An end-cloud cooperation-based predictive maintenance method for medical purification equipment
By employing a predictive maintenance approach that integrates edge and cloud computing, and utilizing an improved TS2Vec model to predict the health status and failure risk of medical purification equipment, this approach addresses the problem of traditional maintenance methods being insensitive to hidden degradation. It enables early identification and proactive maintenance, thereby reducing equipment failure risks and operating costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-08
- Publication Date
- 2026-07-03
AI Technical Summary
Existing maintenance methods for medical purification equipment cannot effectively identify hidden degradation of equipment purification performance and multi-factor coupled failures, resulting in delayed maintenance and the risk of sudden downtime. Traditional periodic maintenance and threshold alarm methods lack sensitivity.
A predictive maintenance method based on edge-cloud collaboration is adopted. Multi-source sensor data is collected and preprocessed at the device end to generate edge feature packages and upload them to the cloud. A historical feature database is established, and an improved TS2Vec model is used to predict health, failure risk and remaining life. Maintenance strategies are generated and distributed to the device end for online reasoning and maintenance closed-loop actions.
It enables early identification of degradation trends in key components of medical purification equipment, reduces the risk of sudden failure of purification performance, reduces unnecessary maintenance costs and downtime, and improves the reliability of equipment operation and the ability to continuously meet purification standards.
Smart Images
Figure CN122337534A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent operation and maintenance of medical equipment and IoT-based edge-cloud collaboration, and particularly to a predictive maintenance method for medical purification equipment based on edge-cloud collaboration. Background Technology
[0002] Medical purification equipment is widely used in clean areas such as hospital operating rooms, intensive care units, negative pressure isolation wards, emergency infusion areas, and laboratory and pharmaceutical preparation areas. It performs functions such as air filtration, aerosol control, disinfection and sterilization, constant air exchange, and differential pressure control to ensure the medical environment meets predetermined cleanliness or anti-diffusion requirements. This type of equipment typically includes fans or pumps, filter assemblies, disinfection modules, valves and duct structures, as well as various sensors and control units. The equipment is characterized by continuous operation, large load fluctuations, hidden component degradation, and high safety risks. Once the purification performance deteriorates or critical components malfunction, it may lead to substandard cleanliness levels, failure of isolation differential pressure, and increased risk of pathogen spread. Therefore, highly reliable, low-latency, and predictable maintenance of medical purification equipment is of great importance.
[0003] The operation and maintenance methods of existing medical purification equipment are mainly divided into two categories: periodic maintenance and threshold-based condition maintenance. Periodic maintenance is usually carried out according to the equipment manual or experience cycle, such as periodic replacement of filter cartridges, periodic maintenance of fans, and periodic inspection of disinfection module output. However, due to the significant influence of factors such as departmental traffic, pollution load, seasonal environment, and switching of operating modes, the fixed cycle does not match the actual degradation process, which can easily lead to over-maintenance or delayed maintenance. Threshold-based condition maintenance relies on setting alarm thresholds for single or a few sensor indicators. When the indicators exceed the limits, maintenance is triggered. However, this method lacks sensitivity to early degradation and multi-factor coupled faults. It often fails to identify trends in the hidden stages such as gradual increase in filtration resistance, gradual decrease in fan efficiency, and slow decay of disinfection intensity. It can only passively alarm after performance has obviously deteriorated or faults have occurred, thus leading to delayed maintenance and the risk of sudden downtime. Summary of the Invention
[0004] One objective of this invention is to propose a predictive maintenance method for medical purification equipment based on edge-cloud collaboration. This invention fully utilizes multi-source sensing and edge reasoning at the equipment end, and the construction and improvement of the TS2Vec model using a historical feature database in the cloud. It performs health assessment, fault risk prediction, and remaining life estimation for the degradation of key components and the decline in purification performance of medical purification equipment. The prediction results are then used to generate maintenance strategies and distribute them to the equipment end to form a closed loop of early warning and maintenance. This method has the advantages of strong real-time performance, high prediction accuracy, low dependence on fault samples, strong edge-cloud collaborative update capability, and low maintenance cost.
[0005] A predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to an embodiment of the present invention includes the following steps:
[0006] Collect and preprocess multi-source sensor data from medical purification equipment;
[0007] Feature vectors are extracted from the preprocessed multi-source sensor data to generate edge-side feature packets;
[0008] The edge-side feature packets are uploaded to the cloud platform through the edge-cloud communication link, and a historical feature database is established on the cloud platform according to device identification and time series.
[0009] Training samples are constructed based on a historical feature database and input into the improved TS2Vec model to output health indicators, failure risk indicators and remaining life prediction results.
[0010] A set of maintenance strategies is generated on the cloud platform based on health indicators, failure risk indicators, and remaining life prediction results.
[0011] The improved TS2Vec model corresponding edge-side inference parameters and maintenance strategy set will be distributed to the device.
[0012] Online reasoning is performed on the real-time collected edge data at the device end to obtain edge health and edge risk assessment results;
[0013] Based on the edge health status, edge risk assessment results, and maintenance strategy set, a closed-loop maintenance action is executed, early warning information and maintenance event logs are output, and the maintenance event logs and corresponding edge feature packets are sent back to the cloud platform to update the historical feature database.
[0014] Optionally, the multi-source sensor data includes air volume data, pressure difference data before and after the filter element, particulate matter concentration, particle count data, temperature and humidity data, operating status data of key actuators, and electrical parameter data. The preprocessing includes noise reduction, outlier removal, and missing value completion.
[0015] Optionally, the generation of the end-side feature packet specifically includes:
[0016] The preprocessed multi-source sensor data is divided into time window data segments according to a pre-set time window length and sliding step size.
[0017] For each sensor variable sampling sequence within each time window data segment, calculate the time domain statistical characteristics, which include the mean, standard deviation, root mean square, skewness, and kurtosis of the sensor variable within the time window;
[0018] For each sensor variable sampling sequence within each time window data segment, a discrete Fourier transform is performed to obtain the corresponding frequency domain amplitude sequence. Based on the frequency domain amplitude sequence, frequency domain features are calculated. The frequency domain features include frequency domain energy and spectral centroid. The frequency domain energy is the cumulative value of the squares of the amplitudes at each frequency point in the frequency domain amplitude sequence. The spectral centroid is the ratio of the cumulative value of the product of each frequency point frequency and its corresponding amplitude to the cumulative value of the amplitudes at each frequency point.
