Cable bridge fault prediction and health management method based on industrial internet

Abstract

The application relates to the technical field of fault prediction and health management, and particularly discloses a cable bridge fault prediction and health management method based on an industrial internet. A lightweight neural network embedded with a cable physical equation is deployed on an edge gateway to generate compressed features containing physical residuals; a conditional triggering mechanism is used to selectively upload original high-frequency signals; a hierarchical ensemble learning model is constructed on the cloud to perform multi-view verification and global trend analysis, and a high-fidelity digital twin model is dynamically downloaded to the edge end when a suspected fault occurs to realize fine local diagnosis; and the cloud and edge results are fused to output a fault type identification and health assessment report. Through cloud-edge collaboration and physical information fusion, the application improves the fault identification accuracy and system generalization ability under complex working conditions, supports whole-process closed-loop predictive maintenance, prolongs the safe service period of the cable bridge and reduces the risk of unplanned downtime.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of fault prediction and health management, specifically relating to a method for fault prediction and health management of cable trays based on the Industrial Internet. Background Technology

[0002] With the continuous maturation of Industrial Internet technology, its application in the operation and maintenance monitoring of power infrastructure has become a core path to ensure industrial power safety and improve system robustness. Cable trays, as key components supporting and protecting cables, directly affect the overall stability of the power transmission system. In the context of digital transformation, utilizing IoT sensing technology and big data analytics for real-time monitoring of cable trays aims to achieve a leap from traditional passive maintenance to proactive predictive maintenance through the precise capture of key operating parameters.

[0003] Fault prediction and health management of cable trays are crucial for ensuring the long-term safe operation of large-scale power networks. This process typically involves sensing various parameters such as partial discharge signals, ambient temperature, and mechanical stress in the cables. Raw data is collected through monitoring nodes deployed in industrial sites and then aggregated to a processing center via communication networks. The fundamental principle is to build predictive models and, through correlation analysis between historical data and real-time monitoring indicators, assess the degree of damage to the cable trays and predict potential evolutionary risks.

[0004] Existing technologies face a trade-off between limited communication bandwidth and computing resources when processing high-frequency partial discharge signals from cable trays, making it difficult to achieve real-time cloud transmission of all raw data. Purely data-driven models often suffer from insufficient prediction accuracy and weak generalization ability when dealing with small-sample faults in industrial settings due to a lack of physical mechanism constraints. Furthermore, the strong electromagnetic interference in industrial environments makes single monitoring methods prone to high false alarm rates, and traditional models struggle to achieve dynamic coordination between lightweight edge computing and high-fidelity global trend analysis in the cloud. This results in insufficient accuracy in identifying various complex types of damage to meet the demands of precision maintenance. Therefore, a cable tray fault prediction and health management method based on the Industrial Internet is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a method for cable tray fault prediction and health management based on the Industrial Internet, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for cable tray fault prediction and health management based on the Industrial Internet, comprising the following specific steps: Step 1: Deploy a lightweight physical information neural network at the edge gateway near the cable tray. This lightweight physical information neural network receives raw monitoring signals from partial discharge sensors, temperature sensors, and stress sensors, and embeds the transmission line equation or thermal diffusion partial differential equation of the cable as physical constraints into the network structure to generate a compressed feature vector containing physical residuals. Step 2: Upload the compressed feature vector to the cloud processing platform through the industrial communication network, while retaining the original high-frequency signal locally cached at the edge, and only activating the high-bandwidth transmission mechanism when a preset condition is triggered; Step 3: Build a hierarchical ensemble learning model on the cloud processing platform. This hierarchical ensemble learning model consists of multiple base classifiers. It uses a voting mechanism to re-verify the compressed feature vectors uploaded from the edge from multiple perspectives, and combines historical running data to perform global trend analysis and output preliminary fault category judgment results. Step 4: When the edge detects that the partial discharge energy exceeds the preset threshold or the impedance continuity index deviates from the normal range, a digital twin-assisted diagnosis request is sent to the cloud. Step 5: The cloud dynamically generates a high-fidelity digital twin model that matches the current cable tray status based on the request, and downloads part of the inference logic of the high-fidelity digital twin model to the edge for fine-grained local diagnosis of the cached original high-frequency signals. Step 6: Integrate the local diagnostic results at the edge with the global judgment results in the cloud to generate the final fault type identification conclusion and health status assessment report, and push them to the operation and maintenance management terminal.

[0007] Preferably, the physical information neural network in step 1 adopts a hybrid architecture of convolution and fully connected layers. Its loss function is composed of a data fitting term and a physical residual term. The physical residual term is calculated by numerically approximating the time derivative and spatial gradient of the input signal and then inputting it into the cable physical equation, ensuring that the network output is within the range allowed by physical laws.

