A medical fabric cleaning, disinfection and inspection method and device

Abstract

The present invention relates to the technical field of medical-related data processing, in particular to a method and device for cleaning, disinfecting and inspecting medical fabrics. By combining VMF, LeNet-5 and an edge learning framework, and integrating spectral analysis and chip recognition technology, it is used for the automatic inspection and sorting of medical fabric cleaning and disinfection. The results show that this method is significantly superior to the existing methods in terms of detection accuracy, speed and efficiency, providing a new technical approach for the standardized supervision and intelligent management of the quality of medical fabric cleaning and disinfection.

Description

Technical Field

[0001] The present invention relates to the technical field of medical-related data processing, and particularly to a method and device for cleaning, disinfecting, and inspecting medical fabrics. Background Art

[0002] Medical fabrics (including surgical gowns, bed sheets, towels, isolation gowns, etc.) play a crucial role in the medical environment, and the quality of their cleaning and disinfection directly affects medical safety and the control effect of nosocomial infections. According to the statistical data of the World Health Organization, about 15 - 30% of hospital-acquired infections (HAIs) are related to the improper handling of medical fabrics. Medical fabrics may carry a variety of pathogenic microorganisms, including bacteria (such as Staphylococcus aureus, Pseudomonas aeruginosa), viruses (such as norovirus, hepatitis A virus), and fungi, etc. If the cleaning and disinfection are not thorough, these pathogens may spread in the medical environment, causing cross-infection.

[0003] Currently, the quality inspection technology for medical fabric cleaning and disinfection mainly includes the following aspects:

[0004] Physical Index Detection:

[0005] Physical indexes mainly include the cleanliness, integrity, color, etc. of the fabric. Traditional detection methods rely on manual visual inspection, which has problems such as strong subjectivity, low efficiency, and inconsistent standards. In recent years, automatic detection systems based on computer vision have begun to be applied. By capturing fabric images with a high-resolution camera and using image processing algorithms to analyze the surface characteristics of the fabric, the detection ability for microscopic contaminants is limited. Research shows that even experienced medical staff can only detect 40 - 60% of trace contaminants (such as highly diluted blood, colorless and transparent body fluids, etc.) in pure visual inspection, which far from meets the requirements of medical safety.

[0006] Another important aspect of physical detection is the integrity inspection of the fabric. Medical fabrics may have physical damages such as wear, tear, or loose stitches during repeated use. These damages not only affect the barrier function of the fabric but also may become places where microorganisms hide. Traditional visual inspection is prone to overlooking subtle damages, while computer vision technology can more objectively evaluate the physical integrity of the fabric through feature extraction and pattern recognition. Currently, advanced computer vision systems can improve the accuracy rate of fabric integrity inspection to 85 - 90%, but there are still detection blind spots for more delicate microscopic structural damages.

[0007] Chemical Index Detection:

[0008] Chemical indicators mainly detect the chemical substances remaining on fabrics, including detergents, disinfectants, drugs, etc. Traditional methods include pH value determination, residual chlorine determination, etc. These methods usually require sampling and analysis, which is a cumbersome process and cannot cover all fabrics. Spectral analysis techniques provide new ideas for the detection of chemical residues. Among them, near-infrared spectroscopy (NIR) and Raman spectroscopy have shown good application prospects.

[0009] Research shows that different types of chemical residues have characteristic absorption or scattering peaks within specific wavelength ranges. For example, commonly used quaternary ammonium salt disinfectants have obvious absorption peaks near 1450 nm and 1950 nm; iodophor disinfectants have characteristic absorption within the range of 400 - 450 nm; while drug residues exhibit unique spectral characteristics at different wavelengths according to their chemical structures. By establishing the quantitative relationship between these characteristic peaks and the residue concentration, non-destructive and rapid detection of chemical residues can be achieved.

[0010] It is worth noting that excessive chemical residues may not only cause skin irritation and allergic reactions but also cross-react with other chemical substances in the medical environment, forming potential hazards. Therefore, accurately assessing chemical residues is an important part of the quality inspection of medical fabrics. However, there is still a lack of systematic solutions for chemical indicator detection at present, especially in real-time and full-coverage detection, which has obvious deficiencies.

[0011] Biological indicator detection:

[0012] Biological indicators mainly focus on the microbial residues on fabrics and are the key to evaluating the disinfection effect. The traditional microbial culture method is the gold standard, but it takes a long time (usually 24 - 72 hours) and cannot meet the needs of real-time monitoring; the ATP bioluminescence detection method can quickly evaluate the biological load, but its specificity is insufficient, and the false positive rate is as high as 15 - 25%; PCR technology can detect specific pathogens with high sensitivity, but it is costly and requires professional equipment and personnel.

[0013] Although the microbial culture method is time-consuming, it provides the most comprehensive information, including the types, quantities, and viability of microorganisms. According to the standards of the Centers for Disease Control and Prevention (CDC), the total number of microbial colonies on medical fabrics should be less than 20 CFU / dm², and specific pathogenic bacteria (such as Staphylococcus aureus, methicillin-resistant Staphylococcus aureus, etc.) should not be detected at all. However, due to the complex hospital environment and diverse types of microbial contamination, a single indicator is difficult to comprehensively reflect the biological safety of fabrics.

[0014] ATP bioluminescence detection is a widely used rapid detection method at present. This method is based on the fact that all living cells contain ATP (adenosine triphosphate). By measuring the intensity of the luciferase luminescence reaction catalyzed by ATP, the biological load on the surface is indirectly evaluated. However, ATP detection cannot distinguish microbial types, and the organic matter residues in the environment can also affect the detection results, leading to a high false positive rate.

[0015] Emerging molecular biology techniques, such as real-time PCR and high-throughput sequencing, provide new means for microbial detection. These techniques can detect specific pathogen nucleic acid sequences with extremely high sensitivity and specificity. However, due to the high cost of equipment and complex operation, they are currently mainly applied to scientific research or monitoring in key areas and have not been widely used in routine medical fabric inspections.

[0016] Process parameter monitoring:

[0017] In addition to the detection of the fabric itself, the monitoring of cleaning and disinfection process parameters is also an important aspect of quality control. Parameters such as temperature, time, disinfectant concentration, and mechanical force directly affect the cleaning and disinfection effect. Traditional methods are difficult to achieve precise control and traceability of the whole process through manual recording or single-sensor monitoring.

[0018] According to the recommendations of the World Health Organization, hot water washing disinfection of medical fabrics requires at least 71°C for 25 minutes or 80°C for 10 minutes; chemical disinfection requires a specific concentration of disinfectant to act within the specified contact time. Any deviation of these parameters may lead to poor disinfection effect. Research shows that more than 30% of medical fabric treatments in actual operations have parameter deviations, mainly due to equipment aging, operation errors, or process omissions.

[0019] The introduction of RFID technology has revolutionized process parameter monitoring. By implanting RFID tags into fabrics, the complete processing history can be recorded, including key parameters such as temperature curves, chemical agent concentrations, and processing times. These data can not only be used for quality assessment but also support problem tracing and process optimization. Currently, RFID tags have been developed to be resistant to high temperature, high pressure, and chemicals, meeting the harsh requirements of the medical fabric processing environment. The latest generation of intelligent tags even integrate temperature sensors to directly record the actual temperature process experienced by the fabric, further improving the monitoring accuracy.

[0020] However, the RFID system mainly provides indirect evidence of the fabric processing process and cannot directly reflect the actual cleaning and disinfection status of the fabric surface. It needs to be combined with other detection techniques to comprehensively evaluate the fabric quality. In addition, the integration, analysis, and utilization of a large amount of RFID data also face challenges, requiring more efficient data processing algorithms and models.

