Edge calculation and adaptive data hierarchical transmission method for overhead line system monitoring

Abstract

The invention relates to the cross technical field of artificial intelligence and engineering monitoring, in particular to an edge calculation and adaptive data hierarchical transmission method for catenary monitoring, which comprises the following steps of: S1, acquiring monitoring data in real time to obtain original monitoring data; s2, preprocessing the original monitoring data to obtain preprocessed monitoring data; s3, constructing an abnormal confidence model; s4, generating an abnormal confidence score of the preprocessed monitoring data according to the abnormal confidence model, and transmitting the preprocessed monitoring data to a central platform according to a shunt transmission strategy; step S5, optimizing a shunting transmission strategy; s6, adjusting the optimization process of the shunting transmission strategy; and S7, carrying out convergence analysis on the preprocessed monitoring data transmitted to the central platform, and correcting the abnormal confidence model. According to the invention, the real-time performance and reliability of overhead line system monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and engineering monitoring, and in particular to an edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring. Background Technology

[0002] The overhead contact line monitoring points are scattered and the environment is harsh. A large number of vibration, galloping, tension and other sensors are deployed. The existing sensor data is simply packaged by the aggregator and usually transmitted back to the central cloud or big data platform as raw data at a fixed frequency. This method has huge bandwidth and cost pressures, and the transmission cost is high. Moreover, the central platform receives a massive amount of raw data, most of which is redundant data without anomalies, resulting in wasted storage resources and invalid calculations. When bandwidth is limited, important abnormal data segments may queue with a large amount of normal data, resulting in early warning delays.

[0003] Chinese Patent Publication No. CN120455498A discloses a method and system for monitoring data transmission via an edge gateway combined with neural networks. The method includes: Step S1, acquiring a large amount of historical network monitoring data to construct a dataset, training and deploying a network state prediction model using the dataset; Step S2, the edge gateway creating a message circular queue and an edge knowledge graph; Step S3, collecting monitoring data and storing it in the message circular queue; Step S5, collecting real-time network monitoring data and inputting it into the network state prediction model to obtain network state prediction results to match the corresponding batch processing strategy; Step S6, the edge gateway, based on the batch processing strategy and the edge knowledge graph, sequentially preprocesses the monitoring data stored in the message circular queue, encrypts the preprocessed monitoring data into encrypted monitoring data, and uploads it to the server. Therefore, the aforementioned method and system for monitoring data transmission via an edge gateway combined with neural networks lacks the ability to intelligently identify and filter the value of the data itself at the edge, resulting in its transmission optimization being based on a low value density. This makes it difficult to resolve the core contradictions of "data pipeline congestion" and "uneven value density" in overhead contact line monitoring, thus hindering the improvement of the real-time performance and reliability of overhead contact line monitoring. Summary of the Invention

[0004] To address this issue, the present invention provides an edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring, which overcomes the problems of low real-time performance and reliability in existing overhead contact line monitoring technologies.

[0005] To achieve the above objectives, the present invention provides an edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring, the method comprising: Step S1: Real-time acquisition of monitoring data to obtain raw monitoring data; Step S2: Preprocess the raw monitoring data to obtain preprocessed monitoring data; Step S3: Construct an anomaly confidence model; Step S4: Generate anomaly confidence scores for the preprocessed monitoring data based on the anomaly confidence model, and transmit the preprocessed monitoring data to the central platform according to the anomaly confidence scores of the preprocessed monitoring data using a diversion transmission strategy. Step S5: Optimize the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method; Step S6: Adjust the optimization process of the off-transmission strategy according to the changes in network congestion after the off-transmission strategy optimization; Step S7: Perform aggregated analysis on the preprocessed monitoring data transmitted to the central platform and correct the anomaly confidence model.