[0019] Based on the sensor variables of key components degrading in the time window data segment, derived indices are constructed. The derived indices include the pressure difference across the filter element, the pressure difference change rate, the fan load factor, and the purification efficiency index. The pressure difference across the filter element is the difference between the average upstream pressure and the average downstream pressure of the filter element. The pressure difference change rate is the difference between the pressure difference across the filter element in the current time window and the pressure difference across the filter element in the previous time window, divided by the sliding step size. The fan load factor is the ratio of the average fan or pump current to the average fan speed. The purification efficiency index is a value minus the ratio of the average outlet particulate matter concentration or particle count to the average inlet particulate matter concentration or particle count.
[0020] The time-domain statistical features, frequency-domain features, and derived indicators are concatenated to form a feature vector of key component degradation and purification performance attenuation. The feature vector is then encapsulated with the corresponding timestamp to generate an end-side feature package.
[0021] Optionally, the establishment of the historical feature database specifically includes:
[0022] On the device side, a unique device identifier is set for each medical purification device, and the end-side feature packet is bound with the device identifier and the corresponding collection timestamp to form a feature unit to be uploaded;
[0023] On the device side, the feature units to be uploaded are batch-collected and encapsulated according to a preset upload cycle to generate upload batches;
[0024] The uploaded batch is sent to the cloud platform via the end-to-cloud communication link and decapsulated. Integrity verification and timestamp consistency verification are performed on each feature unit to be uploaded to obtain valid feature units.
[0025] A corresponding historical feature database is established for each device identifier on the cloud platform. The establishment process involves the cloud creating an independent storage entry for each device as a historical feature database based on the device identifier, and writing the received valid feature units into the historical feature database of the corresponding device according to the timestamp.
[0026] Optionally, the output of the health index, failure risk index, and remaining life prediction results specifically includes:
[0027] The system extracts feature vector sequences arranged in ascending order of timestamps from a historical feature database to form training samples for the corresponding devices. These samples are then input into the improved TS2Vec model. The improved TS2Vec model includes a multi-scale time series construction and enhancement module, a time series feature encoding module, a hierarchical comparative representation learning module, and a maintenance target joint prediction output module. The multi-scale time series construction and enhancement module segments the feature vector sequences in the training samples according to preset multi-scale pruning rules to form a set of positive sample pairs. The time series feature encoding module performs time-step encoding and outputs latent representation subsequences. The hierarchical comparative representation learning module introduces a reversible normalized flow hard negative sample mining mechanism to obtain the device time series representation vector. The maintenance target joint prediction output module inputs the device time series representation vector into a multi-task prediction head and outputs health indicators, fault risk indicators, and remaining life prediction results.
[0028] In the multi-scale time series construction and enhancement module, the feature vector sequence in the training samples is segmented according to the preset multi-scale pruning rules to obtain a set of sub-sequences at different time scales. The first time series enhancement operator and the second time series enhancement operator are applied to each sub-sequence in the set of sub-sequences to generate two corresponding enhanced sub-sequences, forming a set of positive sample pair sub-sequences.
[0029] In the temporal feature encoding module, the two enhanced subsequences in the set of positive sample pair subsequences are encoded step by step over time, and the hidden representation subsequences corresponding one-to-one with the two enhanced subsequences are output.
[0030] In the hierarchical contrastive representation learning module, a reversible normalized flow hard negative sample mining mechanism is introduced to estimate the probability density of the latent representation sequence at the current scale. Based on the probability density estimate, a set of hard negative samples related to positive sample pairs is selected from the latent representation sequences of devices in the same batch. Hierarchical consistency alignment and time-step contrast calculation are performed on two latent representation sequences to obtain the contrast loss at the corresponding time scale. The contrast losses of all time scales and all time steps are accumulated to obtain the hierarchical contrast loss.
[0031] The temporal feature encoding module is self-supervised and optimized based on hierarchical contrastive loss, and the output device temporal representation vector corresponding to each training sample is generated.
[0032] In the joint prediction output module for maintenance targets, the equipment time-series representation vector is input into the multi-task prediction head. The multi-task prediction head includes a health prediction head, a fault risk prediction head, and a remaining lifetime prediction head. The health prediction head performs a first mapping on the equipment time-series representation vector to obtain a raw health prediction value, and then performs a normalized mapping on the raw health prediction value to obtain a health index. The fault risk prediction head performs a second mapping on the equipment time-series representation vector to obtain a raw fault risk prediction value, and then performs a risk probability mapping on the raw fault risk prediction value to obtain a fault risk index. The remaining lifetime prediction head performs a third mapping on the equipment time-series representation vector to obtain a raw remaining lifetime prediction value, and then performs a preset lifetime calibration mapping on the raw remaining lifetime prediction value to obtain a remaining lifetime prediction result.
[0033] Optionally, the generation of the maintenance strategy set specifically includes:
[0034] On the cloud platform, each device is assigned a preset health trigger threshold, risk trigger threshold, and lifespan trigger threshold. Maintenance trigger conditions are generated based on the health index, fault risk index, and remaining lifespan prediction result corresponding to the current time window. The maintenance trigger conditions refer to the health index being less than the health trigger threshold, the fault risk index being greater than the risk trigger threshold, or the remaining lifespan prediction result being less than the lifespan trigger threshold.
[0035] When the maintenance trigger condition is determined to be triggered, the maintenance timing parameters are obtained based on the timestamp corresponding to the current time window and the remaining lifetime prediction result. The maintenance timing parameters include the maintenance start time and the maintenance end time. The maintenance start time is the timestamp corresponding to the current time window, and the maintenance end time is the time when the timestamp corresponding to the current time window is superimposed with the estimated runnable time corresponding to the remaining lifetime prediction result.