[0021] Currently, the quality inspection technology for medical textiles cleaning and disinfection still faces the following key technical problems:

[0022] The quality inspection of medical textiles cleaning and disinfection involves various data types, including spectral data (continuous values, high-dimensional), image data (pixel matrix), process parameters (discrete values, time series), microbial counts (sparse values), etc. These heterogeneous data have different characteristic dimensions, sampling frequencies, and noise characteristics. How to effectively fuse multi-modal data and construct a unified quality assessment model is an urgent problem to be solved. Hospitals need to process a large number of medical textiles every day, requiring the detection system to be highly efficient. However, high-accuracy detection methods (such as microbial culture) usually take a long time, while rapid detection methods (such as ATP detection) have lower accuracy. How to improve the detection speed while ensuring accuracy is the key challenge in system design. Summary of the Invention

[0023] The purpose of the invention is to provide a method and device for inspecting the cleaning and disinfection of medical textiles. By combining the improved Viterbi algorithm (referred to as the VMF algorithm in the present invention), the LeNet-5 convolutional neural network, and the edge learning framework, it solves the problems in the prior art such as the difficulty in fusing multi-modal data, the difficulty in balancing detection accuracy and real-time performance, and the low efficiency of distributed processing under resource constraints, and realizes the efficient, accurate detection and intelligent sorting of the quality of medical textiles cleaning and disinfection.

[0024] To achieve the above purpose, on the one hand, the technical solution adopted by the invention is as follows: Figure 1 As shown, a method for inspecting the cleaning and disinfection of medical textiles, the method includes the following steps:

[0025] Step 1, data acquisition: Collect multi-modal data of the textile through various sensors, including spectral data collected by a spectral analyzer, image data collected by a camera, and chip data read by an RFID reader;

[0026] Step 2, edge processing: Locally on the detection device, perform data preprocessing, feature extraction, and preliminary classification tasks. Deploy a lightweight version of the LeNet-5 network therein to perform real-time analysis on the spectral image and identify common pollutants and residues;

[0027] Step 3, cloud analysis: Utilize the complete VMF algorithm and the full-size LeNet-5 network to be responsible for in-depth analysis of complex samples, historical data mining, and model updating; Based on the edge learning framework, the cloud and edge nodes work together to dynamically adjust the task allocation according to the resource status and task complexity;

[0028] Step 4, application management: Provide a user interface and management functions;

[0029] Among them, the LeNet-5 network in the edge processing step is used to convert spectral data into two-dimensional images and then extract deep features to identify different types of pollutants and residues; the VMF algorithm in the cloud analysis step is used to regard the cleaning and disinfection process as a state transition sequence, and infer the most likely state sequence based on the process parameters recorded by the RFID chip and the spectral analysis results.

[0030] On the other hand, we also provide a medical fabric cleaning, disinfection and inspection device, which includes:

[0031] Data acquisition module: Equipped with a variety of sensors for collecting multimodal data of the fabric, including a spectral analyzer, a camera and an RFID reader / writer, which are used to collect spectral data, image data and chip data respectively;

[0032] Edge processing module: Set locally in the detection device, including a processing unit for performing data preprocessing, feature extraction and preliminary classification tasks, and a lightweight version of the LeNet-5 network is deployed to perform real-time analysis on spectral images;

[0033] Cloud analysis module: Includes a server processing unit, and a complete VMF algorithm and a full-size LeNet-5 network are deployed for deep analysis of complex samples, historical data mining and model update;

[0034] Application management module: Includes a user interface unit and a management unit, which are used to provide functions such as quality assessment result display, alarm prompt, statistical analysis and management decision support;

[0035] The device is used to execute a sorting method for cleaning and disinfecting medical fabrics using spectral and chip identification technologies as described above.

[0036] The beneficial effects of the invention are as follows:

[0037] Combining spectral analysis and chip identification technologies, capturing the surface substance composition and processing process information of medical fabrics, and accurately and quickly realizing the cleaning, disinfection and sorting of single-use fabrics. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a process execution diagram of a sorting method for cleaning and disinfecting medical fabrics using spectral and chip identification technologies provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The following further elaborates on the invention in detail in conjunction with specific embodiments.

[0040] The overall architecture of the intelligent device proposed by the present invention includes four modules: a data acquisition module, an edge processing module, a cloud analysis module and an application management module.

[0041] The data acquisition module is responsible for collecting multi-modal data of fabrics through various sensors, including spectral data collected by a spectral analyzer, image data collected by a camera, chip data read by an RFID reader, etc. This layer directly interacts with the physical world and converts physical information into digital signals. The data acquisition devices are distributed at various key nodes in the fabric processing process, such as the fabric receiving area, preprocessing area, washing area, drying area, disinfection area, and storage area, forming a whole-process monitoring network.

[0042] Specifically, the spectral analyzer uses near-infrared spectroscopy technology, operates in the wavelength range of 900 - 1700 nm, has a spectral resolution better than 10 nm, a sampling interval of about 6.3 nm, and forms spectral data of 128 wavelength points. 5 - 10 spectra are collected for each fabric at different positions to reflect the non-uniformity of the pollution distribution.

[0043] The camera uses an industrial-grade camera with a resolution of no less than 20 million pixels and is equipped with a controllable LED light source to ensure consistent image acquisition conditions. The system automatically triggers shooting and is equipped with a standard color card for subsequent color correction of the images.

[0044] The RFID reader supports the ISO 18000-6C standard, operates at a frequency of 860 - 960 MHz, and has an adjustable reading distance. The RFID tag is designed with high-temperature and high-pressure waterproofing, is embedded in a specific position of the fabric, and can withstand at least 100 high-temperature and high-pressure sterilization treatments. The tag stores information such as the fabric ID, type, usage history, and processing parameters.

[0045] The edge processing module is deployed locally on the detection device and performs data preprocessing, feature extraction, and preliminary classification tasks. A lightweight version of the LeNet-5 network is deployed at this layer to perform real-time analysis of spectral images and identify common pollutants and residues. The main advantage of the edge layer lies in low latency and high real-time performance, which can complete preliminary processing near the data generation location, reduce the amount of raw data transmission, and improve the system response speed.

[0046] The edge processing module adopts an embedded system design and has sufficient computing processing power, storage space, and network connection capabilities. The system runs an operating system optimized for edge computing and supports containerized deployment. Each edge node has the ability to cache data for 24 hours and can continue to work in case of network interruption.

[0047] The cloud analysis module integrates the complete VMF algorithm and the full-scale LeNet-5 network, and is responsible for the in-depth analysis of complex samples, historical data mining, and model updating. Based on the edge learning framework, the cloud and edge nodes work together to dynamically adjust the task allocation according to the resource status and task complexity. The cloud system is usually deployed in the hospital data center or dedicated servers to handle complex analysis tasks that cannot be completed by the edge layer, and is also responsible for the training and optimization of the global model.

[0048] The application management module provides user interfaces and management functions, including the display of quality assessment results, alarm prompts, statistical analysis, management decision support, etc.

[0049] These four levels are connected through standardized interfaces, and the data flow and control flow flow bidirectionally within the system. The data acquisition module transmits the raw data upward; after preliminary processing by the edge layer, the results or suspicious sample information are passed upward; the cloud layer performs in-depth analysis and feeds back the guidance information downward; the application management module receives the analysis results and passes the user operation instructions downward. The entire system constitutes a closed-loop control network to achieve the full-process monitoring and management of the quality of medical textiles.

[0050] The standardized interfaces are used for connection between each layer:

[0051] 1. Between the data acquisition module and the edge processing module, the raw data is mainly transmitted through USB, RS-232 or Ethernet interfaces, in JSON or Protocol Buffers format, and data compression is supported;

[0052] 2. Between the edge processing module and the cloud analysis module, communication is carried out through the TCP / IP network, using the MQTT protocol to transmit the processing results or complex sample feature data, and TLS / SSL encryption is used to ensure data security;

[0053] 3. Between the cloud analysis module and the application management module, communication is carried out through the RESTful API interface to transmit the analysis results and control instructions, and user authentication and session management are supported.

[0054] In the quality inspection of the cleaning and disinfection of medical textiles, the whole process from the use, recycling, pretreatment, washing, disinfection to storage of the textiles is a state transition sequence. Based on the process parameters and environmental data recorded by the RFID chip, combined with the spectral analysis results, a Hidden Markov Model (HMM) can be constructed, and the improved VMF algorithm can be used to infer the most likely state sequence.

