A Real-Time Diagnostic Method and System for Flexible Stent Acceleration Data Based on Edge Computing

Abstract

This invention belongs to the field of engineering structure health monitoring technology, and discloses a real-time diagnostic method and system for acceleration data of flexible supports based on edge computing. Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support. Combined with an event activation mechanism, the system adapts to the dynamic characteristics of acceleration data and avoids invalid transmission, achieving localized and accurate data acquisition. The edge computing nodes process the raw data locally and extract feature vectors, reducing transmission latency. A built-in lightweight diagnostic model performs real-time inference on the feature vectors, completing anomaly detection and fault classification without relying on cloud computing power. The system integrates model results, historical data, and a local rule base to form diagnostic decisions, constructing a localized closed loop of acquisition, processing, inference, and decision-making. This effectively solves the problems of slow diagnostic response and poor adaptability in existing technologies, achieving real-time, accurate, and localized diagnosis of acceleration data for flexible supports.

Description

Technical Field

[0001] This invention belongs to the field of engineering structure health monitoring technology, specifically relating to a real-time diagnostic method and system for acceleration data of flexible supports based on edge computing. Background Technology

[0002] Flexible supports are widely used in photovoltaic power plants, communication base stations, and agricultural facilities due to their advantages such as light weight, large span, and good economic efficiency. However, these structures are prone to significant vibration under dynamic environmental loads such as wind and snow loads. Long-term excessive or abnormal vibration modes can accelerate fatigue damage, leading to loosening of connecting components, degradation of material properties, and even overall instability or collapse. Therefore, real-time and accurate health monitoring and diagnosis of the flexible supports are crucial for ensuring their safe operation, preventing major accidents, and enabling predictive maintenance.

[0003] Current common monitoring methods mainly rely on deploying accelerometer sensor nodes at key parts of the support structure. The collected high-frequency raw acceleration data is continuously transmitted via wireless network to a remote cloud server or data center for centralized storage, processing, and analysis. While this cloud-centric centralized processing model is powerful, it has revealed the following shortcomings in practical engineering applications: First, the data transmission volume and cost are high. Acceleration signals typically require high sampling frequencies to ensure information integrity, leading to the continuous generation of massive amounts of raw data. Real-time, uninterrupted remote transmission of this data generates huge wireless communication bandwidth demands, resulting in high traffic costs and power consumption at terminal nodes, limiting the large-scale, long-term deployment of monitoring systems. Second, the diagnostic real-time performance is poor. Data must go through a lengthy link: terminal acquisition → network upload → cloud processing → result distribution, with an overall latency that can reach several seconds or even minutes. This makes it difficult for the system to achieve millisecond or second-level real-time anomaly detection and early warning, failing to meet the immediate response requirements for sudden, severe vibrations or rapid structural deterioration. Finally, the cloud computing and storage pressure is enormous. When faced with tens of thousands of sensor nodes simultaneously uploading data, the aggregation of massive amounts of data poses a severe challenge to the computing resources, storage resources, and processing capabilities of the cloud platform, limiting the system's scalability and continuously increasing operation and maintenance costs.

[0004] Edge computing, as an emerging computing paradigm, offers a new technical approach to solving the aforementioned problems by shifting data processing and analysis tasks from the network core to the network edge, closer to the data source. It can effectively reduce data transmission, lower latency, and enhance the system's autonomy in network outage conditions. However, when deploying diagnostic functions on resource-constrained edge devices, designing a real-time processing workflow that integrates efficient data preprocessing, lightweight feature extraction, and low-complexity intelligent diagnostic algorithms to achieve localized, rapid anomaly identification and fault diagnosis, given the rich frequency domain characteristics, high noise interference, and variable operating conditions of flexible support acceleration data, remains a key technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of how to achieve localized and rapid anomaly identification and fault diagnosis based on the characteristics of acceleration data of flexible stents in the prior art, and to provide a real-time diagnostic method and system for acceleration data of flexible stents based on edge computing.

[0006] To achieve the above objectives, the present invention employs the following technical solution: The present invention proposes a real-time diagnostic method for acceleration data of flexible stents based on edge computing, comprising the following steps: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, the system enters an event activation state. Edge computing nodes perform local processing and feature extraction on raw acceleration data within a fixed time window to obtain feature vectors. A lightweight diagnostic model is built into the edge computing node to perform real-time reasoning on the generated feature vectors, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. Edge computing nodes integrate the inference results of lightweight diagnostic models, historical data, and local rule bases to form diagnostic decisions, enabling real-time diagnosis of flexible stent acceleration data.

[0007] Preferably, the deployment of accelerometers and edge computing nodes at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration, and the entry into an event activation state when the characteristic value exceeds a preset quiet state threshold, specifically: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support. In each computing cycle, the nodes are configured according to a window... Calculate the root mean square value of the triaxial resultant acceleration. ;in, N It is a window Number of sampling points within, It is the first The composite acceleration value of each sampling point; When the feature value is greater than the quiet state threshold The duration exceeds the activation duration When a significant vibration event occurs, the system determines that the event has occurred and switches to the event activation state. When the eigenvalue < The duration exceeds the hibernation delay time When the event is considered over, the system returns to a quiet standby state, uploading only a status summary and heartbeat packet.