[0008] Preferably, the preset conditions in step 2 include the partial discharge pulse repetition frequency exceeding a predetermined number of times, the local temperature rise rate being greater than a preset slope, or the mechanical stress change amplitude exceeding the historical fluctuation range. Once any one of these conditions is met, the incremental upload mechanism of the original signal is activated, and only low-frequency feature upload is maintained during other time periods to save communication resources.

[0009] Preferably, in step 3, the hierarchical ensemble learning model first trains multiple independent decision tree classifiers using a Bagging strategy, then uses a Stacking strategy to take the outputs of each classifier as secondary inputs, and performs weighted fusion through a logistic regression unit to finally output the probability distribution of nine typical cable tray damage modes.

[0010] Preferably, the impedance continuity index in step 4 is obtained by calculating the rate of change of the high-frequency current phase difference and amplitude ratio between adjacent sampling points in real time at the edge end. When the rate of change exceeds the statistical confidence interval of steady-state operation for several consecutive cycles, it is determined that there is potential insulation degradation or loose connection.

[0011] Preferably, the digital twin model in step 5 is constructed based on a historical fault case library of similar equipment accumulated in the cloud. Its geometric parameters, material properties and boundary conditions are dynamically adjusted according to the environmental and load information uploaded by the current edge terminal to ensure a high degree of consistency between the model and the actual physical object.

[0012] Preferably, the health status assessment report in step 6 includes the current fault confidence level, remaining safe operating time prediction, recommended maintenance priority, and historical similar case reference information. All content is encapsulated in a structured data format and pushed to the designated terminal through a secure protocol.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a physical information neural network at the edge, embedding the physical laws of the cable system into the model training process. Under limited sample conditions, it improves the accuracy and generalization ability of partial discharge signal identification and alleviates the problem that the original data cannot be fully uploaded due to limited communication bandwidth.

[0014] 2. This invention achieves the organic unity of lightweight real-time response on the edge side and high-precision global analysis on the cloud by constructing a layered diagnostic architecture that integrates cloud and edge, thus avoiding the performance bottleneck of a single model in complex industrial scenarios. By dynamically triggering the digital twin model download mechanism, high-fidelity simulation resources are accurately invoked when a suspected fault occurs, ensuring the depth of diagnosis and avoiding continuous high computing power consumption.

[0015] 3. By integrating multi-source sensor information and physical residual characteristics, this invention significantly suppresses false alarms caused by environmental electromagnetic interference, thereby systematically improving the system's accuracy in identifying various complex types of damage. The overall solution supports closed-loop management of the entire process from state perception, feature extraction, fault identification to health assessment, providing reliable technical support for predictive maintenance of cable trays, extending the safe service life of power infrastructure and reducing the risk of unplanned downtime. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the lightweight physical information neural network embedded with cable physical constraints in this invention; Figure 3This is a logical flowchart of the cloud processing platform in this invention that uses layered integrated learning for multi-perspective fault verification. Figure 4 This is a schematic diagram of the reasoning logic download and multi-level interaction relationship and data flow between the edge and cloud processing platforms based on the digital twin model in this invention; Figure 5 This is a flowchart illustrating the logical process of generating the final conclusion and assessing health in this invention, which integrates edge local diagnosis and cloud global judgment. Detailed Implementation

[0017] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0018] In a specific embodiment of the present invention, the core of the cable tray fault prediction and health management method based on the Industrial Internet lies in achieving deep perception and accurate diagnosis of the operating status of power cables through a distributed computing framework that combines cloud and edge computing.

[0019] Step 1 involves deploying a lightweight physical information neural network at an edge gateway located near the cable tray. The edge gateway employs a high-performance industrial-grade processor and is equipped with a dedicated artificial intelligence acceleration unit to meet the computational power requirements of the neural network inference. The edge gateway receives raw monitoring signals in real time from a multi-source sensor array deployed at key locations on the cable tray via an industrial bus, wireless sensor network, or fiber optic communication interface. The multi-source sensor array includes at least a partial discharge sensor, a temperature sensor, and a stress sensor.

[0020] The partial discharge sensor employs an ultra-high frequency sensor or a high-frequency current transformer to capture partial discharge pulse signals with frequencies ranging from 30 MHz to 300 MHz or from 3 MHz to 30 MHz. The temperature sensor uses a distributed fiber optic temperature measurement system or an infrared thermal imaging sensor to acquire real-time temperature field distribution data on the cable surface. The stress sensor employs a fiber optic stress sensor or a piezoelectric strain gauge to monitor stress changes in the cable tray caused by gravity, thermal expansion and contraction, or external mechanical vibration during long-term operation.

[0021] In step 1 above, the lightweight physical information neural network adopts a hybrid architecture of convolutional and fully connected layers. The convolutional layers are used to automatically extract multi-scale spatial features from the original one-dimensional time-series signal, such as the waveform features of partial discharge pulses, the distribution features of temperature gradients, and the frequency features of stress waves. The fully connected layers are used to perform nonlinear mapping on the extracted features to generate feature vectors to be processed. To improve the network's generalization ability under small sample conditions, this embodiment embeds the transmission line equation or thermal diffusion partial differential equation of the cable as a physical constraint into the structure of the neural network.