[0036] When the maintenance trigger condition is determined to be triggered, maintenance content parameters are generated based on health indicators, fault risk indicators and remaining life prediction results. The maintenance content parameters include a set of filter maintenance items, a set of fan or pump maintenance items and a set of disinfection module maintenance items. When the maintenance trigger condition is not determined to be triggered, the maintenance content parameters are an empty set.
[0037] On the cloud platform, maintenance trigger conditions, maintenance timing parameters, and maintenance content parameters are combined according to device identifier and time window sequence number to form a set of maintenance strategies.
[0038] Optionally, the distribution of the edge-side inference parameters and maintenance strategy set specifically includes:
[0039] On the cloud platform, for each device identifier, the edge inference parameters of the improved TS2Vec model corresponding to the device identifier are retrieved, and the set of maintenance strategies corresponding to the device identifier is also retrieved.
[0040] On the cloud platform, the edge inference parameters and maintenance strategy set are encapsulated according to a preset data structure to generate a data packet that corresponds one-to-one with the device identifier, and a data packet generation timestamp is added to the data packet.
[0041] The data packets are sent to the device corresponding to the device identifier via the end-to-cloud communication link.
[0042] Optionally, obtaining the endpoint health status and endpoint risk assessment results specifically includes:
[0043] At the device end, perform end-side feature extraction on the real-time acquired end-side data to obtain the feature vector corresponding to the time window;
[0044] On the device side, the feature vector is input into the edge inference network corresponding to the improved TS2Vec model. The feature vector is then mapped to health and risk, and the original predicted values of edge health and edge fault risk are output.
[0045] On the device side, the original predicted value of the edge health is normalized to obtain the edge health value. The edge health value is then compared with the health trigger threshold issued by the cloud. On the device side, the original predicted value of the edge fault risk is compared with the risk trigger threshold issued by the cloud to obtain the edge risk judgment result. The comparison process is as follows: when the original predicted value of the edge fault risk is greater than the risk trigger threshold, the edge risk judgment result is in a triggered state; when the original predicted value of the edge fault risk is less than the risk trigger threshold, the edge risk judgment result is in a non-triggered state.
[0046] Optionally, updating the historical feature database specifically includes:
[0047] Based on the maintenance trigger conditions in the maintenance strategy set, trigger determination is made on the edge health and edge risk assessment results. When the determination is triggered, maintenance closed-loop action is executed; when the determination is not triggered, maintenance closed-loop action in the current time window is terminated.
[0048] When the state is determined to be triggered, the device generates an early warning information based on the maintenance timing parameters and the maintenance content parameters in the maintenance strategy set. The early warning information includes the device identifier, the triggered maintenance conditions, the terminal health corresponding to the current time window, the terminal risk judgment result corresponding to the current time window, and the corresponding maintenance content parameters.
[0049] The maintenance closed-loop actions in the triggered state are recorded to form a maintenance event log. The maintenance event log includes the device identifier, the timestamp of the trigger time, the terminal health status, the terminal risk assessment result, the maintenance triggering conditions, the maintenance timing parameters, the maintenance content parameters, and the corresponding warning information.
[0050] The device binds and encapsulates the maintenance event log with the edge feature packet corresponding to the trigger time, and sends it back to the cloud platform through the edge-cloud communication link to update the historical feature database.
[0051] The beneficial effects of this invention are:
[0052] This invention collects and preprocesses multi-source sensor data from medical purification equipment at the device end, extracts feature vectors to generate end-side feature packages, and uploads them to the cloud to establish a historical feature database. This allows for a continuous and structured characterization of the equipment's operating status and purification performance degradation process. Training samples are constructed in the cloud based on the historical feature database and input into an improved TS2Vec model. This enables self-supervised multi-scale temporal representation learning of unlabeled or poorly labeled operating sequences. Furthermore, a joint prediction output module for maintenance targets simultaneously obtains health indicators, failure risk indicators, and remaining life prediction results. This allows for the reliable early identification of degradation trends in key components such as filters, fans or pumps, and disinfection modules. This overcomes the shortcomings of existing periodic maintenance and single-threshold maintenance methods, which are insensitive to hidden degradation and have delayed early warnings, significantly reducing the risk of sudden purification performance failure in medical settings.
[0053] This invention further utilizes health indicators, fault risk indicators, and remaining life prediction results to generate a set of maintenance strategies. The improved TS2Vec model's corresponding edge-side inference parameters and the set of maintenance strategies are then distributed to the device. The device performs online inference to obtain edge-side health and risk assessment results. When maintenance trigger conditions are met, it executes a closed-loop maintenance action, generates a maintenance event log, and sends it back to the cloud for model updates, forming an edge-cloud collaborative prediction-early warning-pre-maintenance closed loop. Thus, the highly complex part of the prediction calculation is completed in the cloud, while the edge only performs lightweight inference and immediate triggering. This ensures that medical purification equipment still has low-latency early warning capabilities even under network fluctuations or bandwidth limitations, and utilizes historical data from across devices in the cloud to continuously iterate the model, improving the model's generalization adaptability and stability to different departmental loads, different operating modes, and environmental interference.
[0054] Therefore, without increasing the burden of additional manual inspections, this invention transforms maintenance from passive repair after failure to proactive maintenance based on remaining lifespan and risk trends, reducing unnecessary premature replacements and downtime maintenance, lowering operation and maintenance costs and spare parts waste, while improving the continuous compliance of medical purification equipment with purification effects and operational reliability in key scenarios, meeting the requirements of medical environments for safety, continuity and traceable operation and maintenance. Attached Figure Description
[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0056] Figure 1 This is an overall flowchart of a predictive maintenance method for medical purification equipment based on edge-cloud collaboration proposed in this invention.
[0057] Figure 2 This is a schematic diagram illustrating the construction of the edge-side feature package for a predictive maintenance method for medical purification equipment based on edge-cloud collaboration proposed in this invention.