[0055] Define the state space of medical textiles , including the following aspects:

[0056] 1. Clean state : Indicates the degree of removal of visible stains on the fabric surface, and the following specific scheme is adopted:

[0057] : Severe pollution (pollutant coverage area > 10% or obvious large stains)

[0058] : Moderate pollution (pollutant coverage area 5 - 10%)

[0059] : Light pollution (pollutant coverage area 1 - 5% or a small amount of visible stains)

[0060] : Surface clean (no visible stains to the naked eye)

[0061] 2. Disinfection status : Indicates the microbial load level of the fabric, adopting the following specific scheme:

[0062] : High load (> 100 CFU / dm²)

[0063] : Medium load (10 - 100 CFU / dm²)

[0064] : Low load (1 - 10 CFU / dm²)

[0065] : Qualified (< 1 CFU / dm²)

[0066] 3. Chemical residue status : Indicates the residue degree of chemical substances such as detergents and disinfectants, adopting the following specific scheme:

[0067] : Severe residue (more than 50% exceeding the safety limit)

[0068] : Moderate residue (exceeding the safety limit by 0 - 50%)

[0069] : Slight residue (the residue amount is within the safety limit)

[0070] : No residue (residue cannot be detected or is below the detection limit)

[0071] 4. Fabric integrity status : Indicates the physical integrity of the fabric, adopting the following specific scheme:

[0072] : Unusable (broken and irreparable or function impaired)

[0073] : To be repaired (with damage or severe wear of 5 - 20 mm)

[0074] : Slightly damaged (with damage or wear less than 5 mm)

[0075] : In good condition (without any damage)

[0076] Comprehensive status Can be expressed as a combination of these four aspects: , forming a multi - dimensional state space. This multi - dimensional state representation method enables the system to evaluate quality problems in different aspects separately and adjust the weights of each dimension according to the application scenario. The state space contains a total of possible states, and each state can be represented by a unique quadruple.

[0077] State transition probability matrix Is defined as:

[0078] ;

[0079] Where Represents the probability of transitioning from state To state . State transitions satisfy the following physical constraints:

[0080] 1. The cleanliness and degree of disinfection usually increase monotonically during the treatment process and do not decrease, unless re - contamination occurs;

[0081] 2. Chemical residues usually decrease monotonically in the subsequent steps of washing;

[0082] 3. The integrity state may remain unchanged or decrease, but will not increase by itself. [[ID=4,6]]

[0083] Observation space Includes all directly measurable indicators, which come from the following aspects:

[0084] 1. Spectral characteristics : The absorption or reflection intensity at different wavelengths, which can be represented as a vector , where Is the spectral intensity at wavelength , Is the number of wavelength points;

[0085] 2. Image characteristics : Characteristics such as the color and texture of the fabric surface, which are feature vectors extracted by the LeNet - 5 network;

[0086] 3. Process parameters : Parameters such as washing temperature, time, and disinfectant concentration read from the RFID chip;

[0087] 4. Environmental parameters : Conditions such as temperature and humidity of the washing and storage environment.

[0088] Comprehensive observation Can be expressed as: This multi-modal observation design is based on the complementarity of different sensing data and can capture fabric quality characteristics from different perspectives.

[0089] The observation probability is:

[0090] ;

[0091] Indicates the probability of observing in state . In practical applications, we train a Bayesian network or Gaussian mixture model with a large number of labeled samples to represent the observation probability distribution.

[0092] The traditional Viterbi algorithm finds the most likely state sequence through dynamic programming, and its recurrence formula is:

[0093] ;

[0094] ;

[0095] Among them, represents the maximum probability of being in state j at time t, is the state transition probability, is the observation probability, records the path information.

[0096] To improve the calculation efficiency, we improved the Viterbi algorithm and, on this basis, proposed the Viterbi-Medical-Fabric (VMF) algorithm.

[0097] The core of the VMF algorithm is to define the evaluation function:

[0098] ;

[0099] Among them, represents the negative logarithm probability from the initial state to the current state , is the estimated cost from the current state to the target state. The negative logarithm probability conversion converts probability multiplication into cost addition, which is convenient for calculation and comparison.

[0100] Based on the following three aspects:

[0101] 1. Process parameter analysis: By comparing the recorded parameters such as temperature, time, and disinfectant concentration with the standard treatment requirements, estimate the remaining steps and their costs required to reach the target state;

[0102] 2. Historical sample analogy: By comparing with samples in the historical database with similar characteristics and treatment conditions, predict the possible state change path and cost of the current sample, and use the K-nearest neighbor algorithm to select the most similar historical sample as a reference;

[0103] 3. Expert rule encoding: Encode the empirical knowledge of hospital infection control experts into heuristic rules, such as "If the washing temperature is below 70°C, the disinfection status is unlikely to reach the 'qualified' level". These rules are implemented through decision trees or fuzzy logic to provide cost estimates based on domain knowledge.

[0104] The heuristic function we use has the following complete expression. Generally speaking, meeting the above conditions is sufficient:

[0105] ;

[0106] where , , are weight coefficients, satisfying and are all non-negative values. In this embodiment, the parameters we use are , , .

[0107] The process parameter heuristic function is:

[0108] ;

[0109] where:

[0110] is the number of process parameters

[0111] is the weight of the parameter , satisfying

[0112] is the current parameter value

[0113] is the standard parameter value required to reach the target state

[0114] is the parameter 's normalized distance function

[0115] For the temperature parameter, the distance function is defined as:

[0116] ;

[0117] For the time parameter, the distance function is defined as:

[0118] ;

[0119] For the disinfectant concentration parameter, the distance function is defined as:

[0120] ;

[0121] These distance functions ensure that when the parameter reaches or exceeds the standard value, its contribution to the heuristic function is zero; while when the parameter does not meet the standard, the contribution is proportional to the gap.

[0122] Historical sample heuristic function Using the K-nearest neighbor algorithm, calculate the similarity between the current state and historical samples, and estimate the cost based on the state transition path of historical samples:

[0123] ;

[0124] Where:

[0125] is the number of neighboring samples considered (in this embodiment ), is the -th sample state found in the historical database that is most similar to , is the cost (negative log probability) actually spent by sample to reach the target state ,

[0126] The state similarity is calculated using the weighted Euclidean distance:

[0127] ;

[0128] Where: is the state space dimension (in this example ), is the weight of dimension , is the value of the current state in dimension , is the value of the historical sample state in dimension .

[0129] Expert rule heuristic function Encode the empirical knowledge of hospital infection control experts into a series of rules with penalty values:

[0130] ;

[0131] Wherein: is the total number of rules, is the rule function, which returns 1 when the state violates the rule and returns 0 otherwise. is the penalty value for violating the rule . The system implements the following five core expert rules:

[0132] 1. Temperature - Disinfection Rule:

[0133] ;

[0134] ;

[0135] This rule states that: if the maximum temperature is below 70°C, the disinfection status is unlikely to reach the "qualified" level.

[0136] 2. Time - Disinfection Rule:

[0137] ;

[0138] ;

[0139] This rule states that: if the high - temperature exposure time is insufficient (25 minutes at 71°C, 10 minutes at 80°C), the disinfection status is unlikely to reach the "qualified" level.

[0140] 3. Cleaning - Disinfection Rule:

[0141] ;

[0142] ;

[0143] This rule states that: if the cleaning status is "moderate pollution" or "severe pollution", the disinfection status is unlikely to reach the "qualified" level.

[0144] 4. Disinfectant - Residue Rule:

[0145] ;

[0146] ;

[0147] This rule states that: if the disinfectant concentration exceeds 1.5 times the standard concentration, the chemical residue status is unlikely to be "no residue".

[0148] 5. Rinsing - Residue Rule:

[0149] ;

[0150] ;

[0151] This rule states that if the number of rinsing times is less than 3, the chemical residue state is unlikely to be "no residue".

[0152] These rules guide the algorithm to search along a more reasonable path by penalizing physically unlikely state transitions, improving the search efficiency and result accuracy. The penalty values are set based on expert evaluation and historical data analysis, reflecting the severity of rule violations.