[0008] Preferably, the local processing and feature extraction to obtain the feature vector specifically involves: Local processing involves noise reduction and filtering, using moving averages or digital bandpass filters to remove high-frequency noise and low-frequency drift. Feature extraction is a lightweight extraction of time-domain and frequency-domain feature vectors, including the maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis, and dominant frequency components and frequency band energy proportions extracted through fast Fourier transform.

[0009] Preferably, the lightweight time-domain and frequency-domain feature vector extraction specifically includes: Calculate the time-domain features (maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis) and frequency-domain features (dominant frequency, band energy ratio) for the filtered triaxial data; finally, generate the feature vector for the analysis window. F ;

[0010] in, for x The root mean square value of the axis represents x The effective energy or average intensity of vibration in the axial direction; for x The standard deviation of the axis represents x The degree of dispersion of shaft vibration data around the mean; for x The skewness of the axis indicates x Asymmetry in the probability distribution of shaft vibration signals; for x The kurtosis of the axis indicates the sharpness of the vibration waveform; express x The first dominant frequency of axis data; Indicates that flexible stents are in x The proportion of vibration energy falling within frequency band 1 in the axial direction; For have y The root mean square value of the axis represents y The effective energy or average intensity of vibration in the axial direction; This represents the total intensity of triaxial vibration; Indicates the first dominant frequency of the three-axis data; This indicates the proportion of the vibration energy of the flexible support in the three axial directions that falls within frequency band 1.

[0011] Preferably, the step of embedding a lightweight diagnostic model within the edge computing node to perform real-time inference on the generated feature vectors, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels, specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

[0012] Preferably, the diagnostic decision specifically includes: If the diagnosis is a minor abnormality or a known pattern, a structured diagnostic report is generated and uploaded to the cloud platform or regional gateway in a low-data-volume format. If a serious abnormality or unknown danger mode is diagnosed, a local audible and visual alarm will be triggered directly within milliseconds, and the highest priority alarm information will be sent to the superior node and cloud platform simultaneously. Edge collaboration involves multiple edge nodes within a region communicating and comparing diagnostic results from adjacent measurement points to rule out isolated sensor failures or locate the approximate area of ​​damage.

[0013] Preferably, a cloud platform is used to aggregate reports and alarm information uploaded by each edge computing node for macro-analysis and visualization; the lightweight diagnostic model is iteratively optimized based on the aggregated multi-point data, and the updated model parameters are incrementally pushed to the edge computing nodes to achieve continuous evolution of diagnostic capabilities.

[0014] Preferably, the local rule base specifically comprises: Rule 1: If the anomaly detection model outputs a normal result, it is directly judged as normal and a short log is generated. Rule 2: If the anomaly detection model outputs an anomaly and the fault classification model outputs vortex-induced vibration with a confidence level > 80%, then it is determined to be a known minor anomaly - vortex-induced vibration. Rule 3: If the fault classification model outputs "loose connection" or "structural resonance", it is determined to be a known serious anomaly. Rule 4: If the anomaly detection model outputs an anomaly, but the confidence level of the fault classification model is less than 60%, it is determined to be an unknown danger mode.

[0015] This invention proposes a real-time diagnostic system for acceleration data of flexible stents based on edge computing, comprising: The event triggering module is used to deploy acceleration sensors and edge computing nodes at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, it enters the event activation state. The feature vector acquisition module is used by the edge computing node to perform local processing and feature extraction on the raw acceleration data within a fixed time window to obtain a feature vector. The lightweight diagnostic reasoning module is used to embed a lightweight diagnostic model in the edge computing node, perform real-time reasoning on the generated feature vector, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. The diagnostic decision module is used by the edge computing node to integrate the reasoning results of the lightweight diagnostic model, historical data and local rule base to form a diagnostic decision, thereby realizing real-time diagnosis of the acceleration data of the flexible stent.

[0016] Preferably, the step of embedding a lightweight diagnostic model within the edge computing node to perform real-time inference on the generated feature vectors, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels, specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a real-time diagnostic method for flexible stent acceleration data based on edge computing. First, by deploying acceleration sensors and edge computing nodes at key monitoring points of the flexible stent, real-time statistical characteristics of acceleration are continuously monitored, adapting to the dynamic changes in flexible stent acceleration data and avoiding invalid data transmission. Second, the edge computing nodes perform local processing and feature extraction on the raw acceleration data within a fixed time window, transforming the high-dimensional raw data into feature vectors. This adapts to the high-dimensional and real-time characteristics of acceleration data while reducing data transmission latency through local processing, ensuring rapid diagnosis. Third, the edge computing nodes incorporate a lightweight diagnostic model that performs real-time inference on the feature vectors, simultaneously completing anomaly identification and fault classification and outputting results. This eliminates reliance on cloud computing power, enabling rapid localized diagnosis that meets the actual needs of on-site diagnosis of flexible stents. Finally, the edge computing nodes combine the model inference results, historical data, and local rule base to form a diagnostic decision, further improving diagnostic accuracy. This addresses the problems of existing technologies, such as reliance on the cloud, slow response, and inability to adapt to the characteristics of flexible stent acceleration data, achieving real-time, accurate, and localized diagnosis of flexible stent acceleration data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the real-time diagnostic method for acceleration data of flexible stents based on edge computing according to the present invention.