[0058] Figure 3 This is a schematic diagram of the structure of the improved TS2Vec model of the predictive maintenance method for medical purification equipment based on edge-cloud collaboration proposed in this invention. Detailed Implementation
[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0060] refer to Figures 1-3 A predictive maintenance method for medical purification equipment based on edge-cloud collaboration includes the following steps:
[0061] Collect and preprocess multi-source sensor data from medical purification equipment;
[0062] Feature vectors are extracted from the preprocessed multi-source sensor data to generate edge-side feature packets;
[0063] The edge-side feature packets are uploaded to the cloud platform through the edge-cloud communication link, and a historical feature database is established on the cloud platform according to device identification and time series.
[0064] Training samples are constructed based on a historical feature database and input into the improved TS2Vec model to output health indicators, failure risk indicators and remaining life prediction results.
[0065] A set of maintenance strategies is generated on the cloud platform based on health indicators, failure risk indicators, and remaining life prediction results.
[0066] The improved TS2Vec model corresponding edge-side inference parameters and maintenance strategy set will be distributed to the device.
[0067] Online reasoning is performed on the real-time collected edge data at the device end to obtain edge health and edge risk assessment results;
[0068] Based on the edge health status, edge risk assessment results, and maintenance strategy set, a closed-loop maintenance action is executed, early warning information and maintenance event logs are output, and the maintenance event logs and corresponding edge feature packets are sent back to the cloud platform to update the historical feature database.
[0069] In this embodiment, the multi-source sensor data includes air volume data, pressure difference data before and after the filter element, particulate matter concentration, particle count data, temperature and humidity data, operating status data of key actuators, and electrical parameter data. The preprocessing includes noise reduction, outlier removal, and missing value completion.
[0070] In this embodiment, the generation of the end-side feature packet specifically includes:
[0071] The preprocessed multi-source sensor data is divided into time window data segments according to a pre-set time window length and sliding step size.
[0072] For each sensor variable sampling sequence within each time window data segment, calculate the time domain statistical characteristics, which include the mean, standard deviation, root mean square, skewness, and kurtosis of the sensor variable within the time window;
[0073] For each sensor variable sampling sequence within each time window data segment, a discrete Fourier transform is performed to obtain the corresponding frequency domain amplitude sequence. Based on the frequency domain amplitude sequence, frequency domain features are calculated. The frequency domain features include frequency domain energy and spectral centroid. The frequency domain energy is the cumulative value of the squares of the amplitudes at each frequency point in the frequency domain amplitude sequence. The spectral centroid is the ratio of the cumulative value of the product of each frequency point frequency and its corresponding amplitude to the cumulative value of the amplitudes at each frequency point.
[0074] Based on the sensor variables of key components degrading in the time window data segment, derived indices are constructed. The derived indices include the pressure difference across the filter element, the pressure difference change rate, the fan load factor, and the purification efficiency index. The pressure difference across the filter element is the difference between the average upstream pressure and the average downstream pressure of the filter element. The pressure difference change rate is the difference between the pressure difference across the filter element in the current time window and the pressure difference across the filter element in the previous time window, divided by the sliding step size. The fan load factor is the ratio of the average fan or pump current to the average fan speed. The purification efficiency index is a value minus the ratio of the average outlet particulate matter concentration or particle count to the average inlet particulate matter concentration or particle count.
[0075] The time-domain statistical features, frequency-domain features, and derived indicators are concatenated to form a feature vector of key component degradation and purification performance attenuation. The feature vector is then encapsulated with the corresponding timestamp to generate an end-side feature package.
[0076] In this embodiment, the establishment of the historical feature database specifically includes:
[0077] On the device side, a unique device identifier is set for each medical purification device, and the end-side feature packet is bound with the device identifier and the corresponding collection timestamp to form a feature unit to be uploaded;
[0078] On the device side, the feature units to be uploaded are batch-collected and encapsulated according to a preset upload cycle to generate upload batches;
[0079] The uploaded batch is sent to the cloud platform via the end-to-cloud communication link and decapsulated. Integrity verification and timestamp consistency verification are performed on each feature unit to be uploaded to obtain valid feature units.
[0080] A corresponding historical feature database is established for each device identifier on the cloud platform. The establishment process involves the cloud creating an independent storage entry for each device as a historical feature database based on the device identifier, and writing the received valid feature units into the historical feature database of the corresponding device according to the timestamp.
[0081] In this embodiment, the output of the health index, failure risk index, and remaining life prediction results specifically includes:
[0082] The system extracts feature vector sequences arranged in ascending order of timestamps from a historical feature database to form training samples for the corresponding devices. These samples are then input into the improved TS2Vec model. The improved TS2Vec model includes a multi-scale time series construction and enhancement module, a time series feature encoding module, a hierarchical comparative representation learning module, and a maintenance target joint prediction output module. The multi-scale time series construction and enhancement module segments the feature vector sequences in the training samples according to preset multi-scale pruning rules to form a set of positive sample pairs. The time series feature encoding module performs time-step encoding and outputs latent representation subsequences. The hierarchical comparative representation learning module introduces a reversible normalized flow hard negative sample mining mechanism to obtain the device time series representation vector. The maintenance target joint prediction output module inputs the device time series representation vector into a multi-task prediction head and outputs health indicators, fault risk indicators, and remaining life prediction results.
[0083] In the multi-scale time series construction and enhancement module, the feature vector sequence in the training samples is segmented according to the preset multi-scale pruning rules to obtain a set of sub-sequences at different time scales. The first time series enhancement operator and the second time series enhancement operator are applied to each sub-sequence in the set of sub-sequences to generate two corresponding enhanced sub-sequences, forming a set of positive sample pair sub-sequences.
[0084] In the temporal feature encoding module, the two enhanced subsequences in the set of positive sample pair subsequences are encoded step by step over time, and the hidden representation subsequences corresponding one-to-one with the two enhanced subsequences are output.
[0085] In the hierarchical contrastive representation learning module, a reversible normalized flow hard negative sample mining mechanism is introduced to estimate the probability density of the latent representation sequence at the current scale. Based on the probability density estimate, a set of hard negative samples related to positive sample pairs is selected from the latent representation sequences of devices in the same batch. Hierarchical consistency alignment and time-step contrast calculation are performed on two latent representation sequences to obtain the contrast loss at the corresponding time scale. The contrast losses of all time scales and all time steps are accumulated to obtain the hierarchical contrast loss.