[0153] To ensure the correctness of the VMF algorithm, the heuristic function must satisfy admissibility. The heuristic function designed in the invention is based on the statistical analysis of a large amount of historical data, ensuring that the estimated future cost will not exceed the expected value of the actual cost, and guaranteeing that the algorithm finds the optimal solution. Through comparison and verification, when the actual cost of a shortest path is , the average estimated value of the heuristic function is , and the maximum estimated value does not exceed , meeting the admissibility requirements.

[0154] The specific steps are as follows:

[0155] 1. Define the evaluation function , where , is the heuristic function;

[0156] 2. Initialization: For all states , set , calculate , and add to the priority queue Q;

[0157] 3. Iteration: When the priority queue Q is not empty, execute the following steps:

[0158] a) Take out the state [[ID=5�2]]with the minimum evaluation function value from Q;

[0159] b) If (the length of the observation sequence), then find the optimal path and end the algorithm;

[0160] c) Otherwise, for all possible next states , calculate ;

[0161] d) Calculate the heuristic function ;

[0162] e) Calculate the evaluation function ;

[0163] f) Add to the priority queue Q and record the path information;

[0164] 4. Backtracking: According to the recorded path information, backtrack from the final state to the initial state to obtain the optimal state sequence.

[0165] Through controlled experiments, it is verified that for a medical fabric processing model with 256 combined states and 20 time steps, the VMF algorithm significantly shortens the calculation time compared to the traditional Viterbi algorithm, while maintaining the same inference accuracy. This performance improvement mainly comes from the guidance of the heuristic function, enabling the algorithm to effectively avoid ineffective exploration of low-probability state paths. In the actual application environment, the stability and accuracy are verified, and the processing anomalies and quality risks are reliably identified.

[0166] The pollutants and residues on the surface of medical fabrics often have specific spectral characteristics. By visualizing the spectral data as a two-dimensional image, the invention uses an improved LeNet-5 convolutional neural network to extract deep features and identify different types of pollutants and residues.

[0167] Spectral data is usually a one-dimensional vector , representing the light intensity at K different wavelengths. To utilize the two-dimensional feature extraction ability of LeNet-5, the spectral data needs to be converted into a two-dimensional image. The following method is adopted in this embodiment:

[0168] 1. Data preprocessing: First, perform Savitzky-Golay filtering, baseline correction, and normalization on the spectral data.

[0169] a) Savitzky-Golay filtering uses parameter settings with a window width of 9 and a polynomial order of 3. The selection of this parameter combination is based on systematic testing of 500 standard samples. This configuration can maximize the retention of the morphological characteristics of spectral peaks while effectively removing high-frequency noise. The filtering formula is:

[0170] ;

[0171] where is the smoothed spectral intensity value, is the original spectral intensity value, is the convolution coefficient, is the window half-width.

[0172] b) Baseline correction uses the progressive baseline correction algorithm, which is more robust than polynomial fitting when dealing with non-linear baseline drift in the spectrum and is particularly suitable for the spectra of medical fabrics with non-uniform baseline interference.

[0173] c) Normalization processing uses the z-score method to transform the spectral data into zero mean and unit variance:

[0174] ;

[0175] in, is the mean, is the standard deviation.

[0176] 2. Zigzag Rearrangement: The one-dimensional spectral data is rearranged into a 32×32 two-dimensional matrix using a zigzag path. This zigzag path was determined to be optimal after systematically comparing multiple rearrangement methods because it maximizes the spatial correlation between adjacent wavelength points, facilitating the convolution operation to extract effective features.

[0177] The mathematical expression of the zigzag rearrangement is:

[0178] ;

[0179] in, It is the image Rank The pixel value of the column, It is This method rearranges the 128 wavelength points into a matrix of approximately 12×11 and then pads it to the standard size of 32×32.

[0180] Multi-channel fusion: To provide richer feature representation, spectral data from different preprocessing methods are merged as different channels:

[0181] (raw reflectance data);

[0182] (first derivative data);

[0183] (absorbance data);

[0184] The first-order derivative data is obtained by difference calculation:

[0185] ;

[0186] in is the sampling wavelength interval.

[0187] The absorbance data were logarithmically transformed:

[0188] ;

[0189] The first derivative data emphasizes the rate of change of the spectral curve and is sensitive to the peak position; the absorbance data has a linear relationship with the substance concentration, facilitating quantitative analysis. This multi-channel representation provides complementary information and enhances the separability of features.

[0190] According to the characteristics of the spectral image, the invention makes adaptive adjustments to the traditional LeNet-5 network structure, mainly including:

[0191] 1. Input layer: Receive of the spectral image, where C is the number of channels;

[0192] For multi-channel input processing, the following strategy is adopted:

[0193] Each channel uses an independent set of convolutional kernels for initial feature extraction;

[0194] In the first convolutional layer, for C-channel input, 6C convolutional kernels are used;

[0195] The convolutional kernel parameters are learned independently to capture the unique features of different channels;

[0196] The convolutional output is grouped by channel, and each group of 6 feature maps corresponds to one input channel;

[0197] Inter-channel feature fusion is achieved in the second convolutional layer using cross-channel convolution operations;

[0198] 2. First convolutional layer: 6C convolutional kernels (C-channel input), feature maps of (C channels);

[0199] The mathematical expression of the convolutional layer is:

[0200] ;

[0201] where is the output of the first convolutional layer, is the convolutional kernel weight, is the bias term, is the activation function.

[0202] 3. First pooling layer: Max pooling, feature maps of;

[0203] The max pooling operation is defined as:

[0204] ;

[0205] 4. Second convolutional layer: Use 16 convolutional kernels and output feature maps of;

[0206] For multi-channel inputs, the second convolutional layer implements feature fusion across channels:

[0207] ;

[0208] where are the convolutional kernel weights for fusing features from different channels, adjusting the importance of features from different channels through learning.

[0209] 5. Second pooling layer: Max pooling, outputting feature maps;

[0210] 6. First fully connected layer: 120 neurons, with the ReLU activation function;

[0211] Batch normalization is applied to the input of the fully connected layer:

[0212] ;

[0213] where and are the batch statistics, and are the learnable parameters, is the numerical stability constant.

[0214] 7. Second fully connected layer: 84 neurons, with the ReLU activation function;

[0215] Dropout regularization is applied to the output of the fully connected layer, with a dropout rate p = 0.5:

[0216] ;

[0217] 8. Output layer: The number of neurons is determined according to the task type. The Softmax activation function is used for classification tasks, and the linear activation function is used for regression tasks.

[0218] Compared with the original LeNet-5, the main adjustments include:

[0219] 1. Adjust the input layer to support multi-channel spectral images;

[0220] 2. Replace average pooling with max pooling to improve the sensitivity to feature peaks;

[0221] These adjustments are targeted at the characteristics of spectral images and the requirements of medical fabric quality inspection, enabling the network to effectively extract spectral features and achieve accurate classification. In particular, the combination of max pooling helps capture peak features in spectral data, and these peaks are key clues for identifying specific pollutants.

[0222] The network training adopts the supervised learning method, and the loss function is selected according to the task type:

[0223] For the classification task (pollutant type recognition), the cross-entropy loss is used:

[0224] ;

[0225] where is the true label (0 or 1) that the i-th sample belongs to the j-th class, is the predicted probability, is the number of samples, is the number of classes.

[0226] For the regression task (pollutant concentration estimation), the mean squared error loss is used:

[0227] ;

[0228] where is the true value, is the predicted value.

[0229] The Adam algorithm is selected as the optimizer, the initial learning rate is set to 0.001, and the learning rate decay strategy is used:

[0230] ;

[0231] where is the learning rate of the t-th round of training, is the initial learning rate, and decay is the decay rate (set to 0.01). The Adam optimizer is selected because it shows better convergence speed and stability than SGD in high-dimensional sparse gradient problems such as spectral data.

[0232] To prevent overfitting, the following regularization is adopted:

[0233] 1. Weight decay (L2 regularization), with coefficient ;

[0234] 2. Dropout, with a dropout rate of 0.5;

[0235] 3. Data augmentation, including:

[0236] a) Random noise addition: Gaussian noise is added to the original spectrum, and the noise level is 5% of the standard deviation of the original signal

[0237] The specific implementation is:

[0238] ;

[0239] where , is the standard deviation of the original signal, Represents a normal distribution.