[0022] The loss function of the neural network consists of a data fitting term and a physical residual term. The data fitting term measures the deviation between the network's predicted values ​​and the sensor's measured values. The physical residual term is a regularization term constructed based on the cable's physical equations. When processing temperature signals, the partial differential equation of thermal diffusion is used, i.e., the partial derivative of temperature with respect to time is equal to the thermal diffusivity multiplied by the second-order partial derivative of temperature with respect to spatial location.

[0023] The physical residual term is calculated by numerically approximating the time derivative and spatial gradient of the input signal and then inputting them into the above equation. For example, the spatial gradient is calculated using the central difference method, and the time derivative is calculated using the forward difference method. The calculated values ​​are then substituted into the left and right sides of the equation, and the square of the difference is calculated to obtain the physical residual. During the training process, the neural network is required not only to fit the observed data but also to ensure that its output is within the range allowed by physical laws, thus ensuring that the generated compressed feature vector has a clear physical meaning. Finally, the lightweight physical information neural network outputs a low-dimensional compressed feature vector containing physical residual information.

[0024] Step 2 involves uploading the compressed feature vector to the cloud processing platform via an industrial communication network. This industrial communication network can be based on 5G, broadband power line carrier, or industrial Ethernet. To optimize communication bandwidth utilization, this implementation adopts a differentiated data transmission strategy. Under normal operating conditions, the edge gateway only periodically uploads a very small amount of compressed feature vector to the cloud.

[0025] The edge gateway utilizes its internal large-capacity memory to locally cache raw high-frequency signals from various sensors. The sampling frequency of the raw high-frequency signals is much higher than the generation frequency of the feature vectors, preserving the most complete original waveform information. The edge gateway monitors indicators in real time and only activates a high-bandwidth transmission mechanism when preset conditions are triggered, uploading incremental cached raw high-frequency signals to the cloud.

[0026] The preset conditions in step 2 include the following: First, the partial discharge pulse repetition frequency exceeds a predetermined number, for example, the number of discharge pulses captured within 1 second exceeds 100; second, the local temperature rise rate is greater than a preset slope, for example, the temperature rise rate at a certain point in the cable exceeds 2 degrees Celsius per minute; third, the mechanical stress change amplitude exceeds the historical fluctuation range, for example, the output value of the stress sensor deviates from the average value of the past 24 hours by more than 3 times the standard deviation. When any of the above conditions are met, the system determines that it is currently in a suspected fault state, immediately activates high-bandwidth transmission, and maintains low-frequency characteristic uploading during other periods to save communication resources.

[0027] Step 3 involves constructing a hierarchical ensemble learning model on a cloud processing platform. This platform boasts massive storage capacity and powerful computing capabilities, enabling in-depth analysis of multi-source data. The hierarchical ensemble learning model comprises multiple base classifiers with different structures. In this embodiment, multiple independent decision tree classifiers are trained using a Bagging strategy. Each decision tree classifier is trained on a random subset of the original training set, increasing the model's diversity.

[0028] A stacking strategy is employed for model fusion. The predicted outputs of the multiple decision tree classifiers are used as secondary inputs and fed into the meta-classifier. The meta-classifier uses a logistic regression unit, assigning different weights to each base classifier based on its performance on the validation set for weighted fusion. The hierarchical ensemble learning model utilizes a voting mechanism to perform multi-view re-validation of the compressed feature vectors uploaded from the edge, focusing not only on the current feature values ​​but also combining historical operating data of the cable tray, surrounding environmental parameters, and load current records for global trend analysis.

[0029] The hierarchical ensemble learning model outputs preliminary fault category determination results, which are typically probability distributions for nine typical cable tray damage modes. These nine typical damage modes include, but are not limited to: internal insulation air gap discharge, insulation surface creepage, terminal head stress cone displacement, multi-point grounding of the metal sheath, localized overheating due to excessive contact resistance, mechanical fatigue fracture of the cable support, chemical corrosion damage to the outer sheath, biological damage, and surface flashover caused by excessive environmental humidity.

[0030] In step 4, when the edge terminal detects that the partial discharge energy exceeds a preset threshold or the impedance continuity index deviates from the normal range, a digital twin-assisted diagnostic request is initiated to the cloud. The partial discharge energy is obtained by integrating the acquired high-frequency pulse waveform.