[0086] The temporal feature encoding module is self-supervised and optimized based on hierarchical contrastive loss, and the output device temporal representation vector corresponding to each training sample is generated.
[0087] In the joint prediction output module for maintenance targets, the equipment time-series representation vector is input into the multi-task prediction head. The multi-task prediction head includes a health prediction head, a fault risk prediction head, and a remaining lifetime prediction head. The health prediction head performs a first mapping on the equipment time-series representation vector to obtain a raw health prediction value, and then performs a normalized mapping on the raw health prediction value to obtain a health index. The fault risk prediction head performs a second mapping on the equipment time-series representation vector to obtain a raw fault risk prediction value, and then performs a risk probability mapping on the raw fault risk prediction value to obtain a fault risk index. The remaining lifetime prediction head performs a third mapping on the equipment time-series representation vector to obtain a raw remaining lifetime prediction value, and then performs a preset lifetime calibration mapping on the raw remaining lifetime prediction value to obtain a remaining lifetime prediction result.
[0088] In this embodiment, the generation of the maintenance strategy set specifically includes:
[0089] On the cloud platform, each device is assigned a preset health trigger threshold, risk trigger threshold, and lifespan trigger threshold. Maintenance trigger conditions are generated based on the health index, fault risk index, and remaining lifespan prediction result corresponding to the current time window. The maintenance trigger conditions refer to the health index being less than the health trigger threshold, the fault risk index being greater than the risk trigger threshold, or the remaining lifespan prediction result being less than the lifespan trigger threshold.
[0090] When the maintenance trigger condition is determined to be triggered, the maintenance timing parameters are obtained based on the timestamp corresponding to the current time window and the remaining lifetime prediction result. The maintenance timing parameters include the maintenance start time and the maintenance end time. The maintenance start time is the timestamp corresponding to the current time window, and the maintenance end time is the time when the timestamp corresponding to the current time window is superimposed with the estimated runnable time corresponding to the remaining lifetime prediction result.
[0091] When the maintenance trigger condition is determined to be triggered, maintenance content parameters are generated based on health indicators, fault risk indicators and remaining life prediction results. The maintenance content parameters include a set of filter maintenance items, a set of fan or pump maintenance items and a set of disinfection module maintenance items. When the maintenance trigger condition is not determined to be triggered, the maintenance content parameters are an empty set.
[0092] On the cloud platform, maintenance trigger conditions, maintenance timing parameters, and maintenance content parameters are combined according to device identifier and time window sequence number to form a set of maintenance strategies.
[0093] In this embodiment, the distribution of the edge inference parameters and maintenance strategy set specifically includes:
[0094] On the cloud platform, for each device identifier, the edge inference parameters of the improved TS2Vec model corresponding to the device identifier are retrieved, and the set of maintenance strategies corresponding to the device identifier is also retrieved.
[0095] On the cloud platform, the edge inference parameters and maintenance strategy set are encapsulated according to a preset data structure to generate a data packet that corresponds one-to-one with the device identifier, and a data packet generation timestamp is added to the data packet.
[0096] The data packets are sent to the device corresponding to the device identifier via the end-to-cloud communication link.
[0097] In this embodiment, obtaining the end-side health status and end-side risk assessment results specifically includes:
[0098] At the device end, perform end-side feature extraction on the real-time acquired end-side data to obtain the feature vector corresponding to the time window;
[0099] On the device side, the feature vector is input into the edge inference network corresponding to the improved TS2Vec model. The feature vector is then mapped to health and risk, and the original predicted values of edge health and edge fault risk are output.
[0100] On the device side, the original predicted value of the edge health is normalized to obtain the edge health value. The edge health value is then compared with the health trigger threshold issued by the cloud. On the device side, the original predicted value of the edge fault risk is compared with the risk trigger threshold issued by the cloud to obtain the edge risk judgment result. The comparison process is as follows: when the original predicted value of the edge fault risk is greater than the risk trigger threshold, the edge risk judgment result is in a triggered state; when the original predicted value of the edge fault risk is less than the risk trigger threshold, the edge risk judgment result is in a non-triggered state.
[0101] In this embodiment, the updating of the historical feature database specifically includes:
[0102] Based on the maintenance trigger conditions in the maintenance strategy set, trigger determination is made on the edge health and edge risk assessment results. When the determination is triggered, maintenance closed-loop action is executed; when the determination is not triggered, maintenance closed-loop action in the current time window is terminated.
[0103] When the state is determined to be triggered, the device generates an early warning information based on the maintenance timing parameters and the maintenance content parameters in the maintenance strategy set. The early warning information includes the device identifier, the triggered maintenance conditions, the terminal health corresponding to the current time window, the terminal risk judgment result corresponding to the current time window, and the corresponding maintenance content parameters.
[0104] The maintenance closed-loop actions in the triggered state are recorded to form a maintenance event log. The maintenance event log includes the device identifier, the timestamp of the trigger time, the terminal health status, the terminal risk assessment result, the maintenance triggering conditions, the maintenance timing parameters, the maintenance content parameters, and the corresponding warning information.
[0105] The device binds and encapsulates the maintenance event log with the edge feature packet corresponding to the trigger time, and sends it back to the cloud platform through the edge-cloud communication link to update the historical feature database.
[0106] Example 1:
[0107] This embodiment selects a newly built comprehensive building of a tertiary hospital as the application scenario. The building includes two Class 100 laminar flow operating rooms, a negative pressure isolation ward, an intensive care unit, and a supporting central purification room. The core medical purification equipment in the building includes: laminar flow air supply purification units, negative pressure ward dedicated purification units, and several terminal air disinfection and circulation purification devices. All of the above equipment operates continuously for a long time, in high-load mode during the day and in low-noise, low-power mode at night, with significant load fluctuations on weekends and holidays. In the past, the hospital's maintenance department mainly performed filter replacement, fan maintenance, and disinfection module calibration according to fixed cycles. In actual operation, there were many instances where "the purification effect had decreased but maintenance was not triggered until the differential pressure or particulate matter index exceeded the limit, at which point maintenance was passively shut down." Especially in the negative pressure isolation ward, a partial blockage of the filter caused unstable differential pressure control, resulting in a short-term deviation of the negative pressure from the target value, posing a safety hazard. It can be seen that the most prominent problem in this scenario is that the degradation of key components and the decline in purification performance are insidious and gradual. Traditional periodic maintenance and threshold alarms are unable to identify trends in time, leading to maintenance delays and sudden risks.