[0240] b) Wavelength offset: Randomly offset the spectral data along the wavelength axis, with an offset of ±50% of the spectral resolution

[0241] The specific implementation is:

[0242] ;

[0243] in is a random offset, is the spectral resolution.

[0244] c) Intensity scaling: Randomly scale the overall intensity of the spectrum to 90%-110% of the original intensity

[0245] The specific implementation is:

[0246] ;

[0247] in is the random scaling factor.

[0248] d) Baseline drift simulation: Add random low-frequency baseline drift with a drift amplitude of ±1% of the signal intensity range

[0249] The specific implementation is:

[0250] ;

[0251] in is the random drift coefficient, and are the maximum and minimum values of the original signal respectively.

[0252] The data augmentation method was selected based on an analysis of spectral variations that can occur in real-world environments. These variations primarily stem from instrument noise, environmental interference, and sample preparation differences. By simulating these variations, the model's generalization capabilities were significantly improved, resulting in higher recognition accuracy on the validation set.

[0253] To meet the deployment requirements of edge devices, the LeNet-5 network is lightweight:

[0254] 1. Depthwise separable convolution: decompose the standard convolution into depthwise convolution and pointwise convolution to reduce the number of parameters and computation. The computational complexity of standard convolution is , and the depthwise separable convolution is ,in is the convolution kernel size, is the number of input channels, is the number of output channels, is the size of the feature map.

[0255] 2. Knowledge Distillation: Use the trained complete model (teacher network) to guide the training of the lightweight model (student network). The distillation loss function is:

[0256] ;

[0257] where is the output of the student network, is the output of the teacher network, is the temperature parameter, is the balance parameter, is the KL divergence loss. The temperature parameter is chosen as 3 based on the systematic experiment comparison of different temperature parameters, and this value achieves the best balance between knowledge transfer efficiency and model independence.

[0258] 3. Weight Quantization: Quantize 32-bit floating-point weights into 8-bit integers, and the quantization formula is:

[0259] ;

[0260] where is the quantized integer weight, is the bit width.

[0261] 4. Channel Pruning: Remove the convolutional channels that contribute less to the final prediction. The channel importance is evaluated based on the L1 norm:

[0262] ;

[0263] where is the channel importance score, is the weight associated with this channel. Retain the top 70% of the channels with importance scores and perform fine-tuning later to restore the accuracy.

[0264] The medical fabric cleaning and disinfection quality inspection system is usually deployed at multiple spatially dispersed detection points, and the computing capabilities and network conditions of different nodes vary. Based on the edge learning concept, the invention designs an adaptive distributed processing framework to achieve efficient utilization of computing resources and real-time response. The system adopts an edge-cloud collaborative architecture. The edge layer includes multiple detection nodes, and each node is equipped with a basic computing unit, responsible for data collection, preprocessing, and preliminary analysis; the cloud layer provides powerful computing resources, responsible for in-depth analysis of complex data, model training, and knowledge base maintenance.

[0265] Edge nodes are usually deployed at key positions in the medical fabric processing flow: receiving stations, pre-washing areas, main washing areas, drying areas, disinfection areas, and storage areas. Each edge node is equipped with different sensing devices and computing modules according to its function. The system supports a variety of edge devices, including industrial computers, embedded systems, or intelligent gateway devices, as long as they meet the minimum computing power requirements.

[0266] The cloud system is usually deployed in a hospital data center or on a dedicated server, providing the following functions: high-performance computing, large-capacity storage, model training, data integration, and management interfaces. The cloud system supports horizontal scaling and can add server nodes according to the hospital scale and processing requirements.

[0267] The two layers are connected through a bidirectional communication channel: the edge node sends processing requests and compressed feature data to the cloud; the cloud returns analysis results and updated model parameters to the edge node.

[0268] The data exchange between the edge and the cloud adopts a unified JSON format, including the following main fields:

[0269] 1. Device ID (uniquely identifies the edge node);

[0270] 2. Timestamp (millisecond precision);

[0271] 3. Sample ID (fabric identification);

[0272] 4. Data type (raw data, feature data, result data);

[0273] 5. Payload (contains different content according to the data type);

[0274] 6. Priority (1 - 10, used for task scheduling);

[0275] 7. Checksum (ensures data integrity);

[0276] The computing tasks in the system can be classified into the following categories:

[0277] 1. Data preprocessing tasks: such as signal filtering, image enhancement, feature extraction, etc.;

[0278] 2. Inference tasks: using the trained model for classification or regression;

[0279] 3. Learning tasks: model parameter update, transfer learning, etc.;

[0280] 4. Decision-making tasks: result fusion, quality assessment, anomaly detection, etc.

[0281] The task scheduling strategy is based on the following objective function:

[0282] ;

[0283] ;

[0284] ;

[0285] Among them, is the execution time of task i on node j, is the energy consumption, is a binary decision variable (1 means task i is assigned to node j, 0 means not assigned), is the resource requirement of task i, is the resource capacity of node j, is the number of tasks, is the number of nodes.

[0286] The task scheduling problem is solved by an improved greedy algorithm, which considers task priorities, dependencies, and resource status, and balances between near-optimal solutions and computational efficiency. The complete pseudocode of the improved greedy algorithm is as follows:

[0287] Input: Task set , Node set

[0288] Output: Task allocation scheme

[0289] 1: / / Initialization

[0290] 2: For each task in :

[0291] 3: Calculate the priority based on urgency, dependencies, and importance

[0292] 4: Sort in descending order of priority

[0293] 5: Initialize 6:

[0295] 7: / / Build the task dependency graph

[0296] 8: Graph 9:

[0298] 10: / / Topological sorting

[0299] 11: List 12:

[0301] 13: / / Resource status tracking

[0302] 14: For each node in : :

[0303] 15: 16:

[0305] 17: / / Main scheduling loop

[0306] 18: For each task in : :

[0307] 19: / / Calculate the cost metric for each node

[0308] 20: For each node in : :

[0309] 21: If :

[0310] 22:

[0311] 23: Otherwise:

[0312] 24: 25:

[0314] 26: / / Find the node with the lowest cost

[0315] 27: 28:

[0317] 29: / / If all nodes are infeasible, try task splitting

[0318] 30: If :

[0319] 31:

[0320] 32: Insert at the appropriate position in

[0321] 33: Continue 34:

[0323] 35: / / Allocate tasks

[0324] 36:

[0325] 37: 38:

[0327] 39: / / Dynamically adjust the priority of subsequent tasks

[0328] 40: 41:

[0330] 42: / / Optimize local load balancing

[0331] 43: For each overloaded node :

[0332] 44: Identify tasks that can be migrated with minimal impact

[0333] 45: Reassign the selected tasks to nodes with lower load 46:

[0335] 47: Return

[0336] In the above pseudocode, the weight parameters , , and control the importance of execution time, energy consumption, network latency, and load balancing respectively. The load balancing factor is defined as the ratio of the current load of node to the system average load.

[0337] Specifically, the system maintains a global task queue and a node resource status table. The scheduler works based on a multi-level priority queue mechanism: Emergency tasks (such as real-time detection requests) enter the high-priority queue; periodic tasks (such as model updates) enter the medium-priority queue; background tasks (such as historical data analysis) enter the low-priority queue.

[0338] The task splitting strategy follows the principles of data locality, computation-data balance, emergency priority, failure recovery, and energy efficiency balance. For long-running complex tasks, a progress tracking and dynamic adjustment mechanism is adopted. When it is detected that the task execution time far exceeds the expectation or the resource usage is abnormal, the system rearranges or migrates the task.

[0339] For computationally intensive tasks such as spectral analysis, the system provides a multi-level processing strategy:

[0340] 1. Lightweight processing: Perform basic preprocessing and feature extraction at the edge node;

[0341] 2. Intermediate processing: Use the lightweight LeNet-5 model on the edge node for preliminary classification;

[0342] 3. Deep processing: For complex or uncertain samples, send the data to the cloud for comprehensive analysis.