[0031] The impedance continuity index in step 4 is a key parameter reflecting the integrity of cable insulation. The edge terminal obtains this index by calculating the rate of change of the phase difference and amplitude ratio of the high-frequency current between adjacent sampling points in real time. The specific operation process is as follows: The edge gateway synchronously acquires the current waveforms at both ends of the cable or different detection points, and extracts the phase and amplitude of the frequency components through discrete Fourier transform. It calculates the change in phase difference between adjacent moments and the fluctuation range of the amplitude ratio. When this rate of change exceeds the statistical confidence interval of steady-state operation for several consecutive monitoring cycles, the system determines that there is potential insulation degradation or loose connection. At this time, the edge terminal sends a request command containing the current real-time parameters to the cloud.

[0032] In step 5, the cloud dynamically generates a high-fidelity digital twin model that matches the current state of the cable tray based on the request. The digital twin model is a full-element mapping of the physical entity in digital space, including not only the geometric three-dimensional model, but also material properties, physical field coupling equations, and operational mechanism models.

[0033] The digital twin model is built upon a historical fault case library of similar equipment accumulated in the cloud. The cloud server first obtains static information about the cable segment, such as its model, laying method, service life, and historical maintenance records, based on the asset number uploaded from the edge device. Then, based on real-time environmental temperature, humidity, load current, and soil thermal resistance information uploaded from the edge device, the boundary conditions of the digital twin model are dynamically adjusted. For example, the thermal conductivity parameters of the materials in the finite element mesh are adjusted to match the actual aging state of the cable.

[0034] After generating a high-fidelity digital twin model, the cloud does not directly run a time-consuming full simulation. Instead, it extracts key inference logic specific to the current suspected fault type and downloads it to the edge. This inference logic can be a trained fast correlation operator for waveforms or a lookup table model based on simplified physical equations. Upon receiving this logic, the edge performs refined local diagnosis using locally cached raw high-frequency signals. In this way, the edge can perform deep feature comparisons of signals using prior knowledge generated in the cloud, without needing to upload all the raw data.

[0035] Step 6 integrates the local diagnostic results at the edge with the global judgment results in the cloud to generate the final fault type identification conclusion and health status assessment report.

[0036] The fusion process employs a fusion algorithm based on evidence theory. Local diagnostic results from the edge are considered local observational evidence, while global trend analysis results from the cloud are considered prior evidence. By calculating the degree of conflict between the two, fusion rules are used to calculate the final confidence distribution of each fault mode. The final output includes not only the identified fault types and their probabilities but also a comprehensive assessment of the health status.

[0037] The health status assessment report includes the following core elements: First, the current fault confidence level, expressed as a percentage to indicate the system's certainty regarding the conclusion; second, the remaining safe operating time prediction, based on the cable insulation aging model and current operating conditions, estimating the remaining lifespan before the fault deteriorates to failure; third, recommended maintenance priorities, categorizing maintenance needs into three levels—urgent, important, and general—based on the severity, scope of impact, and speed of development of the fault; and fourth, historical similar case reference information, automatically linking the three historical cases in the cloud database with the highest matching degree to the current signal characteristics, providing maintenance plan references. All content is encapsulated in a structured data format and pushed to the mobile terminals of operation and maintenance personnel or the large screen in the monitoring center via a secure transmission protocol.

[0038] Furthermore, in the specific implementation of the above method, the cable transmission line equation described in step 1 is transformed into a textual logical form: the partial derivative of the voltage at any point on the cable with respect to spatial location is equal to the resistance multiplied by the current plus the inductance multiplied by the partial derivative of the current with respect to time, which is the negative value; simultaneously, the partial derivative of the current with respect to spatial location is equal to the conductance multiplied by the voltage plus the capacitance multiplied by the partial derivative of the voltage with respect to time, which is the negative value. The physical residual term is the residual deviation on both sides of the above equation. By adding this residual deviation term to the loss function, the neural network is forced to satisfy the electromagnetic characteristics of electromagnetic wave propagation when learning features, thus eliminating a large amount of interference from environmental white noise.

[0039] Furthermore, in the specific implementation of the above method, the hierarchical ensemble learning model described in step 3 includes not only decision trees as its base classifiers, but also support vector machines and deep residual networks. Through this heterogeneous integration, partial discharge signals can be analyzed from multiple perspectives, including the frequency domain, time domain, and time-frequency domain, to identify fault modes such as "point discharge" and "surface discharge," which have similar waveforms but completely different physical mechanisms.

[0040] Furthermore, in the specific implementation of the above method, the inference logic download process described in step 5 is implemented using a dynamic link library or script container. After receiving the download instruction, the edge gateway starts a sandbox process in a restricted operating environment and runs the lightweight inference code sent from the cloud. This mechanism ensures that when the system deals with new failure modes, it does not need to perform a full update of the edge firmware; it only needs to download the corresponding model parameters or discrimination algorithm logic to achieve functional upgrades.

[0041] Example 2: As a further supplement and refinement to Example 1, this example describes in detail an application scenario for high-voltage cable trays. In high-voltage scenarios, the electromagnetic interference environment is extremely complex, placing higher demands on the accuracy of fault prediction.