[0020] Figure 2 This is a detailed flowchart of the real-time diagnostic method for acceleration data of flexible stents based on edge computing according to the present invention.

[0021] Figure 3 This is a diagram of the real-time diagnostic system for acceleration data of a flexible stent based on edge computing according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 The purpose of this invention is to address the problem of how to achieve rapid localized anomaly identification and fault diagnosis based on the characteristics of acceleration data from flexible stents in existing technologies. This invention provides a real-time diagnostic method for acceleration data of flexible stents based on edge computing. This method achieves localized intelligent monitoring through edge computing nodes deployed at key points of the stent. First, the system operates an adaptive event-triggered mechanism, continuously monitoring the root mean square value of the acceleration sliding window. The full processing flow is activated only after the value exceeds a preset quiet state threshold and remains so for a specified period. Normally, only heartbeat packets are sent to save energy and bandwidth. After event activation, the edge nodes perform noise reduction filtering on the raw data within a fixed time window and extract lightweight feature vectors, including time-domain and frequency-domain features, compressing high-dimensional data into a low-dimensional representation. Subsequently, the feature vectors are input into a built-in lightweight diagnostic model for inference. The anomaly detection model is used to determine whether the vibration deviates from the normal range, and the fault classification... The model accurately classifies the fault type and outputs confidence scores after an anomaly is triggered. Based on the inference results and the local rule base, edge nodes form diagnostic decisions, uploading structured reports for minor anomalies and triggering local audible and visual alarms and simultaneously uploading the highest priority alarm for severe or unknown dangerous modes within milliseconds. It also supports collaborative judgment between adjacent nodes through simple communication to eliminate false alarms. Finally, the cloud platform aggregates data from all nodes, performs visualization analysis and model retraining, and incrementally pushes the optimized, lightweight model to edge nodes using pruning, quantization, and other techniques, achieving continuous evolution of diagnostic capabilities. This method, through a closed-loop process of "edge-side triggering - local processing - intelligent inference - collaborative decision-making - cloud-edge optimization," achieves ultra-low latency response to flexible support vibration, massive data compression, availability during network outages, and adaptive, accurate diagnosis.

[0027] To achieve the above objectives, the present invention employs the following technical solution: A real-time diagnostic method for acceleration data of flexible stents based on edge computing, such as Figure 1 As shown, it includes the following steps: Step 1: Deploy accelerometers and edge computing nodes at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, it will enter the event activation state. The deployment of accelerometers and edge computing nodes at key monitoring points of the flexible support continuously monitors the real-time statistical characteristics of acceleration. If the characteristic value exceeds a preset quiet state threshold, an event activation state is entered. Specifically: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support. In each computing cycle, the nodes are configured according to a window... Calculate the root mean square value of the triaxial resultant acceleration. ;in, N It is a window Number of sampling points within, It is the first The composite acceleration value of each sampling point; When the feature value is greater than the quiet state threshold The duration exceeds the activation duration When a significant vibration event occurs, the system determines that the event has occurred and switches to the event activation state. When the eigenvalue < The duration exceeds the hibernation delay time When the event is considered over, the system returns to a quiet standby state, uploading only a status summary and heartbeat packet.

[0028] Step 2: The edge computing nodes perform local processing and feature extraction on the raw acceleration data within a fixed time window to obtain feature vectors; The local processing and feature extraction to obtain the feature vector are as follows: Local processing involves noise reduction and filtering, using moving averages or digital bandpass filters to remove high-frequency noise and extremely low-frequency drift. Feature extraction is a lightweight extraction of time-domain and frequency-domain feature vectors, including the maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis, and dominant frequency components and frequency band energy proportions extracted through fast Fourier transform.

[0029] The lightweight time-domain and frequency-domain feature vector extraction is specifically as follows: Calculate the time-domain features (maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis) and frequency-domain features (dominant frequency, band energy ratio) for the filtered triaxial data; finally, generate the feature vector for the analysis window. F ;

[0030] in, for x The root mean square value of the axis represents x The effective energy or average intensity of vibration in the axial direction; for x The standard deviation of the axis represents x The degree of dispersion of shaft vibration data around the mean; for x The skewness of the axis indicates x Asymmetry in the probability distribution of shaft vibration signals; for x The kurtosis of the axis indicates the sharpness of the vibration waveform; expressx The first dominant frequency of axis data; Indicates that flexible stents are in x The proportion of vibration energy falling within frequency band 1 in the axial direction; For have y The root mean square value of the axis represents y The effective energy or average intensity of vibration in the axial direction; This represents the total intensity of triaxial vibration; Indicates the first dominant frequency of the three-axis data; This indicates the proportion of the vibration energy of the flexible support in the three axial directions that falls within frequency band 1.

