarios with insufficient inventory, the system automatically calculates the optimal combination of instruments required to meet the requested quantity based on the individual instrument identification results and the standard instrument pack list. Determining the optimal combination requires evaluating and scoring all feasible spare instrument combinations using a scoring function. The aim is to maximize device compatibility and minimize relative packaging costs. The calculation formula is:

[0118]

[0119] in: This represents the matching score between the currently evaluated device combination and the standard device package list. The matching score is calculated based on factors such as device model consistency and device sterility expiration date. Represents all the items needed to complete the current combination. The total distance traveled by the scattered instruments Represents picking up the first The estimated travel distance required for each piece of miscellaneous equipment. It is a predefined reference distance baseline, such as the distance from the central dispensing workstation to the farthest scattered instrument cabinet. and Dimensionless weighting coefficients are used to balance matching priority and efficiency priority; both terms on the right-hand side of the formula are dimensionless to ensure dimensional consistency. System Selection The combination of devices with the highest score is the optimal combination.

[0120] In practical implementation, based on the optimal combination scheme, the dynamic assembly module generates an assembly task instruction. This instruction includes picking instruments "LC-001,8812" and "LC-102,7715" from "Loose Instrument Cabinet D-07" and combining them with two existing complete "Laparoscopic Basic Packages" in inventory to fulfill a three-package requisition request. The assembly task instruction guides staff or automated robotic arm units to perform the assembly operation. After the assembly operation is completed, the dynamic assembly module updates the inventory status of the relevant instrument packages and loose instruments. For example, it updates the status of the loose instruments "LC-001,8812" and "LC-102,7715" used for assembly from "Loose Inventory" to "Assembled and Issued," and correspondingly reduces the total issueable quantity of "Laparoscopic Basic Packages." It can be understood that when generating the picking list and optimal picking path, the system calculates the shortest physical path from the current issuance workstation to each inventory location and back. Optionally, a reference distance benchmark value can be used. Configuration and optimization can be tailored to the actual department layout. In some embodiments, the matching score... The score is composed of sub-items such as the instrument model matching score and the instrument sterilization date freshness score. The weighting coefficients in the formula can be understood. and The value is set after analyzing historical packing efficiency data. Optionally, if the inventory of scattered sterile instruments cannot meet the packing demand, the system will generate an inventory replenishment warning.

[0121] See Figure 5In the trend analysis of surgical instrument lifecycle management, the correlation between instrument health degradation and usage frequency can be intuitively presented through multi-dimensional time-series curves. Specifically, with the number of days of use as the horizontal axis, three core indicators are simultaneously quantified on the vertical axis: cumulative cleaning times (dot curve), cumulative sterilization times (box curve), and instrument health score (diamond curve). The evolution of the curves reveals typical lifecycle characteristics: a negative correlation between usage intensity and health degradation: as the number of days of use increases, the cumulative cleaning and sterilization times show a continuous monotonically increasing trend, while the instrument health score shows an approximately linear decreasing trend, reflecting the cumulative damage to instrument material fatigue, surface finish, and structural integrity caused by repeated high-temperature and high-pressure sterilization and mechanical cleaning. The stage characteristics of processing frequency: both the cleaning and sterilization frequency curves show a rapid increase in the early stage and a slowdown in the later stage, reflecting the high-frequency turnover of instruments during peak clinical use and the management strategy of gradually reducing usage intensity in the later stage due to declining health status and increased failure risk. Visual reference for warning thresholds: When the device health score drops to about 40 points, the cumulative number of cleaning and sterilization times reaches about 45 and 35 times, respectively. This can be used as a quantitative reference threshold for triggering preventive maintenance or replacement warnings, providing intuitive feature engineering basis for the life cycle prediction module.