[0042] In this embodiment, in step 1 of the cable tray fault prediction and health management method based on the Industrial Internet, the edge gateway employs multi-channel synchronous sampling technology. For partial discharge signals, the sampling rate is set to 500 mega-samples per second to ensure complete reproduction of the nanosecond-level rising edge of the partial discharge pulse. The physical information neural network adds an attention mechanism module after the convolutional layer. This attention mechanism module automatically focuses on key time windows containing fault features in the signal by calculating the correlation scores at different locations in the feature map, thus shielding against cluttered pulses generated by background electromagnetic interference.

[0043] In step 2, the triggering conditions for the incremental upload mechanism are further refined. The system establishes a baseline database based on normal operating conditions. The edge gateway measures the shift in probability distribution by calculating the Kübk-Leibler divergence (KL divergence) between the current signal and the baseline signal in real time. When the Kübk-Leibler divergence value continuously exceeds a set threshold for 10 statistical periods, it is immediately determined to be an abnormal state, triggering the complete upload of the original data.

[0044] In step 3, the global trend analysis in the cloud specifically incorporates environmental and climatic coupling factors. The system automatically acquires rainfall, atmospheric pressure, and air humidity in the area where the cable tray is located. During the rainy season, increased air humidity may alter the surface charge distribution of the termination heads, inducing weak partial discharge. By learning the periodic characteristics of these environmental variables, the hierarchical ensemble learning model in the cloud can distinguish such recoverable anomalies caused by the environment from true internal insulation degradation, reducing the system's false alarm rate.

[0045] In step 4, the impedance continuity calculation considers not only high-frequency components but also the phase of the power frequency current. By monitoring the ratio of circulating current to main current, a nonlinear step change in this ratio indicates a grounding system fault. The digital twin-assisted diagnostic request includes a high-precision GPS timestamp, ensuring that the cloud can align data from different monitoring points at the nanosecond level, achieving precise fault location based on the traveling wave principle.

[0046] In step 5, the generated digital twin model includes a detailed electrothermal coupling field model. The cloud simulation demonstrates the evolution of the local temperature field over time under the current load, assuming an abnormal increase in contact resistance at a certain point. The logic downloaded to the edge contains a predefined set of temperature evolution curves. By matching the measured temperature rise curves with these simulated curves, the edge can directly infer the physical size and severity of internal defects.

[0047] In step 6, the generated report not only provides conclusions but also includes an augmented reality (AR)-based 3D visualization interface. Maintenance personnel can scan tags on cable trays using mobile devices to view a digital twin model overlaid on the physical structure on a screen, intuitively observing the internal temperature distribution, electric field distribution, and the location of warning points.

[0048] Example 3: In this example, the distributed deployment architecture and large-scale collaboration method of the present invention in a cross-regional industrial internet platform are described in detail.

[0049] When deployed in a provincial cable network covering hundreds of substations, the edge gateways in step 1 are managed using containerization technology. Weight updates for the physical information neural network are achieved through a federated learning mechanism. Each site's edge gateway trains its model locally using private data, only anonymizing the model parameter updates before uploading them to the cloud. The cloud aggregates the parameters and then distributes the optimized model to all gateways. This approach protects power data privacy while enabling the sharing of fault characteristic knowledge across the entire network, improving the system's ability to identify rare faults.

[0050] In step 2, the industrial communication network uses edge computing nodes (MECs) for secondary aggregation. At the edge of the access network, the data undergoes initial cleaning to remove abnormal packets caused by communication link jitter. The compressed feature vector employs a binary encapsulation format based on a lossless compression algorithm to further reduce transmission latency in narrowband wireless environments.

[0051] In step 3, the cloud processing platform employs a massively parallel computing framework based on distributed computing power scheduling. When multiple cable trays simultaneously initiate diagnostic requests, the load balancer dynamically allocates processing tasks based on the current resource occupancy of each computing node. The base classifiers in the hierarchical ensemble learning model are deployed in different virtual machines or containers, and intermediate prediction results are exchanged and synchronized through message brokers.

[0052] In step 4, a topology identification algorithm is introduced to monitor the impedance continuity index. The edge end can automatically identify the location of cable joints and branch structures, and adjust the threshold value for impedance mutation judgment accordingly. At cable branches, impedance discontinuities naturally exist. The system learns the characteristics of these structural discontinuities through training and eliminates them from impedance fluctuations caused by faults.

[0053] In step 5, the digital twin model generation process incorporates automated 3D modeling technology. By calling Building Information Modeling (BIM) data from the cable engineering construction phase and combining it with point cloud data obtained from UAV inspections, the cloud can quickly construct a digital twin of the cable tray with geographic coordinate information. The downloaded inference logic undergoes model compression processing, converting 32-bit floating-point operations into 8-bit fixed-point operations, enabling the edge gateway to complete refined diagnostic tasks with lower energy consumption.