[0031] Step 3: Integrate a lightweight diagnostic model into the edge computing node, perform real-time inference on the generated feature vectors, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. The lightweight diagnostic model built into the edge computing node performs real-time inference on the generated feature vectors, completes anomaly detection and fault classification, and outputs preliminary diagnostic results and confidence levels. Specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

[0032] Step 4: The edge computing node integrates the reasoning results of the lightweight diagnostic model, historical data, and local rule base to form a diagnostic decision, thereby realizing real-time diagnosis of the acceleration data of the flexible stent.

[0033] The diagnostic decision is specifically as follows: If the diagnosis is a minor abnormality or a known pattern, a structured diagnostic report is generated and uploaded to the cloud platform or regional gateway in a low-data-volume format. If a serious abnormality or unknown danger mode is diagnosed, a local audible and visual alarm will be triggered directly within milliseconds, and the highest priority alarm information will be sent to the superior node and cloud platform simultaneously. Edge collaboration involves multiple edge nodes within a region communicating and comparing diagnostic results from adjacent measurement points to rule out isolated sensor failures or locate the approximate area of ​​damage.

[0034] The cloud platform aggregates reports and alarm information uploaded by various edge computing nodes for macro-analysis and visualization. Based on the aggregated multi-point data, the lightweight diagnostic model is iteratively optimized, and the updated model parameters are incrementally pushed to the edge computing nodes to achieve continuous evolution of diagnostic capabilities.

[0035] The local rule base specifically refers to: Rule 1: If the anomaly detection model outputs a normal result, it is directly judged as normal and a short log is generated. Rule 2: If the anomaly detection model outputs an anomaly and the fault classification model outputs vortex-induced vibration with a confidence level > 80%, then it is determined to be a known minor anomaly - vortex-induced vibration. Rule 3: If the fault classification model outputs "loose connection" or "structural resonance", it is determined to be a known serious anomaly. Rule 4: If the anomaly detection model outputs an anomaly, but the confidence level of the fault classification model is less than 60%, it is determined to be an unknown danger mode.

[0036] The following is combined Figure 2 A detailed description is provided, including the following steps: S1, edge-side adaptive data acquisition and triggering, involves deploying triaxial accelerometers and edge computing nodes at key monitoring points of the flexible support. The edge computing nodes do not continuously upload raw data; instead, they run an adaptive event-triggered mechanism to continuously monitor the real-time statistical characteristics of acceleration.

[0037] When the feature value exceeds the preset quiet state threshold, the event is identified as active, and the complete processing flow of subsequent steps is initiated. When the feature value falls below the threshold and remains below it for a period of time, the system returns to a quiet standby state, only uploading heartbeat packets and status summaries.

[0038] S1.1 Initial State and Parameter Configuration: After the system is powered on, the edge nodes enter a quiet standby state. The nodes sample the triaxial accelerometer at a low fundamental frequency for status monitoring.

[0039] Pre-configure key parameters: monitoring features, selecting the root mean square value of the sliding window of the acceleration signal as the core feature for triggering judgment, and the window length. Set to 0.5~2s. Quiet state threshold. The activation duration was obtained through statistical analysis of long-term monitoring data of the support under static conditions such as no wind or light wind, using the 95th percentile of the historical root mean square acceleration values. To prevent accidental triggering due to momentary interference, the event must be considered activated only after the characteristic value exceeds the threshold for a specified duration. (Sleep hysteresis time) After the event ends, the feature value must remain below the threshold for that period of time before the system returns to standby mode, in order to avoid frequent switching around the threshold.

[0040] S1.2, Real-time monitoring and status assessment In each computation cycle, nodes are arranged by window. Calculate the root mean square value of the triaxial resultant acceleration. .

[0041] For triaxial data , , Composite acceleration

[0042] Root mean square value within the sliding window:

[0043] in, N It is a window T w Number of sampling points within, It is the first i The composite acceleration value of each sampling point.

[0044] State machine logic: Standby → Activation: When continuously calculated... a rms > The duration exceeded When a significant vibration event occurs, the system determines that the event has occurred and switches to the event activation state.

[0045] Activation → Standby: In the active state, when continuously calculated... a rms < The duration exceeded When the event is considered over, the system returns to a quiet standby state.

[0046] S1.3, State Behavior: In the quiet standby state, heartbeat packets are sent to the cloud platform at an extremely low frequency, containing the device ID, battery voltage, signal strength, and status identifier (normal standby), with almost no data traffic. The event activation state initiates the complete data processing and diagnostic pipeline of subsequent S2-S4, and sends a status summary to the cloud at a higher frequency, notifying the cloud that this node has entered a working state.