[0343] This multi-level processing method dynamically adjusts the processing depth according to the sample complexity and system load, achieving a balance between resource utilization and response time.

[0344] To adapt to the capacity differences of different edge nodes, we adopt the following deployment mechanisms:

[0345] 1. Model layering: Divide the LeNet-5 network into a feature extraction layer and a classification layer, and determine the splitting point according to the node capacity;

[0346] 2. Dynamic quantization: Dynamically adjust the weight precision according to the node computing power and task accuracy requirements;

[0347] 3. Selective computing: For complex samples, execute the complete computing process; for simple samples, use a lightweight model for fast processing;

[0348] 4. Incremental learning: Edge nodes use local data to fine-tune the model to adapt to the local feature distribution.

[0349] Model layering is the key technology for flexible deployment. The LeNet-5 network can be divided into four functional modules: front-end feature extraction, intermediate feature abstraction, feature integration, and decision-making layer. According to the computing power of the edge node, the system can flexibly determine the model splitting point. For example, nodes with strong computing power can execute the complete model; nodes with medium capacity execute the first three modules and send the feature vector to the cloud for final classification; nodes with limited capacity only execute basic feature extraction and send the intermediate features to the cloud for processing.

[0350] The dynamic quantization technology selects different quantization strategies according to the capabilities of edge devices and task requirements:

[0351] 32-bit floating point: For complex tasks with high-precision requirements (such as diluted blood detection);

[0352] 16-bit floating point: General tasks that balance precision and efficiency (such as regular pollution identification);

[0353] 8-bit integer: Suitable for most regular detection tasks, with a good balance between precision and efficiency;

[0354] 4-bit / 2-bit integer: For simple identification tasks (such as obvious pollution detection), giving priority to speed.

[0355] The selection of different quantization levels is based on the system's comprehensive evaluation of the task importance and the current resource status of the node. For example, when the node CPU utilization is below 30%, a higher precision representation can be used; in high-load (>70%) situations, automatically switch to a low-precision representation to ensure system responsiveness. In practical applications, dynamic quantization increases the peak processing capacity of the system by approximately 40% while maintaining the average accuracy drop within 1.5%.

[0356] Incremental learning enables edge nodes to adapt to local environmental characteristics. The specific implementation adopts the following steps:

[0357] 1. Basic model deployment: Deploy the basic model trained in the cloud to each edge node;

[0358] 2. Local data collection: Edge nodes collect and label local data during daily operation;

[0359] 3. Model fine-tuning: Regularly fine-tune the model using local data with a small learning rate (0.0001);

[0360] 4. Forgetting protection: Adopt Elastic Weight Consolidation (EWC) to prevent catastrophic forgetting;

[0361] 5. Performance evaluation: Regularly evaluate the performance of the fine-tuned model to ensure that the accuracy does not decrease;

[0362] 6. Model merging: Regularly feedback the knowledge of local fine-tuning to the cloud global model.

[0363] Incremental learning significantly improves the environmental adaptability of the system. In the cross-department test, the accuracy of the model with local incremental learning on the fabrics of specific departments has increased compared to the basic model, especially for the detection of department-specific pollution types (such as specific drug residues in the operating room, specific body fluids in the pediatrics department, etc.).

[0364] To further reduce communication overhead and data inequality, the present invention introduces a federated learning framework to enable edge nodes to co-train the model. The process is as follows:

[0365] ⒈ The cloud initializes the global model parameters and distributes them to all edge nodes;

[0366] ⒉ Each edge node calculates the gradient using local data and uploads it to the cloud;

[0367] ⒊ The cloud aggregates all gradients and updates the global model: ;

[0368] ⒋ Repeat steps 2 and 3 until the model converges.

[0369] In view of the characteristics of the quality inspection of medical fabric cleaning and disinfection, the invention proposes a federated aggregation algorithm based on credibility:

[0370] ;

[0371] where the weight is related not only to the data volume but also to the historical performance and data quality of the node:

[0372] ;

[0373] in, is a node The historical performance coefficient is based on the historical forecast accuracy; is the data quality coefficient, based on data distribution and diversity. Calculated using the exponentially weighted moving average method:

[0374] ;

[0375] in, Is the current round node The verification accuracy of is the historical performance coefficient of the previous round, is the weighting factor.

[0376] Data quality factor Calculated by evaluating the diversity and balance of data distribution at each node:

[0377] ;

[0378] in, is the data diversity index, is the data balance index. The diversity index is calculated based on the category entropy:

[0379] ;

[0380] in, is a category At the node The proportion of the data, is the total number of categories. The balance index is based on the inverse of the coefficient of variation of the data volume of each category:

[0381] ;

[0382] in, is the coefficient of variation, is the standard deviation of the number of samples in each category, is the average number of samples in each category.

[0383] In this way, the data quality coefficient can evaluate the comprehensiveness and representativeness of the node data and avoid model bias caused by data imbalance or incomplete coverage. The value will be lower, reducing its influence weight in the global model update.

[0384] To reduce the edge-cloud communication overhead, the following techniques are adopted:

[0385] 1. Sparse update: Only transmit the significantly changed parameters, and define the importance metric for parameter updates:

[0386] ;

[0387] where is the change amount of the parameter and is the absolute value of the parameter. Only transmit the parameters with the top-k importance, and the k value is dynamically set to the minimum value that reduces the transmission volume by 90%.

[0388] 2. Quantization compression: The parameter update values are represented with a low bitwidth. The specific quantization scheme is:

[0389] ;

[0390] This scheme only transmits the sign bit and the exponent bit, significantly reducing the number of transmitted bits.

[0391] 3. Asynchronous update: Nodes can accumulate a certain amount of parameter updates locally and then transmit them all at once, reducing the number of communication times. The cumulative update threshold is dynamically adjusted based on the node computing power and network conditions:

[0392] ;

[0393] where and are the weight coefficients, and are the minimum and maximum accumulation thresholds respectively.

[0394] These strategies are applied comprehensively, greatly reducing the communication volume of parameter updates while keeping the model performance almost unchanged.

[0395] Spectral analysis and chip identification are two key technical means for the quality inspection of medical fabric cleaning and disinfection. The former provides direct evidence of the surface substance composition of the fabric, and the latter provides indirect information on the fabric treatment process. The described invention uses a multi-level fusion strategy to organically combine the two technologies.

[0396] There is an inherent correlation between the process data recorded by the RFID chip and the spectral analysis results. Parameters such as washing temperature, time, and disinfectant concentration directly affect the microbial residue amount, and these residues will show characteristic absorption peaks on the spectrum at specific wavelengths. Establishing this correlation model helps improve the reliability and interpretability of detection.

[0397] The correlation model can be expressed as: , where is the theoretical expected value of the spectral characteristics, is a process parameter, and is a model parameter. When there is a significant deviation between the actual spectrum and the theoretical expectation, it may indicate an abnormal processing process or false recording.

[0398] Specifically, the system has established the following correlation models:

[0399] 1. Temperature - protein denaturation relationship model: The washing temperature affects the degree of protein denaturation, which in turn affects the characteristics of the near-infrared spectrum in the 1450 - 1650 nm region. This relationship can be expressed by the Arrhenius equation:

[0400] ;

[0401] where, is the protein denaturation rate constant, is the pre-exponential factor, is the activation energy, is the gas constant, is the absolute temperature. From this, the change in spectral intensity can be deduced:

[0402] ;

[0403] where, is the wavelength at which the spectral intensity is, is the initial spectral intensity, is the processing time. In this model, the value of blood protein is 287 kJ / mol, is s^-1.

[0404] 2. Disinfectant - residue relationship model: Different disinfectants have characteristic absorption peaks at specific wavelengths. The model uses the modified Beer-Lambert law:

[0405] ;

[0406] where, is the absorbance at wavelength , is the molar absorptivity of substance at this wavelength, is the concentration of substance , is the optical path length, is the background absorption. The value of quaternary ammonium salt disinfectant at 1450 nm is 0.012 L·mol^-1·cm^-1, and the The value is 8500 - 9200 L·mol^-1·cm^-1.