[0108] When deploying the method of this invention on the purification equipment in this building, the equipment first connects to the existing sensing system to collect multi-source sensor data, including the pressure difference before and after the filter element, air volume, fan speed, motor current, cavity pressure, inlet and outlet particle counts, temperature and humidity, and disinfection module operating power. The device's built-in edge computing unit performs time alignment, noise removal, and outlier filtering to ensure data stability and reliability. Subsequently, the equipment extracts time-domain and frequency-domain features from the pre-processed multi-source data and constructs derived indicators such as the filter element pressure difference change rate, fan load coefficient, and purification efficiency index to form feature vectors that characterize component degradation and purification attenuation. These feature vectors, along with timestamps, are then encapsulated into an end-side feature package. The end-side feature package is uploaded to the cloud platform via the hospital's intranet end-to-cloud communication link. The cloud platform establishes a historical feature database based on device identification and time sequence, continuously accumulating operational features across departments and operating conditions.
[0109] The cloud platform constructs training samples based on a historical feature library and inputs them into an improved TS2Vec model for self-supervised multi-scale comparative representation learning. In the multi-scale time series construction and enhancement module, the model forms sub-sequence views at different time granularities. In the time series feature encoding module, it obtains latent representation sequences. Furthermore, in the hierarchical comparative representation learning stage, a reversible normalized flow hard-to-bear sample mining mechanism is introduced, enabling the model to learn more robust health representations even under complex operating conditions such as "high load / low load / sudden pollution load changes." Finally, the joint prediction output module for maintenance targets outputs health indicators, failure risk indicators, and remaining life prediction results. Based on these three outputs, the cloud platform generates a set of maintenance strategies. The maintenance triggering conditions are directly given by the health, risk, and life thresholds. The maintenance timing parameters correspond to the operable window, and the maintenance content parameters are automatically mapped to the filter element maintenance item set, the fan or pump maintenance item set, and the disinfection module maintenance item set. The maintenance strategy set and the corresponding end-side inference parameters of the improved TS2Vec model are then sent to the equipment. During daily operation, the equipment uses the end-side inference parameters to perform online inference on real-time data, obtaining end-side health and risk assessment results. When the risk assessment result on the device side matches the maintenance trigger condition and the device is in a triggered state, the device generates an early warning message based on the maintenance timing parameters and maintenance content parameters in the maintenance strategy set. After outputting the early warning message, the device records this closed-loop action and generates a maintenance event log. The maintenance event log is bound and encapsulated with the device-side feature packet corresponding to the trigger time and then sent back to the cloud. The cloud writes this information back to the historical feature library and uses it to improve the TS2Vec model update, thus forming a continuously iterative end-cloud collaborative closed loop.
[0110] After a six-month trial run, the hospital maintenance department compared the six months before and after the implementation of this invention method and obtained the following key findings: First, the system can provide a risk trend warning before the filter resistance slowly increases but still does not exceed the traditional differential pressure threshold, with an average early warning time of 8 to 14 days. This allows maintenance personnel to schedule filter replacement and maintenance without affecting the continuous operation of surgical or isolation wards. Second, the combined degradation of a slight increase in fan current and a slight decrease in airflow would not trigger an alarm in the traditional single-threshold strategy. However, the combined output of the health index and fault risk index of this invention can identify fan degradation as medium to high risk in the early stages and indicate that the maintenance content parameters include a set of fan maintenance items, thereby avoiding purification failure caused by a sudden drop in fan efficiency. Third, because the improved TS2Vec model uses self-supervised learning, it can make full use of a large number of normal operation sequences. The lack of fault samples did not significantly affect the prediction stability, and the model maintained high consistency after migration between different departments and different operating modes. Finally, significant improvements were demonstrated in the availability of purification effect, the number of sudden shutdowns, and maintenance costs.
[0111] Table 1 Comparison of Predictive Maintenance Effects of Medical Purification Equipment
[0112] index Before launch (6 months) After launch (6 months) range of change Number of sudden equipment shutdowns (times) 9 2 ↓77.8% Clean / negative pressure exceeding limits (times) caused by purification performance degradation 6 1 ↓83.3% Average early warning time for filter cartridges (days) 0 (No advance warning) 11.2 +11.2 days Accuracy rate of fault risk warning (%) 58.4 86.7 +28.3 percentage points Average response time for maintenance work orders (hours) 17.5 6.3 ↓64.0% Percentage of filters replaced prematurely without prior replacement (%) 23.1 7.4 ↓68.0% Equipment purification effect meets the standard availability rate (%) 96.2 99.1 +2.9 percentage points Total maintenance cost (RMB 10,000 / 6 months) 41.6 32.9 ↓20.9%
[0113] As shown in Table 1, after implementing the method of this invention, the number of sudden shutdowns and clean / negative pressure over-limit events decreased from 9 to 2 within six months, and the number of over-limit events decreased from 6 to 1. This indicates that the combination of end-side online inference and maintenance triggering conditions enables early intervention against hidden degradation, significantly reducing the risk of "short-term performance collapse." The average early warning time reached 11.2 days, demonstrating the predictive ability of the improved TS2Vec model for degradation trends, enabling maintenance to shift from passive response to planned proactive maintenance, which is particularly suitable for scenarios such as operating rooms and negative pressure wards where arbitrary shutdowns are not allowed.
[0114] In terms of early warning accuracy and response time, the accuracy rate of fault risk early warning has improved to 86.7%, while the average response time for maintenance work orders has decreased from 17.5 hours to 6.3 hours. This reflects that the risk signals output by the cloud-based multi-scale representation learning are more stable and usable. Furthermore, the maintenance content and timing parameters enable maintenance personnel to quickly locate specific maintenance items such as filters, fans or pumps, and disinfection modules, reducing the time spent on repeated troubleshooting and waiting for confirmation. The improved early warning accuracy also means fewer false alarms, thus avoiding the interference of frequent and ineffective maintenance on clinical operations.