[0047] S2, Edge-side Data Preprocessing and Feature Engineering: In event-activated states, edge computing nodes locally process the raw acceleration data within a fixed time window: Noise reduction and filtering using moving averages or digital bandpass filters to remove high-frequency noise and extremely low-frequency drift. Feature extraction: Calculating a set of lightweight time-domain and frequency-domain feature vectors, including the maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis, and dominant frequency components and frequency band energy proportions extracted through Fast Fourier Transform. The original high-dimensional data stream is compressed into low-dimensional feature vectors.

[0048] When the event is active, the node performs local processing on a high-quality data window to extract lightweight features that characterize the nature of the vibration.

[0049] S2.1, Data Caching and Window Partitioning. Each node allocates a circular buffer to continuously cache raw acceleration data for all three axes. When an event is activated, a fixed-length analysis time window is taken, starting from the trigger time. T ana The data is used as the object of this processing.

[0050] S2.2, Noise Reduction and Filtering. Detrending term removal: This removes linear or slowly varying baseline drift from the signal, typically achieved by subtracting the window mean or using high-pass filtering. Bandpass filtering: This applies a zero-phase digital filter to preserve the frequency bands relevant to the vibration of the flexible support structure.

[0051] S2.3, Lightweight Feature Vector Extraction. For the filtered three-axis data, calculate the time-domain features: maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, and kurtosis; and the frequency-domain features: dominant frequency and band energy proportion. Finally, generate the feature vector for the analysis window.

[0052] S3, Lightweight Intelligent Diagnostic Model Inference on the Edge Side: One or more lightweight diagnostic models are built into the edge computing nodes to perform real-time inference on the generated feature vectors. Model 1: Anomaly Detection Model. This model is trained based on historical data of the stent's health status and is used to determine whether the current vibration pattern falls within the normal range. If an anomaly is detected, a local alarm is immediately triggered.

[0053] eigenvectors F Input an Isolation Forest model, and the model outputs anomaly scores. s Alternatively, in decision labeling, the Isolation Forest model calculates the path length required for a sample to be isolated; the shorter the path, the higher the probability of an anomaly.

[0054] Set an abnormal threshold ,if If the current state is determined to be abnormal, an alarm flag is triggered locally on the edge node, and the subsequent fault classification process is activated.

[0055] Model 2: Fault Classification Model. When the anomaly detection model is triggered, or periodically started, the model performs a preliminary classification of the anomaly type, such as "vortex-induced vibration," "buffeting," "loose connection," "structural resonance," and "external impact."

[0056] A lightweight gradient boosting tree is used, with the model input being the feature vector from step 2. F This is combined with short time series segments. The model output is a probability vector. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class. The classification decision takes the class with the highest probability as the preliminary diagnosis result. Simultaneously record the confidence level. .

[0057] S4: Edge collaboration and diagnostic decision-making. Edge computing nodes combine model reasoning results, historical data, and local rule bases to form preliminary diagnostic decisions.

[0058] If the diagnosis is a minor anomaly or a known pattern, a structured diagnostic report (including timestamp, location, feature value, anomaly type, and confidence level) is generated and sent to the cloud platform or regional gateway in a low-data-volume format.

[0059] If a serious abnormality or unknown danger mode is diagnosed, a local audible and visual alarm will be triggered directly within milliseconds, and the highest priority alarm information will be sent to the superior node and cloud platform simultaneously.

[0060] Edge collaboration allows multiple edge nodes within a region to communicate simply and compare diagnostic results from adjacent measurement points, which can be used to rule out isolated sensor failures or locate the approximate area of ​​damage.

[0061] S5: Cloud-edge collaboration and model optimization. The cloud platform is responsible for macro-management and model iteration: receiving and aggregating summary reports and alarm information uploaded by each edge node, and performing visualization and macro-analysis. Incremental model updates: the cloud platform uses the aggregated multi-point data to retrain and optimize the diagnostic model, and pushes the updated lightweight model parameters incrementally to the edge nodes, enabling continuous evolution of diagnostic capabilities.

[0062] S5.1, Data Aggregation and Model Retraining: The cloud platform receives structured reports and raw data from edge nodes across the entire network; it uses the aggregated massive amounts of multi-condition data to periodically initiate model retraining.

[0063] Training process: For the fault classification model, use newly collected labeled data (manually labeled from historical alarms or automatically clustered) for full or incremental training to optimize model parameters. For the anomaly detection model, use newly collected normal data to refit the boundaries of normal states.

[0064] S5.2, Model Lightweighting and Push, transforms high-performance models trained in the cloud into lightweight versions suitable for edge devices through model pruning, quantization, and knowledge distillation. The new model parameters are packaged into incremental update packages and distributed to relevant edge nodes via a secure channel.

[0065] S5.3 enables hot updates of edge node models. Edge nodes receive and verify update packages when idle. A dual-backup strategy is employed to download the new model to a backup storage area. After verification, the system switches to the new model when there are no critical tasks, achieving seamless hot updates and continuously improving the diagnostic capabilities of edge nodes across the entire network.