[0407] 3. Time - microbial inactivation relationship model: The treatment time affects the microbial survival rate, and the logarithmic decay model is adopted:

[0408] ;

[0409] where is the number of surviving microorganisms, is the initial number of microorganisms, is the inactivation rate constant, is the time, is the shape parameter. The microbial load has a logarithmic correlation with the scattering intensity at 1550 nm of the near - infrared spectrum as follows:

[0410] ;

[0411] where and are calibration constants. For common hospital flora, , .

[0412] 4. Process integrity relationship model: Constructed based on the Bayesian network, define the set of process variables and the set of spectral features , and the conditional probability is expressed as:

[0413] ;

[0414] where is the set of parent nodes of the feature in the Bayesian network, including the process variables that directly affect this feature. The key dependency relationships include: the washing temperature affects the protein spectrum, the rinsing times affect the residue spectrum, and the disinfection method affects the microbial load spectrum.

[0415] The fusion of spectral data and chip data adopts a multi - level strategy:

[0416] 1. Feature - level fusion: Combine spectral features and process parameters into an extended feature vector and input it into the LeNet - 5 network for comprehensive analysis;

[0417] 2. Decision - level fusion: Make preliminary judgments based on the spectrum and process parameters respectively, and then fuse the results through weighted voting or Dempster - Shafer evidence theory;

[0418] 3. Model - level fusion: Design a dual - input network, and the spectral data and process parameters pass through specific processing paths respectively and are fused in the deep layer of the network;

[0419] 4. Causal Fusion: Establish a Bayesian network to capture the causal relationships among process parameters, physicochemical changes, and spectral features.

[0420] The fusion weights can be solved through the following optimization problem:

[0421] ;

[0422] ;

[0423] where is the weight of the j-th model, is the prediction of the j-th model for sample i, is the true label, is the regularization parameter, and s.t. means satisfying the conditions that follow.

[0424] In practice, the system adopts a dynamic fusion strategy to adjust the fusion weights in real time according to data availability and quality:

[0425] ;

[0426] where is the dynamic weight of the j-th model at time t, is the data quality index, is the model confidence.

[0427] To verify the effectiveness of the invention, a series of experiments were designed, including dataset construction, evaluation metric definition, comparative experiments, and ablation experiments, etc.

[0428] We collected data on different types of medical fabrics, including spectral data, image data, process parameters, and microbial test results. The dataset was divided into a training set, a validation set, and a test set in a ratio of 7:2:1. To increase sample diversity, the dataset included fabrics from different departments and different types of contamination.

[0429] To verify the performance of the system under different conditions, the following special test sets were specifically constructed:

[0430] 1. Low-concentration contamination set: Contains highly diluted (1:500 to 1:2000) contaminants such as blood, drugs, etc.;

[0431] 2. Multiple contamination set: Each fabric has 2 - 5 different types of contaminants;

[0432] 3. Environmental interference set: Data collected under different environmental conditions (high humidity, low temperature, etc.);

[0433] 4. Process anomaly set: Samples with abnormal process parameters during washing or disinfection.

[0434] The annotation of all samples was jointly completed by experienced technicians from the Central Sterile Supply Department and infection control experts. Each sample was independently inspected by at least two experts to ensure the quality of annotation. For numerical annotations such as pollutant concentration, standard laboratory methods (such as enzyme-linked immunosorbent assay, high performance liquid chromatography, etc.) were used for quantitative analysis to ensure the accuracy of annotation.

[0435] The joint algorithm proposed by the present invention was compared with the following benchmark methods:

[0436] 1. Traditional method: ATP bioluminescence detection + visual inspection;

[0437] 2. Single machine learning methods: Support Vector Machine (SVM), Random Forest (RF);

[0438] 3. Deep learning methods: CNN, RNN, traditional HMM-Viterbi;

[0439] The experimental results are shown in Table 1 below:

[0440] ;

[0441] Note: Accuracy, precision and recall are based on the mean ± standard deviation of the test set; the average detection time includes data acquisition and processing time;

[0442] As can be seen from Table 1, the joint algorithm proposed by the present invention is superior to the control methods in terms of both accuracy and efficiency.

[0443] To analyze the contribution of each component to the system performance, ablation experiments were conducted by sequentially removing or replacing each key component:

[0444] 1. Replace VMF with the traditional Viterbi algorithm;

[0445] 2. Replace LeNet-5 with a simple CNN or traditional feature extraction method;

[0446] 3. Remove the edge learning framework and change it to centralized processing;

[0447] 4. Only use spectral data without combining RFID chip information;

[0448] 5. Analysis of the contribution of different spectral bands.

[0449] The experimental results are shown in Table 2:

[0450] ;

[0451] Note: The complete algorithm refers to all the technical solutions of the present invention; other rows represent the performance after replacing or removing specific components; the values are mean ± standard deviation.

[0452] As can be seen from Table 2, each component contributes to the system performance to varying degrees:

[0453] 1. After replacing the VMF algorithm with the traditional Viterbi algorithm, the accuracy slightly decreases, but the processing time significantly increases. This indicates that the VMF algorithm greatly improves the computational efficiency while maintaining the accuracy.

[0454] 2. After replacing LeNet-5 with a simple CNN, the accuracy drops by 4.3 percentage points and the processing time slightly decreases. This shows that LeNet-5 has obvious advantages in feature extraction ability, especially for fine features in spectral data.

[0455] 3. After removing the edge learning framework and changing to centralized processing, the accuracy basically remains the same, but the processing time increases by nearly 4 times. This verifies the important role of the edge learning framework in improving the system real-time performance and reducing the communication burden.

[0456] 4. When only using spectral data, the accuracy drops by 7.5 percentage points; when only using chip data, it drops by 19.2 percentage points. This indicates that data fusion brings significant performance improvement, and the two modal data have strong complementarity.

[0457] 5. The comparison of different spectral bands shows that the near-infrared band (900 - 1700 nm) performs better than the mid-infrared band (2500 - 5000 nm), but both are inferior to the full spectral range.

[0458] Further analysis of the performance differences of specific tasks reveals that spectral data contributes the most to the identification of organic matter pollution (such as blood and body fluids), and chip data is more effective in detecting process parameter-related problems (such as temperature damage to thermosensitive materials). This shows that multimodal fusion not only improves the overall accuracy but also enhances the comprehensive detection ability of the system, enabling it to handle various quality problems.

Claims

1. A method for cleaning, disinfecting and inspecting medical fabrics, characterized in that, It includes the following steps: Data acquisition step: Through a sensor network deployed at multiple points in the medical fabric processing flow, multi-modal data of the medical fabric to be inspected is obtained. The multi-modal data includes spectral data reflecting the chemical composition of the fabric surface collected by a spectral analyzer, image data reflecting the physical appearance of the fabric collected by a camera, and chip data read by an RFID reader, stored in the internal chip of the fabric, and recording the processing history and process parameters of the fabric; Edge processing step: Initial processing is performed on the collected multi-modal data locally on the detection device, including data preprocessing, key feature extraction, and real-time analysis of the converted spectral data using a lightweight convolutional neural network deployed at the edge node to achieve preliminary identification and classification of common pollutants and residues; Cloud analysis step: The data or complex sample information after edge processing is uploaded to the cloud platform. Using the improved Viterbi algorithm deployed in the cloud, combined with the process parameters recorded by the chip and the spectral analysis results, state sequence inference and quality assessment of the cleaning and disinfection process of the medical fabric are carried out. At the same time, a full-size convolutional neural network is used to perform in-depth analysis of the sample, execute historical data mining, and model iterative update. The edge node and the cloud platform work together based on the edge learning framework to dynamically allocate computing tasks; Application management step: The comprehensive quality assessment results are displayed through the user interface, providing unqualified alarms, data statistical analysis, and decision support information to achieve full-process intelligent management and sorting decision-making for the cleaning and disinfection quality of medical fabrics.