[0054] In step 6, the health status assessment report is deeply integrated with the enterprise's asset management system (ERP) and production scheduling system. When the assessment results indicate that a certain cable segment has an extremely high risk of failure, the system automatically generates a pre-maintenance work order and provides an optimal outage maintenance time window suggestion based on the current grid load and maintenance personnel schedule. Simultaneously, the report includes an analysis of the potential impact of the failure on grid reliability indicators.

[0055] Example 4: This example focuses on describing the specific implementation details of the present invention in response to extreme environmental conditions.

[0056] In the aforementioned extreme environments, sensors are highly susceptible to damage and difficult to repair. The lightweight physical information neural network in step 1 incorporates a sensor state integrity monitoring module at the input layer. This module analyzes the statistical characteristics of the sensor's output signal to determine in real-time whether the sensor has malfunctioned, drifted, or experienced poor contact. If a sensor fails, the network automatically switches to a robust inference mode, using the constraints of physical equations and combining the remaining temperature and partial discharge signals to virtually reconstruct the missing stress characteristics. For example, using the principle of thermal expansion, equivalent mechanical strain is derived from temperature changes, ensuring the continuity of the diagnostic process.

[0057] In step 2, to address the extremely low bandwidth of deep-sea communication, a "key feature extraction" mode was added to the preset conditions. The edge gateway utilizes autoencoder technology to compress the high-dimensional original waveform to an extremely low dimension, retaining only key parameters such as the waveform's envelope and spectral centroid. Only when the identified fault confidence exceeds and reaches the level of a "critical fault" is the upload of the original segment initiated via satellite link or underwater acoustic communication.

[0058] In step 3, the cloud-based model specifically enhances its ability to identify "compound damage." In extreme environments, cables often suffer from corrosion, mechanical compression, and overload thermal damage simultaneously. The hierarchical ensemble learning model introduces a multi-task learning framework, simultaneously outputting scores for multiple fault dimensions, rather than simple mutually exclusive classifications.

[0059] In step 4, the impedance continuity calculation incorporates an attenuation compensation algorithm for extremely long-distance cables. Since high-voltage cables can reach lengths of tens of kilometers, high-frequency signals experience severe attenuation and distortion during transmission. The edge gateway performs reverse compensation on the signals acquired from the remote end based on preset cable parameters. The compensated signals are then used to calculate the phase difference and amplitude ratio, ensuring the sensitivity of long-distance monitoring.

[0060] In step 5, the digital twin model includes an environmental stress simulation unit. For submarine cables, the model incorporates real-time data on ocean current velocity, seawater temperature, and salinity; for high-altitude and cold regions, it incorporates data on snow depth and permafrost depth. These real-time simulation results of external physical fields are mapped into the cable's mechanical analysis model. The downloaded logic includes fatigue life assessment operators for external loads, enabling the edge device to adjust its diagnostic sensitivity and warning frequency in real time based on the current external environmental intensity.

[0061] In step 6, the health status assessment report includes emergency response guidelines. Addressing maintenance challenges in extreme environments, the report lists the required special tools and equipment, the shortest repair route, and the anticipated consumption of spare parts. In this way, the system not only achieves "prediction" but also "closed-loop management."

[0062] In summary, the method provided in this embodiment achieves a deep fusion of mechanistic models and data-driven models through a physical information neural network. The physical equations, as a form of prior knowledge, regulate the search space of the neural network, enabling the model to maintain extremely high discrimination accuracy even in the constrained computational environment of the edge.

[0063] The cloud-edge collaborative architecture cleverly balances the needs of "real-time response" and "deep analysis." Preliminary diagnosis at the edge ensures timely fault detection and avoids the ineffective transmission of massive amounts of useless data; the global perspective and digital twin-assisted diagnosis in the cloud guarantee the authority and depth of the diagnostic conclusions. This approach is not only applicable to cable tray management in ordinary industrial plants and commercial buildings, but also plays a crucial role in fields with extremely high reliability requirements, such as high-voltage power grids, petrochemicals, rail transportation, and offshore wind power. Through full lifecycle health management, this method can reduce the operation and maintenance costs of cable systems, minimize the huge economic losses caused by unplanned downtime, and improve the overall intelligence level of power infrastructure.

[0064] The structural parameters, coefficients of the physical equations, and preset thresholds of the neural network can all be adjusted and optimized according to the specific cable type and application environment. For cross-linked polyethylene insulated cables and oil-paper insulated cables, the loss term coefficients in their transmission line equations will differ significantly. For cables of different voltage levels, the preset thresholds for partial discharge energy should also be set in a corresponding stepped manner. These parameter fine-tuning based on practical engineering experience all fall within the scope of protection of this invention.