[0066] Example 2 A system for implementing the above method includes: The sensing layer consists of several triaxial accelerometers deployed at key monitoring points of the flexible support, used to collect raw vibration acceleration data of the flexible support in real time. The edge computing layer includes multiple edge computing nodes that are connected in a one-to-one correspondence or in close proximity to the triaxial accelerometer. Each edge computing node includes a processor, a memory, and a communication module. When the processor executes a computer program, it implements the following modules: The adaptive acquisition triggering module is used to run the adaptive event triggering mechanism: continuously monitor the real-time statistical characteristics of acceleration, and when the characteristic value exceeds the preset quiet state threshold, it is determined to be an event activation state and the complete processing flow of the subsequent modules is started; when the characteristic value is lower than the threshold and continues for a preset time, it returns to the quiet standby state and only uploads heartbeat packets and status summaries through the communication module. The data preprocessing and feature engineering module is used to perform local processing on the raw acceleration data within a fixed time window when the event is active. This includes: noise reduction and filtering using a moving average or digital bandpass filter, and calculating lightweight time-domain and frequency-domain feature vectors. The feature vectors include the maximum value, minimum value, mean, root mean square value, standard deviation, amplitude, skewness, kurtosis, and the dominant frequency components and frequency band energy ratios extracted by fast Fourier transform within the window. A lightweight intelligent diagnostic inference module, which has at least one built-in lightweight diagnostic model, is used for real-time inference of the feature vector, including: The anomaly detection model, trained based on historical data of the stent's health status, is used to determine whether the current vibration mode is within the normal range. If it is determined to be abnormal, a local alarm is immediately triggered. The fault classification model is used to perform preliminary classification of anomaly types after the anomaly detection model is triggered or periodically started, and outputs fault categories and confidence levels including vortex-induced vibration, chattering, loose connection, structural resonance, and external impact. The edge collaboration and diagnostic decision module integrates model inference results, historical data, and the local rule base to form a preliminary diagnostic decision and execute corresponding actions based on the diagnostic results: if the diagnosis is a minor anomaly or a known pattern, a structured diagnostic report is generated and sent to the communication module in a low-data-volume format; if the diagnosis is a severe anomaly or an unknown dangerous pattern, a local audible and visual alarm is triggered directly within milliseconds, and the highest priority alarm information is sent to the superior node simultaneously; at the same time, this module supports simple communication between multiple edge nodes in the area to compare the diagnostic results of adjacent measuring points to rule out isolated sensor failures or to initially locate damaged areas; The network transport layer is used to enable data communication between edge computing nodes and between edge computing nodes and the cloud service platform. The cloud service platform, including cloud servers and databases, is used to receive and aggregate summary reports and alarm information uploaded by each edge node, perform visualization and macro-analysis, and use the aggregated multi-point data to retrain and optimize the diagnostic model. The updated lightweight model parameters are incrementally pushed to the edge nodes through the network transport layer to achieve continuous evolution of diagnostic capabilities.

[0067] Example 3 This invention proposes a real-time diagnostic system for acceleration data of flexible stents based on edge computing, such as... Figure 3 As shown, it includes: The event triggering module is used to deploy acceleration sensors and edge computing nodes at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, it enters the event activation state. The deployment of accelerometers and edge computing nodes at key monitoring points of the flexible support continuously monitors the real-time statistical characteristics of acceleration. If the characteristic value exceeds a preset quiet state threshold, an event activation state is entered. Specifically: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support. In each computing cycle, the nodes are configured according to a window... Calculate the root mean square value of the triaxial resultant acceleration. ;in, N It is a window Number of sampling points within, It is the first The composite acceleration value of each sampling point; When the feature value is greater than the quiet state threshold The duration exceeds the activation duration When a significant vibration event occurs, the system determines that the event has occurred and switches to the event activation state. When the eigenvalue < The duration exceeds the hibernation delay time When the event is considered over, the system returns to a quiet standby state, uploading only a status summary and heartbeat packet.

[0068] The feature vector acquisition module is used by the edge computing node to perform local processing and feature extraction on the raw acceleration data within a fixed time window to obtain a feature vector. The local processing and feature extraction to obtain the feature vector are as follows: Local processing involves noise reduction and filtering, using moving averages or digital bandpass filters to remove high-frequency noise and extremely low-frequency drift. Feature extraction is a lightweight extraction of time-domain and frequency-domain feature vectors, including the maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis, and dominant frequency components and frequency band energy proportions extracted through fast Fourier transform.