[0115] In terms of maintenance economy and purification availability, the proportion of unnecessary premature replacements decreased from 23.1% to 7.4%, and the total operation and maintenance cost decreased by approximately 20.9%. This indicates that the present invention, through joint decision-making based on remaining life prediction results and health indicators, significantly reduces the waste of "premature replacement according to fixed cycles." The purification effect compliance rate increased from 96.2% to 99.1%. Although the increase may seem small in percentage terms, in medical scenarios with high cleanliness levels and extremely high infection control requirements, any percentage point improvement means more robust environmental safety assurance and a lower risk of hospital-acquired infections.
[0116] In summary, this embodiment clearly demonstrates the implementation method and effect of the present invention in a real hospital medical purification equipment group: through the edge-cloud collaborative architecture and improved TS2Vec self-supervised temporal representation learning, the problems of existing technologies being insensitive to early degradation, maintenance lag, and insufficient edge-cloud prediction closed loop are solved, and quantifiable beneficial effects are achieved in terms of security, real-time performance, prediction accuracy, and operation and maintenance economy.
Claims
1. A predictive maintenance method for medical purification equipment based on edge-cloud collaboration, characterized in that, Includes the following steps: Collect and preprocess multi-source sensor data from medical purification equipment; Feature vectors are extracted from the preprocessed multi-source sensor data to generate edge-side feature packets; The edge-side feature packets are uploaded to the cloud platform through the edge-cloud communication link, and a historical feature database is established on the cloud platform according to device identification and time series. Training samples are constructed based on a historical feature database and input into the improved TS2Vec model to output health indicators, failure risk indicators and remaining life prediction results. A set of maintenance strategies is generated on the cloud platform based on health indicators, failure risk indicators, and remaining life prediction results. The improved TS2Vec model corresponding edge-side inference parameters and maintenance strategy set will be distributed to the device. Online reasoning is performed on the real-time collected edge data at the device end to obtain edge health and edge risk assessment results; Based on the edge health status, edge risk assessment results, and maintenance strategy set, a closed-loop maintenance action is executed, early warning information and maintenance event logs are output, and the maintenance event logs and corresponding edge feature packets are sent back to the cloud platform to update the historical feature database.
2. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The multi-source sensor data includes air volume data, pressure difference data before and after the filter element, particulate matter concentration, particle count data, temperature and humidity data, operating status data of key actuators, and electrical parameter data. The preprocessing includes noise reduction, outlier removal, and missing value completion.
3. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the end-side feature packet specifically includes: The preprocessed multi-source sensor data is divided into time window data segments according to a pre-set time window length and sliding step size. For each sensor variable sampling sequence within each time window data segment, calculate the time domain statistical characteristics, which include the mean, standard deviation, root mean square, skewness, and kurtosis of the sensor variable within the time window; For each sensor variable sampling sequence within each time window data segment, a discrete Fourier transform is performed to obtain the corresponding frequency domain amplitude sequence. Based on the frequency domain amplitude sequence, frequency domain features are calculated. The frequency domain features include frequency domain energy and spectral centroid. The frequency domain energy is the cumulative value of the squares of the amplitudes at each frequency point in the frequency domain amplitude sequence. The spectral centroid is the ratio of the cumulative value of the product of each frequency point frequency and its corresponding amplitude to the cumulative value of the amplitudes at each frequency point. Based on the sensor variables of key components degrading in the time window data segment, derived indices are constructed. The derived indices include the pressure difference across the filter element, the pressure difference change rate, the fan load factor, and the purification efficiency index. The pressure difference across the filter element is the difference between the average upstream pressure and the average downstream pressure of the filter element. The pressure difference change rate is the difference between the pressure difference across the filter element in the current time window and the pressure difference across the filter element in the previous time window, divided by the sliding step size. The fan load factor is the ratio of the average fan or pump current to the average fan speed. The purification efficiency index is a value minus the ratio of the average outlet particulate matter concentration or particle count to the average inlet particulate matter concentration or particle count. The time-domain statistical features, frequency-domain features, and derived indicators are concatenated to form a feature vector of key component degradation and purification performance attenuation. The feature vector is then encapsulated with the corresponding timestamp to generate an end-side feature package.
4. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The establishment of the historical feature database specifically includes: On the device side, a unique device identifier is set for each medical purification device, and the end-side feature packet is bound with the device identifier and the corresponding collection timestamp to form a feature unit to be uploaded; On the device side, the feature units to be uploaded are batch-collected and encapsulated according to a preset upload cycle to generate upload batches; The uploaded batch is sent to the cloud platform via the end-to-cloud communication link and decapsulated. Integrity verification and timestamp consistency verification are performed on each feature unit to be uploaded to obtain valid feature units. A corresponding historical feature database is established for each device identifier on the cloud platform. The establishment process involves the cloud creating an independent storage entry for each device as a historical feature database based on the device identifier, and writing the received valid feature units into the historical feature database of the corresponding device according to the timestamp.
5. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The outputs of the health index, failure risk index, and remaining life prediction results specifically include: The system extracts feature vector sequences arranged in ascending order of timestamps from a historical feature database to form training samples for the corresponding devices. These samples are then input into the improved TS2Vec model. The improved TS2Vec model includes a multi-scale time series construction and enhancement module, a time series feature encoding module, a hierarchical comparative representation learning module, and a maintenance target joint prediction output module. The multi-scale time series construction and enhancement module segments the feature vector sequences in the training samples according to preset multi-scale pruning rules to form a set of positive sample pairs. The time series feature encoding module performs time-step encoding and outputs latent representation subsequences. The hierarchical comparative representation learning module introduces a reversible normalized flow hard negative sample mining mechanism to obtain the device time series representation vector. The maintenance target joint prediction output module inputs the device time series representation vector into a multi-task prediction head and outputs health indicators, fault risk indicators, and remaining life prediction results. In the multi-scale time series construction and enhancement module, the feature vector sequence in the training samples is segmented according to the preset multi-scale pruning rules to obtain a set of sub-sequences at different time scales. The first time series enhancement operator and the second time series enhancement operator are applied to each sub-sequence in the set of sub-sequences to generate two corresponding enhanced sub-sequences, forming a set of positive sample pair sub-sequences. In the temporal feature encoding module, the two enhanced subsequences in the set of positive sample pair subsequences are encoded step by step over time, and the hidden representation subsequences corresponding one-to-one with the two enhanced subsequences are output. In the hierarchical contrastive representation learning module, a reversible normalized flow hard negative sample mining mechanism is introduced to estimate the probability density of the latent representation sequence at the current scale. Based on the probability density estimate, a set of hard negative samples related to positive sample pairs is selected from the latent representation sequences of devices in the same batch. Hierarchical consistency alignment and time-step contrast calculation are performed on two latent representation sequences to obtain the contrast loss at the corresponding time scale. The contrast losses of all time scales and all time steps are accumulated to obtain the hierarchical contrast loss. The temporal feature encoding module is self-supervised and optimized based on hierarchical contrastive loss, and the output device temporal representation vector corresponding to each training sample is generated. In the joint prediction output module for maintenance targets, the equipment time-series representation vector is input into the multi-task prediction head. The multi-task prediction head includes a health prediction head, a fault risk prediction head, and a remaining lifetime prediction head. The health prediction head performs a first mapping on the equipment time-series representation vector to obtain a raw health prediction value, and then performs a normalized mapping on the raw health prediction value to obtain a health index. The fault risk prediction head performs a second mapping on the equipment time-series representation vector to obtain a raw fault risk prediction value, and then performs a risk probability mapping on the raw fault risk prediction value to obtain a fault risk index. The remaining lifetime prediction head performs a third mapping on the equipment time-series representation vector to obtain a raw remaining lifetime prediction value, and then performs a preset lifetime calibration mapping on the raw remaining lifetime prediction value to obtain a remaining lifetime prediction result.
6. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the maintenance strategy set specifically includes: On the cloud platform, each device is assigned a preset health trigger threshold, risk trigger threshold, and lifespan trigger threshold. Maintenance trigger conditions are generated based on the health index, fault risk index, and remaining lifespan prediction result corresponding to the current time window. The maintenance trigger conditions refer to the health index being less than the health trigger threshold, the fault risk index being greater than the risk trigger threshold, or the remaining lifespan prediction result being less than the lifespan trigger threshold. When the maintenance trigger condition is determined to be triggered, the maintenance timing parameters are obtained based on the timestamp corresponding to the current time window and the remaining lifetime prediction result. The maintenance timing parameters include the maintenance start time and the maintenance end time. The maintenance start time is the timestamp corresponding to the current time window, and the maintenance end time is the time when the timestamp corresponding to the current time window is superimposed with the estimated runnable time corresponding to the remaining lifetime prediction result. When the maintenance trigger condition is determined to be triggered, maintenance content parameters are generated based on health indicators, fault risk indicators and remaining life prediction results. The maintenance content parameters include a set of filter maintenance items, a set of fan or pump maintenance items and a set of disinfection module maintenance items. When the maintenance trigger condition is not determined to be triggered, the maintenance content parameters are an empty set. On the cloud platform, maintenance trigger conditions, maintenance timing parameters, and maintenance content parameters are combined according to device identifier and time window sequence number to form a set of maintenance strategies.
7. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The distribution of the edge inference parameters and maintenance strategy set specifically includes: On the cloud platform, for each device identifier, the edge inference parameters of the improved TS2Vec model corresponding to the device identifier are retrieved, and the set of maintenance strategies corresponding to the device identifier is also retrieved. On the cloud platform, the edge inference parameters and maintenance strategy set are encapsulated according to a preset data structure to generate a data packet that corresponds one-to-one with the device identifier, and a data packet generation timestamp is added to the data packet. The data packets are sent to the device corresponding to the device identifier via the end-to-cloud communication link.
8. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The results of the end-side health assessment and end-side risk determination are obtained specifically in the following ways: At the device end, perform end-side feature extraction on the real-time acquired end-side data to obtain the feature vector corresponding to the time window; On the device side, the feature vector is input into the edge inference network corresponding to the improved TS2Vec model. The feature vector is then mapped to health and risk, and the original predicted values of edge health and edge fault risk are output. On the device side, the original predicted value of the edge health is normalized to obtain the edge health value. The edge health value is then compared with the health trigger threshold issued by the cloud. On the device side, the original predicted value of the edge fault risk is compared with the risk trigger threshold issued by the cloud to obtain the edge risk judgment result. The comparison process is as follows: when the original predicted value of the edge fault risk is greater than the risk trigger threshold, the edge risk judgment result is in a triggered state; when the original predicted value of the edge fault risk is less than the risk trigger threshold, the edge risk judgment result is in a non-triggered state.
9. The predictive maintenance method for medical purification equipment based on edge-cloud collaboration according to claim 1, characterized in that, The updating of the historical feature database specifically includes: Based on the maintenance trigger conditions in the maintenance strategy set, trigger determination is made on the edge health and edge risk assessment results. When the determination is triggered, maintenance closed-loop action is executed; when the determination is not triggered, maintenance closed-loop action in the current time window is terminated. When the state is determined to be triggered, the device generates an early warning information based on the maintenance timing parameters and the maintenance content parameters in the maintenance strategy set. The early warning information includes the device identifier, the triggered maintenance conditions, the terminal health corresponding to the current time window, the terminal risk judgment result corresponding to the current time window, and the corresponding maintenance content parameters. The maintenance closed-loop actions in the triggered state are recorded to form a maintenance event log. The maintenance event log includes the device identifier, the timestamp of the trigger time, the terminal health status, the terminal risk assessment result, the maintenance triggering conditions, the maintenance timing parameters, the maintenance content parameters, and the corresponding warning information. The device binds and encapsulates the maintenance event log with the edge feature packet corresponding to the trigger time, and sends it back to the cloud platform through the edge-cloud communication link to update the historical feature database.