[0122] In one embodiment of the present invention, the intelligent transport robot module is equipped with a navigation and drive unit, a cargo platform, and an identification unit. The identification unit integrates an RFID reader and a QR code scanner. Driven by transport task instructions issued by the central control platform, the intelligent transport robot module moves within a preset channel between the pretreatment area, cleaning and decontamination area, inspection and packaging area, sterilization area, sterile storage area, and clinical departments. When transporting instrument packs or cleaning racks, the intelligent transport robot module approaches the object to be transported. The identification unit scans the RFID tag or QR code attached to the instrument pack or cleaning rack to confirm the identity of the object to be transported and verifies it against the task instructions issued by the central control platform. After verification, the robotic arm or lifting platform of the intelligent transport robot module performs a loading operation, placing the object to be transported on the cargo platform, and then runs along the planned path to the target area.

[0123] In some embodiments, the point-to-point transportation tasks of the intelligent transport robot module include contaminated instrument retrieval and transfer, cleaning rack transfer, sterilized instrument pack delivery, and dispensing. For contaminated instrument retrieval and transfer, the intelligent transport robot module receives a retrieval instruction from a clinical department or operating room, moves to a designated retrieval point, loads a sealed retrieval container containing contaminated instruments, and transfers it to the pretreatment area of ​​the central sterilization supply center. For cleaning rack transfer, after the cleaning equipment completes its work cycle, the intelligent transport robot module receives a cleaning rack loaded with clean instruments from the cleaning unloading point and transfers it to a designated workbench in the inspection and packaging area. For sterilized instrument pack delivery, after the sterilization process is completed and the equipment has cooled, the intelligent transport robot module receives a sterilization cart loaded with sterile instrument packs from the sterilization loading point and transfers it to a designated shelf location in the sterile storage area according to the storage strategy of the central control platform. For dispensing, the intelligent transport robot module receives a dispensing instruction from the central control platform, picks the target instrument pack from a designated location in the sterile storage area, and delivers it to the clinical department or operating room that requested the instrument pack.

[0124] In practical implementation, the equipment interface module establishes data connections with various processing equipment in the disinfection supply center via standard industrial communication protocols. The module connects to the pretreatment cleaning machine, collecting its operating status, current water temperature, cleaning agent concentration, and program stage signals. It connects to the automatic cleaning and disinfection machine, collecting its door open / close status, cleaning cycle parameters, disinfection temperature, duration, and fault alarm codes. It connects to the sterilizer, collecting its batch number, sterilization program type, temperature, pressure, time, and other process parameters, as well as sterilization cycle start and end signals. It connects to the automatic packaging machine, collecting its operating status, packaging material type, sealing temperature, and packaging task completion signals. The module uploads the collected equipment operating status, process parameters, and task completion signals to the central control platform in real time.

[0125] Optionally, the data interaction function of the device interface module is bidirectional. The central control platform can send control commands or parameter settings to the connected processing equipment through the device interface module. The central control platform can issue start commands to the pretreatment cleaning machine or automatic cleaning and disinfection machine and set the cleaning program number. The central control platform can issue sterilization cycle plans to the sterilizer and trigger the sterilization program to start. The device interface module ensures that the central control platform can monitor the real-time operating status of the equipment and schedule equipment to work together according to the overall workflow.

[0126] In implementation, the central control platform, as the core scheduling unit of the system, maintains continuous communication with the vision tracking module, automatic operation and update module, intelligent transport robot module, equipment interface module, and RFID readers deployed at key nodes. The central control platform receives real-time information from the vision tracking module, including individual device identification results, current working node location, and preliminary surface status of the device. It also receives equipment operating status, process parameters, and task completion signals from the equipment interface module. Finally, it receives device identification tags and location timestamps scanned and uploaded by each RFID reader.

[0127] It is understandable that the central control platform, based on the aforementioned multi-source information, uniformly schedules the identification, transfer, equipment recall, and status updates of medical devices throughout the entire process of retrieval, cleaning, inspection, packaging, sterilization, storage, distribution, and damage reporting. The central control platform plans point-to-point transportation routes for the intelligent transport robot module and issues transportation instructions based on the current working node location and target node of the medical device. Based on the equipment status and idle time reported by the equipment interface module, the central control platform schedules the robotic arm automated operation unit in the automatic operation and update module to perform medical device loading, unloading, transfer, and distribution operations. For example, after the cleaning machine completes cleaning, the robotic arm automated operation unit is instructed to unload the cleaning rack. The central control platform correlates the cleaning and sterilization parameters collected from the equipment interface module with the medical device identity and operation time information obtained from the RFID reader and visual tracking module, achieving automatic collection and archiving of medical device information throughout the entire process.