2. The medical fabric cleaning, disinfection and inspection method according to claim 1, wherein, The improved Viterbi algorithm in the cloud analysis step realizes the inference of the state sequence of the medical fabric cleaning and disinfection process in the following way: Define a unified multi-dimensional state space representing the quality of the medical fabric, which is jointly composed of the cleaning state dimension, disinfection state dimension, chemical residue state dimension, and fabric integrity state dimension of the fabric; Construct a comprehensive observation space, which integrates the spectral feature vector from the spectral analyzer, the image-derived features from the camera, the detailed process control parameters read from the RFID chip, and the relevant environmental monitoring data as the direct evidence input for the algorithm to perform state inference; Adopt a search strategy based on dynamic programming and combine an evaluation function to determine the most likely fabric state transition sequence. The evaluation function includes the cost of the actually occurred path and the heuristic cost of the estimated future path. The calculation of the heuristic cost comprehensively considers the compliance analysis of the process parameters recorded by the chip with the preset standard operating procedures, the statistical analogy inference of the processing results of similar samples in the historical database, and the guidance of the encoded domain expert knowledge rule system for the rationality of state transitions, so as to significantly improve the search efficiency of the algorithm in the complex state space while ensuring the inference accuracy.

3. A medical fabric cleaning, disinfection and inspection method according to claim 1, characterized in that, The process of using a convolutional neural network to analyze the spectral data in the edge processing step and the cloud analysis step includes: Systematically preprocess the collected one-dimensional raw spectral data, including filtering, baseline correction, and normalization, and then convert the preprocessed one-dimensional spectral data into a two-dimensional structured representation; Construct a multi-channel input mechanism to organize spectral information from the same spectral measurement but processed by different mathematical transformations or representing different physical meanings into multiple parallel input channels, jointly forming a multi-channel two-dimensional spectral image to provide a richer feature dimension for the convolutional neural network; Adopt an adaptively adjusted convolutional neural network structure. The input layer of this network receives the aforementioned multi-channel two-dimensional spectral image. Its main structure includes several alternately stacked two-dimensional convolutional layers and max-pooling layers. The convolutional layers are used to hierarchically extract local and abstract features in the spectral image, and the max-pooling layers are used to enhance the network's sensitivity to spectral feature peaks and achieve data dimensionality reduction. The deep layer of the network is connected with fully connected layers for feature integration and final classification or regression output, and techniques aiming to improve the model's performance and stability are applied during training and inference; Train the convolutional neural network through supervised learning. During the training process, data augmentation methods designed for the characteristics of spectral data are used to expand sample diversity and improve the model's generalization ability. For the network version deployed on edge processing nodes, model compression techniques are also applied to reduce the model size and computational complexity, ensuring its efficient operation in resource-constrained environments.

4. A medical fabric cleaning, disinfection and inspection method according to claim 1, characterized in that, In the cloud analysis step, the cloud and edge nodes cooperate based on the edge learning framework, and the fusion of spectral data and chip data is achieved through the following mechanisms: The system is built and runs on an edge-cloud collaborative distributed computing architecture, where edge processing nodes are responsible for performing real-time analysis and preliminary decision-making near the data source, and the cloud platform undertakes complex computing tasks, training and maintenance of global models, and large-scale data management; Implement dynamic task scheduling and resource management strategies. This strategy intelligently determines whether a computing task is to be executed locally on an edge node, migrated between edge nodes, or submitted to the cloud platform for processing based on task characteristics, data locality, network conditions, and the real-time load and capabilities of each computing node; Adopt model adaptive deployment and collaborative learning techniques to support the differential deployment of model components or adjustment of model accuracy according to the capabilities of edge nodes, and establish a collaborative model training and update mechanism between edge nodes and the cloud platform, allowing edge nodes to fine-tune the model using local data and safely contribute learning results to the iteration of the global model; Achieve multi-level intelligent fusion of spectral data and process parameter data recorded by the chip, including merging the features extracted from the two types of data at the feature level and inputting them into the analysis model, integrating the preliminary judgment results based on their respective data sources at the decision level, designing a dedicated network structure at the model level that can directly process and fuse the two types of heterogeneous inputs, and establishing and utilizing the internal correlation model between process parameters and expected spectral features.

5. A medical fabric cleaning, disinfection and inspection device, characterized in that, Including: The data acquisition module is configured with a spectral analyzer, a camera, and an RFID reader / writer. The spectral analyzer is used to collect spectral data reflecting the surface chemical composition of medical fabrics. The camera is used to collect image data reflecting the physical appearance of the fabrics. The RFID reader / writer is used to read the processing history and process parameter information stored in the internal chip of the fabric, jointly constituting a multi-modal data source; The edge processing module is set locally in the detection device, with a built-in processing unit and a lightweight convolutional neural network deployed. The edge processing module is connected to the data acquisition module and is used to perform local preprocessing, feature extraction on the collected multi-modal data, and perform real-time analysis on the spectral data to complete the preliminary identification of common pollutants and residues; The cloud analysis module includes a server-level processing unit and is deployed with an improved Viterbi algorithm and a full-size convolutional neural network. The cloud analysis module communicates with the edge processing module through the network and is used to perform in-depth analysis on complex sample data, infer the state sequence and quality assessment of the cleaning and disinfection process based on process parameters and spectral results, and is responsible for the training, management, and update of the global model, and collaborates with the edge processing module under the edge learning framework for task allocation and processing; The application management module includes a user interface and a management function unit, and is connected to the cloud analysis module, and is used to display detailed quality assessment results to users, issue unqualified alarms, provide data statistical analysis reports, and support management decisions.

6. The medical fabric cleaning, disinfection and inspection device according to claim 5, wherein, The improved Viterbi algorithm deployed in the cloud analysis module is configured to accurately infer the state sequence of the medical fabric cleaning and disinfection process in the following ways: First, a unified multi-dimensional state space definition of medical fabrics is constructed and maintained inside the cloud analysis module. This state space is jointly characterized by multiple key quality dimensions including the cleaning state, disinfection state, chemical residue state, and fabric integrity state of the fabric. Each dimension is refined into multiple discrete levels to comprehensively depict various quality conditions that the fabric may present in the cleaning and disinfection process; Second, the cloud analysis module configures the improved Viterbi algorithm to receive and process a comprehensive observation data vector, which is formed by fusing spectral feature data provided by the data acquisition module, image-derived feature data, complete process parameter data recorded by the chip, and relevant environmental monitoring data, as the direct input basis for state inference; Then, the cloud analysis module uses the improved Viterbi algorithm to find the most likely fabric state evolution path in the multi-dimensional state space through an optimization search mechanism that combines the actual path cost and forward-looking heuristic evaluation. Among them, the heuristic evaluation comprehensively considers the deviation degree of actual process parameters from the standard procedures, the prediction of similar case results based on historical big data analysis, and the judgment of the rationality of state transitions by the expert system rules embedded in the algorithm logic, so as to efficiently guide the search direction and ensure the accuracy of the inference results.

7. The medical fabric cleaning, disinfection and inspection device according to claim 5, characterized in that, Through the collaborative working mechanism of the edge learning framework between the cloud analysis module and the edge processing module, as well as their ability to fuse and process spectral data and chip data, efficient distributed intelligent analysis and decision-making are further achieved through the following configurations: The device is constructed and runs on an edge-cloud hierarchical collaborative computing architecture, where the edge processing module performs immediate analysis and preliminary judgment near the data source, and the cloud analysis module performs complex calculations, global model management, and in-depth analysis; The device integrates dynamic task scheduling and resource management functions, and intelligently allocates and schedules computing tasks between the edge and the cloud according to task requirements, data characteristics, and the status of each computing node to optimize the overall processing efficiency of the system; The device supports the adaptive deployment of the model at the edge node and collaborative learning with the cloud platform, including the on-demand adjustment of model parameters, the model fine-tuning by the edge node using local data, and the secure aggregation and feedback update of the learning results of each node to the global model, and adopts communication optimization strategies to reduce the data transmission load; The device realizes multi-strategy intelligent fusion of spectral data and process parameter data recorded by the chip. Through integrated analysis at the feature level, decision level, and model level, and using the inherent correlation between the established process parameters and expected spectral features for cross-validation, to improve the comprehensiveness, accuracy, and reliability of the quality assessment of medical fabric cleaning and disinfection. At the same time, it supports dynamic fusion weight adjustment based on data quality and model confidence.

Images