[0065] The method described in this invention is not limited to processing the three types of sensor signals mentioned above. In practical deployments, circulating current sensors, leakage current sensors, vibration sensors, and even microphones can be connected as needed. All multi-source data can undergo feature extraction, physical constraint embedding, and cloud-based fusion verification according to the aforementioned cloud-edge collaborative processing framework. The digital twin model can also be extended as needed to include peripheral facilities such as cable branch boxes and terminal towers, constructing a more complete digital twin scenario and realizing collaborative management of the entire power distribution network.

[0066] In terms of technical implementation details, the interaction between the edge gateway and the cloud platform can adopt standard protocols such as Message Queuing Telemetry Transport Protocol (MQTT), Constrained Application Protocol (CoAP), or Advanced Message Queuing Protocol (AMQP) to ensure the system's openness and compatibility. Data is encrypted using asymmetric encryption algorithms during transmission to ensure the security of industrial production data. The cloud platform can be deployed in private, public, or hybrid cloud environments, leveraging container cloud technology to achieve elastic scaling of computing resources.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for cable tray fault prediction and health management based on the Industrial Internet, characterized in that, Includes the following steps: A lightweight physical information neural network is deployed at an edge gateway near the cable tray. The edge gateway receives raw monitoring signals from partial discharge sensors, temperature sensors, and stress sensors, and embeds the physical equations of the cable as physical constraints into the network structure to generate a compressed feature vector containing physical residuals. The compressed feature vector is uploaded to the cloud processing platform through the industrial communication network. At the same time, the original high-frequency signal is cached locally at the edge gateway. When a preset condition is triggered, a high-bandwidth transmission mechanism is activated to upload the original high-frequency signal. The cloud processing platform constructs a hierarchical integrated learning model, uses a voting mechanism to re-verify the compressed feature vector from multiple perspectives, and combines historical operating data to perform trend analysis, outputting preliminary fault category determination results. When the edge gateway detects that the partial discharge energy exceeds a preset threshold or the impedance continuity index deviates from the normal range, it initiates a digital twin-assisted diagnosis request to the cloud processing platform. The cloud processing platform generates a digital twin model based on the digital twin-assisted diagnosis request, and downloads part of the reasoning logic of the digital twin model to the edge gateway for refined local diagnosis of the cached original high-frequency signal; By integrating the local diagnostic results of the edge gateway with the preliminary fault category determination results of the cloud processing platform, a fault type identification conclusion and a health status assessment report are generated.

2. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The lightweight physical information neural network adopts a hybrid architecture of convolution and fully connected layers. The convolutional layer is used to extract multi-scale spatial features from the original time series signal, and the fully connected layer is used to perform nonlinear mapping on the extracted features to generate the compressed feature vector. The loss function of the lightweight physical information neural network consists of a data fitting term and a physical residual term. The data fitting term is used to measure the deviation between the network prediction and the sensor measurement. The physical residual term is calculated by numerically approximating the time derivative and spatial gradient of the input signal and then inputting it into the physical equation of the cable. The calculation process of the physical residual term is as follows: the spatial gradient is calculated using the central difference method, the time derivative is calculated using the forward difference method, the calculated value is substituted into both sides of the physical equation of the cable, the square of the difference between the values ​​on both sides of the equation is calculated, and the physical residual term is used as a regularization term to constrain the network output.

3. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 2, characterized in that: The physical equations of the cable include transmission line equations or partial differential equations of thermal diffusion. The transmission line equation is expressed as follows: the partial derivative of the voltage at any specified location on the cable with respect to the spatial location is equal to the negative value of the result of multiplying the resistance by the current and then adding the partial derivative of the inductance by the current with respect to time; the partial derivative of the current with respect to the spatial location is equal to the negative value of the result of multiplying the conductance by the voltage and then adding the partial derivative of the capacitance by the voltage with respect to time. The partial differential equation for thermal diffusion is expressed in the following textual logic: the partial derivative of temperature with respect to time is equal to the thermal diffusivity multiplied by the second partial derivative of temperature with respect to spatial location. The lightweight physical information neural network, by introducing the physical residual term into the loss function, forces its output to be within the range defined by physical laws, and uses an attention mechanism module to calculate the correlation score at different positions in the feature map, focusing on the time window in the original monitoring signal that contains fault features.

4. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The preset conditions include any one of the following: the partial discharge pulse repetition frequency exceeds a predetermined number of times, the local temperature rise rate is greater than a preset slope, or the mechanical stress change amplitude exceeds the historical fluctuation range. The partial discharge pulse repetition frequency exceeding a predetermined number is determined when the number of discharge pulses captured within a unit time period is greater than a preset number. The local temperature rise rate being greater than the preset slope is determined when the temperature rise rate at the monitoring point on the cable surface is greater than the preset time change rate. The mechanical stress variation exceeding the historical fluctuation range is determined when the deviation of the real-time output value of the stress sensor from the historical average value is greater than a preset standard deviation. The edge gateway calculates the Kübk and Leibler divergence between the probability distribution of the current monitored signal and the probability distribution of the baseline database signal. When the Kübk and Leibler divergence exceeds a set threshold for multiple consecutive statistical periods, it activates the incremental upload mechanism of the original high-frequency signal.

5. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The hierarchical ensemble learning model is constructed through a heterogeneous ensemble strategy, and its base classifiers include multiple independent decision tree classifiers, support vector machines, and deep residual networks. The hierarchical ensemble learning model first trains the base classifier through an ensemble strategy, enabling it to perform feature analysis on the compressed feature vector from multiple perspectives, including the frequency domain, time domain, and time-frequency domain. A secondary fusion strategy is adopted, in which the predicted outputs of each base classifier are used as secondary inputs to the meta classifier. The meta classifier uses a logistic regression unit and assigns different weights to each base classifier based on its performance on the validation set for weighted fusion. The output is the probability distribution of nine types of cable tray damage modes, including internal air gap discharge, surface creepage of insulation, stress cone displacement of terminal head, multi-point grounding of metal sheath, contact resistance overheating, mechanical fatigue fracture of support, chemical corrosion damage of outer sheath, biological damage, and environmental humidity flashover.

6. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The impedance continuity index is calculated through the following steps: the edge gateway acquires the current waveforms at different detection points of the cable, and extracts the phase and amplitude of the frequency components through discrete Fourier transform; Calculate the changes in phase and amplitude at adjacent time points to obtain the change in phase difference and amplitude of high-frequency current between adjacent sampling points. When the change in phase difference or the fluctuation of amplitude ratio exceeds the statistical confidence interval of steady-state operation for multiple consecutive monitoring cycles, it is determined that there is insulation deterioration or loose connection. When monitoring long-distance cables, the edge gateway performs reverse attenuation compensation on the original monitoring signal based on preset cable resistance, inductance, capacitance and conductance parameters, and uses the compensated signal to calculate the phase difference and amplitude ratio.

7. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The digital twin model is dynamically constructed based on the historical fault case library of similar equipment, static asset information, and real-time environmental information stored on the cloud processing platform. The static asset information includes cable type, laying method, service life, and historical maintenance records; The real-time environmental information includes ambient temperature, ambient humidity, load current, and soil thermal resistance. The cloud processing platform dynamically adjusts the boundary conditions of the digital twin model based on the real-time environmental information. By adjusting the thermal conductivity parameters of the material in the finite element mesh, it matches the actual aging state of the current cable and constructs an electrothermal coupling field model to simulate the temperature field evolution process caused by the abnormal increase in contact resistance under the current load. The digital twin model also includes an environmental stress simulation unit, which receives external physical field data in real time and maps it into the mechanical analysis model of the cable to achieve fatigue life assessment for external loads.

8. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The reasoning logic download process is implemented using script containers or dynamic link libraries. Upon receiving the download instruction, the edge gateway initiates a sandbox process in a restricted operating environment and runs the inference code issued by the cloud processing platform. The inference code includes fast operators for waveforms or lookup table models based on simplified physical equations; The inference code undergoes model compression before downloading, and floating-point operations are converted into fixed-point operations through pruning and quantization. The edge gateway uses the inference code to perform fine-grained feature comparison on the locally cached original high-frequency signal. Without needing to upload all the original data, it uses the prior knowledge generated by the cloud processing platform to complete a deep-level judgment of the fault mode.

9. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The fusion process employs a fusion algorithm based on evidence theory, using the local diagnostic results of the edge gateway as local observation evidence and the preliminary fault category determination results of the cloud processing platform as prior evidence. By calculating the degree of conflict between the local observation evidence and the prior evidence, the final confidence distribution of each fault mode is calculated. The health status assessment report includes the current fault confidence level, remaining safe operating time prediction, recommended maintenance priority, and historical similar case reference information; The remaining safe operating time prediction is based on the cable insulation aging model and the current operating conditions. The recommended maintenance priority is divided into three levels: urgent, important and general, according to the severity of the fault. The historical similar case reference information is obtained from the historical cases with the highest feature matching degree in the association database.

10. The cable tray fault prediction and health management method based on the Industrial Internet according to claim 1, characterized in that: The weight update of the lightweight physical information neural network is achieved through a federated learning mechanism. Each edge gateway at each site trains the model locally using private data, and uploads the gradient information of the model parameters to the cloud processing platform after desensitizing it. The cloud processing platform then aggregates the parameters and distributes the optimized weight parameters to each edge gateway. The lightweight physical information neural network is equipped with a sensor state integrity monitoring module in the input layer, which determines whether the sensor has malfunctioned or drifted by analyzing the mean, variance and autocorrelation function statistical characteristics of the signal. If a sensor failure is detected, the lightweight physical information neural network switches to robust inference mode, using the constraints of the physical equations in combination with the signals from the remaining sensors to virtually reconstruct the missing physical features.

Images