[0069] The lightweight time-domain and frequency-domain feature vector extraction is specifically as follows: Calculate the time-domain features (maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis) and frequency-domain features (dominant frequency, band energy ratio) for the filtered triaxial data; finally, generate the feature vector for the analysis window. F ;

[0070] in, for x The root mean square value of the axis represents x The effective energy or average intensity of vibration in the axial direction; for x The standard deviation of the axis represents x The degree of dispersion of shaft vibration data around the mean; for x The skewness of the axis indicates x Asymmetry in the probability distribution of shaft vibration signals; for x The kurtosis of the axis indicates the sharpness of the vibration waveform; express x The first dominant frequency of axis data; Indicates that flexible stents are in x The proportion of vibration energy falling within frequency band 1 in the axial direction; For have y The root mean square value of the axis represents y The effective energy or average intensity of vibration in the axial direction; This represents the total intensity of triaxial vibration; Indicates the first dominant frequency of the three-axis data; This indicates the proportion of the vibration energy of the flexible support in the three axial directions that falls within frequency band 1.

[0071] The lightweight diagnostic reasoning module is used to embed a lightweight diagnostic model in the edge computing node, perform real-time reasoning on the generated feature vector, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. The lightweight diagnostic model built into the edge computing node performs real-time inference on the generated feature vectors, completes anomaly detection and fault classification, and outputs preliminary diagnostic results and confidence levels. Specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

[0072] The diagnostic decision module is used by the edge computing node to integrate the reasoning results of the lightweight diagnostic model, historical data and local rule base to form a diagnostic decision, thereby realizing real-time diagnosis of the acceleration data of the flexible stent.

[0073] The diagnostic decision is specifically as follows: If the diagnosis is a minor abnormality or a known pattern, a structured diagnostic report is generated and uploaded to the cloud platform or regional gateway in a low-data-volume format. If a serious abnormality or unknown danger mode is diagnosed, a local audible and visual alarm will be triggered directly within milliseconds, and the highest priority alarm information will be sent to the superior node and cloud platform simultaneously. Edge collaboration involves multiple edge nodes within a region communicating and comparing diagnostic results from adjacent measurement points to rule out isolated sensor failures or locate the approximate area of ​​damage.

[0074] The cloud platform aggregates reports and alarm information uploaded by various edge computing nodes for macro-analysis and visualization. Based on the aggregated multi-point data, the lightweight diagnostic model is iteratively optimized, and the updated model parameters are incrementally pushed to the edge computing nodes to achieve continuous evolution of diagnostic capabilities.

[0075] The local rule base specifically refers to: Rule 1: If the anomaly detection model outputs a normal result, it is directly judged as normal and a short log is generated. Rule 2: If the anomaly detection model outputs an anomaly and the fault classification model outputs vortex-induced vibration with a confidence level > 80%, then it is determined to be a known minor anomaly - vortex-induced vibration. Rule 3: If the fault classification model outputs "loose connection" or "structural resonance", it is determined to be a known serious anomaly. Rule 4: If the anomaly detection model outputs an anomaly, but the confidence level of the fault classification model is less than 60%, it is determined to be an unknown danger mode.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A real-time diagnostic method for acceleration data of a flexible stent based on edge computing, characterized in that, Includes the following steps: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, the system enters an event activation state. Edge computing nodes perform local processing and feature extraction on raw acceleration data within a fixed time window to obtain feature vectors. A lightweight diagnostic model is built into the edge computing node to perform real-time reasoning on the generated feature vectors, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. Edge computing nodes integrate the inference results of lightweight diagnostic models, historical data, and local rule bases to form diagnostic decisions, enabling real-time diagnosis of flexible stent acceleration data.

2. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 1, characterized in that, The deployment of accelerometers and edge computing nodes at key monitoring points of the flexible support continuously monitors the real-time statistical characteristics of acceleration. If the characteristic value exceeds a preset quiet state threshold, an event activation state is entered. Specifically: Accelerometers and edge computing nodes are deployed at key monitoring points of the flexible support. In each computing cycle, the nodes are configured according to a window... Calculate the root mean square value of the triaxial resultant acceleration. ;in, N It is a window Number of sampling points within, It is the first The composite acceleration value of each sampling point; When the feature value is greater than the quiet state threshold The duration exceeds the activation duration When a significant vibration event occurs, the system determines that the event has occurred and switches to the event activation state. When the eigenvalue < The duration exceeds the hibernation delay time When the event is considered over, the system returns to a quiet standby state, uploading only a status summary and heartbeat packet.

3. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 1, characterized in that, The local processing and feature extraction to obtain the feature vector are as follows: Local processing involves noise reduction and filtering, using moving averages or digital bandpass filters to remove high-frequency noise and low-frequency drift. Feature extraction is a lightweight extraction of time-domain and frequency-domain feature vectors, including the maximum, minimum, mean, root mean square, standard deviation, amplitude, skewness, kurtosis, and dominant frequency components and frequency band energy proportions extracted through fast Fourier transform.

4. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 3, characterized in that, The lightweight time-domain and frequency-domain feature vector extraction is specifically as follows: Calculate the time-domain characteristics of the filtered triaxial data: maximum value, minimum value, mean, root mean square value, standard deviation, amplitude, skewness, and kurtosis. Frequency domain characteristics: dominant frequency, frequency band energy ratio; ultimately generating the feature vector of the analysis window. F ; in, for x The root mean square value of the axis represents x The effective energy or average intensity of vibration in the axial direction; for x The standard deviation of the axis represents x The degree of dispersion of shaft vibration data around the mean; for x The skewness of the axis indicates x Asymmetry in the probability distribution of shaft vibration signals; for x The kurtosis of the axis indicates the sharpness of the vibration waveform; express x The first dominant frequency of axis data; Indicates that flexible stents are in x The proportion of vibration energy falling within frequency band 1 in the axial direction; For have y The root mean square value of the axis represents y The effective energy or average intensity of axial vibration; This represents the total intensity of triaxial vibration; Indicates the first dominant frequency of the three-axis data; This indicates the proportion of the vibration energy of the flexible support in the three axial directions that falls within frequency band 1.

5. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 1, characterized in that, The lightweight diagnostic model built into the edge computing node performs real-time inference on the generated feature vectors, completes anomaly detection and fault classification, and outputs preliminary diagnostic results and confidence levels. Specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

6. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 1, characterized in that, The diagnostic decision is specifically as follows: If the diagnosis is a minor abnormality or a known pattern, a structured diagnostic report is generated and uploaded to the cloud platform or regional gateway in a low-data-volume format. If a serious abnormality or unknown danger mode is diagnosed, a local audible and visual alarm will be triggered directly within milliseconds, and the highest priority alarm information will be sent to the superior node and cloud platform simultaneously. Edge collaboration involves multiple edge nodes within a region communicating and comparing diagnostic results from adjacent measurement points to rule out isolated sensor failures or locate the approximate area of ​​damage.

7. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 6, characterized in that, The cloud platform aggregates reports and alarm information uploaded by various edge computing nodes for macro-analysis and visualization. Based on the aggregated multi-point data, the lightweight diagnostic model is iteratively optimized, and the updated model parameters are incrementally pushed to the edge computing nodes to achieve continuous evolution of diagnostic capabilities.

8. The real-time diagnostic method for acceleration data of flexible stents based on edge computing according to claim 1, characterized in that, The local rule base specifically refers to: Rule 1: If the anomaly detection model outputs a normal result, it is directly judged as normal and a short log is generated. Rule 2: If the anomaly detection model outputs an anomaly and the fault classification model outputs vortex-induced vibration with a confidence level > 80%, then it is determined to be a known minor anomaly - vortex-induced vibration. Rule 3: If the fault classification model outputs "loose connection" or "structural resonance", it is determined to be a known serious anomaly. Rule 4: If the anomaly detection model outputs an anomaly, but the confidence level of the fault classification model is less than 60%, it is determined to be an unknown danger mode.

9. A real-time diagnostic system for acceleration data of a flexible stent based on edge computing, characterized in that, include: The event triggering module is used to deploy acceleration sensors and edge computing nodes at key monitoring points of the flexible support to continuously monitor the real-time statistical characteristics of acceleration. If the characteristic value exceeds the preset quiet state threshold, it enters the event activation state. The feature vector acquisition module is used by the edge computing node to perform local processing and feature extraction on the raw acceleration data within a fixed time window to obtain a feature vector. The lightweight diagnostic reasoning module is used to embed a lightweight diagnostic model in the edge computing node, perform real-time reasoning on the generated feature vector, complete anomaly detection and fault classification, and output preliminary diagnostic results and confidence levels. The diagnostic decision module is used by the edge computing node to integrate the reasoning results of the lightweight diagnostic model, historical data and local rule base to form a diagnostic decision, thereby realizing real-time diagnosis of the acceleration data of the flexible stent.

10. The real-time diagnostic system for acceleration data of flexible stents based on edge computing according to claim 9, characterized in that, The lightweight diagnostic model built into the edge computing node performs real-time inference on the generated feature vectors, completes anomaly detection and fault classification, and outputs preliminary diagnostic results and confidence levels. Specifically: Lightweight diagnostic models include anomaly detection models and fault classification models; The anomaly detection model adopts the isolated forest model, which is trained based on the historical health status data of the flexible stent. The feature vector is input into the isolated forest model, and the anomaly score is output by calculating the path length of the isolated sample. The shorter the path length, the higher the probability of an anomaly. The anomaly score is compared with the preset anomaly threshold. If the anomaly score exceeds the preset anomaly threshold, the current state is determined to be abnormal, the edge computing node triggers the local alarm flag and activates the fault classification process. The fault classification model employs a lightweight gradient boosting tree model, which is activated after the anomaly detection model identifies an anomaly or periodically. It is used to classify fault types, including vortex-induced vibration, buffeting, loose connections, structural resonance, and external impact. Feature vectors combined with short time series segments are input into the fault classification model, which outputs probability vectors for each fault type. The fault type with the highest probability in the probability vector is selected as the preliminary diagnostic result. And use the maximum probability value as the diagnostic confidence level. Where C is the preset number of fault categories, Indicates belonging to the first c The probability of a class.

Images