[0128] In some embodiments, the scheduling logic of the central control platform is based on preset rules and optimization algorithms. The central control platform monitors the inventory quantity, storage location, and sterilization date of various types of instrument packs in the sterile storage area in real time. When the central control platform receives an instrument pack requisition request from the operating room, it queries the inventory status. If the inventory is sufficient, the central control platform generates a picking list and optimal route based on the first-in-first-out principle and the distance to the storage location, and dispatches the intelligent transport robot module or instructs staff to perform the distribution. When the instrument status changes, such as from "cleaned" to "packaged," the central control platform updates the instrument status record in the system database in real time to ensure that information and physical flow are synchronized. After the anomaly detection and handling module generates an anomaly detection report, the central control platform triggers information push to the review workstation, sends a request to the maintenance management system, or generates an asset loss application form according to the severity level of the report, and simultaneously updates the instrument status to "pending repair" or "pending scrapping." At the same time, it can dispatch the robotic arm automated operation unit to move the abnormal instrument to the isolation area.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A surgical instrument AI management system for a sterilization supply center, characterized in that: The system includes: The instrument identification management module, based on the surgical instrument requisition ledger of the Equipment Department, establishes an instrument package and a unique instrument identification mark that includes the instrument model code, factory serial number and initial activation time, and binds the instrument package and the unique instrument identification mark to the surgical instrument package and the physical surgical instrument through radio frequency identification tags; The visual tracking module uses a distributed image acquisition terminal to capture visual data of surgical instruments at key nodes in the flow of surgical instruments. The captured visual data of the instruments is compared with a pre-trained instrument feature recognition model. The comparison process outputs the individual instrument recognition result, the current working node position, and the preliminary state information of the instrument surface. The automatic operation and update module calls the automated operation unit based on the robotic arm, and executes the preset physical operation actions at the corresponding key nodes according to the individual identification results of the instruments and the current working node position. It scans the embedded radio frequency identification tags in real time through the radio frequency identification reading and writing device, and records and updates the status change information of each instrument bound with radio frequency identification tags during the circulation process. The anomaly detection and handling module performs a difference analysis between the preliminary surface condition information of the instrument and the preset instrument integrity standard map, generates an anomaly detection report on instrument structural damage, functional defects or contamination residue, and triggers a manual review process or an automatic loss reporting process according to the severity level of the anomaly detection report. The intelligent transport robot module performs the collection and transfer of contaminated instruments, the transfer of cleaning racks, the delivery and distribution of sterilized instrument packages under the scheduling of the central control platform, and completes the point-to-point logistics closed loop according to the preset or optimized path. The system includes an equipment interface module and a central control platform. The equipment interface module is connected to the pretreatment cleaning machine, automatic cleaning and disinfection machine, sterilizer, and automatic packaging machine, respectively, and is used to collect equipment operating status, process parameters, and task completion signals. The central control platform, based on the equipment operating status, the current working node position of the instruments, and transportation task information, schedules the robotic arm automated operation unit and the intelligent transportation robot module to perform instrument loading, unloading, transfer, and distribution operations, and automatically collects and archives cleaning parameters, sterilization parameters, and packaging information. The lifecycle prediction module periodically extracts the complete circulation record of each instrument from the system database. The complete circulation record includes the cumulative number of cleanings, the cumulative number of sterilizations, the contents of each abnormal detection report, and the maintenance history. Data cleaning and feature extraction are performed on the complete circulation record to construct a life cycle feature vector that reflects the intensity of instrument use, maintenance frequency, and performance degradation trend; The life cycle feature vector is input into a pre-trained device life prediction model, which is a regression model trained based on historical scrapped device data. The pre-trained device life prediction model outputs the predicted remaining safe service life or the predicted time of next failure for each device. The system compares the predicted remaining safe service life with the preset warning period, and generates preventive replacement recommendations in advance for devices that are about to reach the end of their service life. The dynamic assembly module receives instrument pack requisition requests from the operating room in real time. The requisition requests include the required instrument pack type, quantity, and planned usage time. Check the current inventory status of the sterile storage area, including the inventory quantity, storage location, and earliest sterilization date of each type of instrument pack; If the inventory is sufficient, a picking list and optimal picking path for the medical kits requested will be generated based on the first-in, first-out principle and storage location, guiding staff or transport robots to carry out the distribution. If there is insufficient stock of a certain type of instrument pack, a dynamic repacking task is triggered. The system queries the available stock of scattered sterile instruments and their locations. The stock of scattered sterile instruments comes from spare instruments or instruments temporarily transferred from other instrument packs. Based on the individual device identification results and the standard list of device packages, the system automatically calculates the optimal combination of devices required to meet the application request, generates a packaging task instruction, guides staff or robotic arm automated operation units to perform the packaging operation, and updates the inventory status of relevant device packages and scattered devices after the packaging is completed.

2. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 1, characterized in that, The method of using a distributed image acquisition terminal to capture visual data of surgical instruments at key nodes in the flow of surgical instruments includes: The key nodes include at least the equipment department requisition node, the operating room usage node, the instrument recycling and receiving point, the cleaning and unloading point, the packaging operation table, and the sterilization loading point. A first set of image acquisition terminals is deployed above the conveyor belt at the instrument recycling receiving point. When a recycling container containing used instruments passes by, the first set of image acquisition terminals is triggered to capture a global top-down image containing the stacked state of all instruments in the container. A second set of image acquisition terminals is deployed at the outlet of the cleaning machine at the cleaning unloading point. When the cleaning rack is removed from the cleaning machine, the second set of image acquisition terminals is triggered to scan each instrument on each shelf of the cleaning rack to obtain a detailed image of the surface of each instrument after cleaning. A third set of image acquisition terminals is deployed above the workstation of the packaging operation table. When the operator or the automated operation unit of the robotic arm performs packaging operations, the operation area is continuously photographed, capturing multiple angle images of each instrument that is picked up, inspected and placed into packaging materials. A fourth set of image acquisition terminals is deployed at the sterilization vehicle parking position at the sterilization loading point. During the process of loading instruments into the sterilization vehicle, images of each basket or tray entering the sterilization vehicle are recorded to generate a loading process sequence image.

3. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 2, characterized in that, The process of comparing the captured visual data of the device with the pre-trained device feature recognition model includes: The pre-trained instrument feature recognition model is obtained by deep learning training based on multi-angle instrument image samples; The global top-view image is processed by image segmentation to separate the contour regions of individual instruments from the stacked state, and visual feature fragments are extracted for each contour region. The surface detail image is subjected to feature enhancement and noise filtering to extract microscopic visual features that reflect the cleanliness, smoothness, and presence of water stains on the instrument. The images from multiple angles are subjected to three-dimensional structural reconstruction processing to synthesize a three-dimensional appearance model of the instrument, and three-dimensional feature data reflecting the geometric shape and structural integrity of the instrument are extracted from the three-dimensional appearance model. The visual feature fragments, the microscopic visual features, and the three-dimensional feature data are input in parallel into the pre-trained instrument feature recognition model; The pre-trained instrument feature recognition model includes multiple parallel feature analysis sub-networks, which process the input visual feature fragments, microscopic visual features, and three-dimensional feature data respectively. The outputs of the feature analysis sub-networks are integrated in the model fusion layer, and finally output the individual instrument recognition result including the instrument model code, individual instrument serial number, and the standard instrument package type to which it belongs. At the same time, it outputs the current working node position of the instrument, whether it is being recycled, cleaned, packaged, or sterilized and loaded, as well as the preliminary surface state information of the instrument, including the overall cleanliness score and obvious stain markings.

4. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 3, characterized in that, The invocation of the robotic arm-based automated operation unit, based on the individual device identification result and the current working node position, executes preset physical operation actions at the corresponding key node, including: The physical operations include sorting instruments from the recycling container, loading instruments onto the washing rack, arranging and combining instruments on the packaging table, and unloading instruments from the sterilization cart. When the current working node is a device recycling and receiving point, and the device identification result indicates that it is a specific type of device, the robotic arm automated operation unit located in the recycling area is controlled to grab the identified device from the recycling container according to the device model code in the device identification result, and place it in the pre-processing classification basket corresponding to the device model code. When the current working node is a cleaning and unloading point, and the cleanliness score in the preliminary status information of the instrument surface is higher than a preset threshold, the robotic arm automated operation unit located at the cleaning and unloading point controls the instrument to be transferred from the cleaning rack to the packaging conveyor belt or temporary storage vehicle corresponding to the individual instrument identification result according to the individual instrument identification result. When the current working node position is the packaging operation table and the system receives the packaging task instruction, the robotic arm automated operation unit located at the packaging operation table controls the corresponding instruments to be picked up from the conveyor belt or temporary storage vehicle in sequence according to the instrument package list required by the packaging task instruction, and arranges the instruments on the packaging material according to the preset instrument package placement template. When the current working node position is the sterilization loading point, the robotic arm automated operation unit located at the sterilization loading point controls the packaged instrument packs to be picked up and stacked into the sterilization cart according to the sterilization batch plan and the preset loading mode.

5. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 4, characterized in that, The process of scanning the embedded RFID tags in real time using an RFID reader / writer to record and update the status change information of each device bound with an RFID tag during its circulation includes: The status change information includes the time of entry into the recycling area, the cleaning batch number, the sterilization cycle parameters, the storage shelf location, and the target department for distribution. When the robotic arm automated operation unit places the device into the pre-processing sorting basket, it triggers the radio frequency identification reading and writing device set above the pre-processing sorting basket to scan the embedded radio frequency identification tags of all the devices in the basket, records the timestamp of the device entering the pre-processing area, and updates the status of the devices in the system to "recovered and awaiting cleaning". After the robotic arm automated operation unit transfers the instrument from the cleaning rack, it triggers the radio frequency identification (RFID) reader / writer located at the conveyor belt entrance to scan the embedded RFID tags of the passing instruments, record the corresponding cleaning machine number and cleaning cycle number, and update the status of the instruments in the system to "cleaned and awaiting inspection". After the robotic arm automated operation unit arranges the instruments on the packaging material, it triggers the radio frequency identification (RFID) reader / writer set on the packaging table to scan the embedded RFID tags of all instruments in the instrument package, binds the unique traceability barcode of the instrument package, records the packaging personnel and packaging time, and updates the status of the relevant instruments in the system to "packaged and awaiting sterilization". After the robotic arm automated operation unit places the instrument pack code into the sterilization cart, it triggers the radio frequency identification (RFID) reader / writer device installed in the sterilization cart to scan the embedded RFID tags of all instrument packs in the sterilization cart, associates the sterilization batch number, sterilization program parameters and planned sterilization time, and updates the status of the relevant instruments in the system to "loaded and ready for sterilization".

6. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 5, characterized in that, The preliminary surface condition information of the instrument is compared with a preset instrument integrity standard map to generate an anomaly detection report on structural damage, functional defects, or residual contamination of the instrument, including: From the preliminary surface state information of the instrument, extract edge feature data reflecting the overall contour of the instrument, key point angle data reflecting joint flexibility, and pixel grayscale distribution data reflecting surface cleanliness. From the preset instrument integrity standard map, retrieve the standard edge features, standard joint angle range, and standard cleanliness grayscale threshold corresponding to the instrument model that matches the individual instrument identification result; The edge feature data is compared pixel by pixel with the standard edge features to calculate the contour overlap. When the contour overlap is lower than the structural integrity threshold, it is determined that there is damage to the instrument structure, and the location and degree of deviation of the damage are recorded. The key point angle data is compared with the standard joint angle range. If the angle data exceeds the standard range, it is determined that there is a functional defect in the device, and the defective joint and the angle of limited movement are recorded. The pixel grayscale distribution data is compared with the standard cleanliness grayscale threshold. The area and distribution of abnormal pixel regions exceeding the threshold are counted. If the area of ​​abnormal regions exceeds the contamination threshold, it is determined that there is contamination residue, and the contamination area and severity level are recorded. Based on the combined results of the assessment of device structural damage, device functional defects, and contamination residues, a structured anomaly detection report is generated. The anomaly detection report includes the anomaly type, anomaly location, a quantitative value of the anomaly severity, and the corresponding individual device identification result.

7. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 6, characterized in that, Based on the severity level of the anomaly detection report, a manual review process or an automatic damage reporting process is triggered, including: The rules for determining the severity of anomalies are set. These rules are based on the comparison between the quantitative value of the anomaly degree in the anomaly detection report and a preset threshold, and divide the anomalies into three levels: mild, moderate and severe. When the anomaly detection report is determined to be slightly abnormal, the system pushes the anomaly detection report and the device image to the review workstation interface, prompting the staff to conduct manual visual review and confirmation, and updates the device status in the system based on the manual review results; When the anomaly detection report is determined to be a moderate anomaly, the system controls the robotic arm automated operation unit to sort the instruments into independent isolation containers awaiting repair, and sends a repair request containing the anomaly detection report to the maintenance management system, while updating the instrument status to "awaiting repair". When the abnormal detection report is determined to be a severe abnormality, or when the same device accumulates a predetermined number of moderate abnormalities, the system automatically generates a damage report in the device asset database. The damage report is associated with the complete circulation history of the device and all previous abnormal detection reports, and the device status is updated to "pending scrapping". At the same time, the system controls the robotic arm automated operation unit to transfer the device to the scrap temporary storage area. When the automatic loss reporting process is triggered, the system pushes the equipment name, model, specifications, anomaly type, anomaly image evidence, and circulation history to the equipment department management platform for the execution of repair, scrapping, or replenishment procurement decisions; when the target equipment inventory is lower than the safety threshold, the central control platform further generates replenishment or procurement reminders.

8. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 7, characterized in that, The complete circulation record is cleaned and its features are extracted to construct a lifecycle feature vector reflecting the intensity of instrument use, maintenance frequency, and performance degradation trend, including: The average treatment frequency per unit time is calculated from the cumulative number of cleaning times and the cumulative number of sterilization times, and is used as a usage intensity characteristic. From the various anomaly detection reports, the frequency of each anomaly, the changing trends of the time interval between occurrences, and the quantitative value of the anomaly severity of the most recent anomaly are statistically analyzed as performance degradation trend characteristics. From the maintenance history, the number of maintenance operations, the parts repaired, and the usage time after maintenance are extracted as characteristics of maintenance frequency and repair effectiveness. The usage intensity characteristics, the performance degradation trend characteristics, and the maintenance frequency and repair effect characteristics are arranged in chronological order and normalized to form a time-series feature sequence. Sliding window feature statistics are performed on the time-series feature sequence to calculate the mean, variance, maximum value and slope of change within the window, and the statistics are combined into a fixed-dimensional life cycle feature vector.

9. The AI ​​management system for surgical instruments in a sterilization supply center according to claim 1, characterized in that, The system also includes: The intelligent transport robot module is used to perform point-to-point transport of instrument packs, cleaning racks or sterile items between the pretreatment area, cleaning and decontamination area, inspection and packaging area, sterilization area, sterile storage area and clinical departments, and to confirm the identity of the object to be transported through radio frequency identification and / or QR code identification. The equipment interface module is used to interact with the pretreatment cleaning machine, automatic cleaning and disinfection machine, sterilizer and automatic packaging machine to collect equipment operating status, process parameters and task completion signals. The central control platform is connected to the visual tracking module, automatic operation and update module, intelligent transport robot module, equipment interface module and radio frequency identification reading and writing device. It is used to uniformly schedule the identification, transfer, equipment call and status update of instruments during the processes of recycling, cleaning, inspection, packaging, sterilization, storage, distribution and damage reporting.

Images