[0006] Further, in step S3, when constructing the anomaly confidence model, the anomaly confidence model is constructed using an anomaly confidence model construction method, wherein the anomaly confidence model construction method includes: Step A01: Obtain model construction simulation data, which includes normal operating condition data and labeled abnormal data; Step A02, structuring the teacher model, including: The visual branch of the teacher model is set to use a deep residual network to process visible light and infrared images to extract appearance and thermal features; The time-series signal branch of the teacher model is set to use a one-dimensional deep convolutional network to process time-series data and extract time-frequency features; The feature fusion and decision layer of the teacher model is set to fuse the appearance and thermal features and time-frequency features extracted from the visual branch and the temporal signal branch of the teacher model, and the classification output layer outputs the probability distribution vector representing the possibility of various anomalies. Step A03: Transmit the model construction simulation data to the teacher model for supervised learning to obtain the trained teacher model; Step A04: Use knowledge distillation to transfer the parameters of the trained teacher model to the student model; Step A05: Perform low-bit quantization on the student model, converting the student model weight activation values ​​from 32-bit floating-point numbers to 8-bit fixed-point numbers to obtain the quantized student model. Step A06 involves using structured pruning to optimize the quantized model, obtaining an anomaly confidence model, and deploying it in the intelligent edge computing gateway.

[0007] Furthermore, in step S4, when generating the anomaly confidence score of the preprocessed monitoring data according to the anomaly confidence model, the preprocessed monitoring data is input into the anomaly confidence model, and the anomaly confidence model outputs the anomaly confidence score Z. In step S4, when transmitting the preprocessed monitoring data to the central platform according to the anomaly confidence score of the preprocessed monitoring data and the diversion transmission strategy, the anomaly confidence score Z of the transmitted preprocessed monitoring data is compared with the preset confidence score Z0. The priority of the preprocessed monitoring data is determined based on the comparison result, and the preprocessed monitoring data is output based on the determination result. Wherein: The preset reliability score Z0 includes a first preset reliability score Zmax and a second preset reliability score Zmin, where Zmax = 0.8 and Zmin = 0.3. When Zmax < Z, the priority of the preprocessed monitoring data is determined to be level one, and the preprocessed monitoring data is output as level one data. When Zmin<Z≤Zmax, the priority of the preprocessed monitoring data is determined to be level two, and the preprocessed monitoring data is output as level two data; When Z≤Zmin, the priority of the preprocessed monitoring data is determined to be level three, and the preprocessed monitoring data is output as level three data.

[0008] Further, in step S4, when the preprocessed monitoring data is transmitted to the central platform according to the anomaly confidence score of the preprocessed monitoring data using a diversion transmission strategy, the diversion transmission strategy includes a primary transmission strategy, a secondary transmission strategy, and a tertiary transmission strategy, wherein: The primary transmission strategy is to generate alarm messages containing time, location, type, and confidence level summaries for the primary data, and then upload the alarm messages to the central platform in real time through a high-priority channel. The secondary transmission strategy involves extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain secondary compressed feature data, and uploading the secondary compressed feature data to the central platform at a frequency of once per hour. The three-level transmission strategy involves randomly selecting a preset number B of three-level data samples from the three-level data. Data that is not randomly selected is considered local data samples. The three-level data samples are uploaded to the central platform once a day, and the local data samples are stored in the local storage buffer for a preset period C.

[0009] Furthermore, in step S5, when optimizing the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method, the traffic splitting strategy is optimized using the traffic splitting optimization method, wherein the traffic splitting optimization method includes: Step D01: Obtain end-to-end transmission delay, packet loss rate, uplink bandwidth utilization, and transmission jitter. Step D02: Calculate the real-time network congestion level value E based on end-to-end transmission delay E1, packet loss rate E2, uplink bandwidth utilization E3, and transmission jitter E4. Set E = w1×1 + w2×E2 + w3×E3 + w4×E4, and the range of the real-time network congestion level value E is [0,1]. Here, w1 is the end-to-end transmission delay weight coefficient, w2 is the packet loss rate weight coefficient, w3 is the uplink bandwidth utilization weight coefficient, and w4 is the transmission jitter weight coefficient, and w1 + w2 + w3 + w4 = 1.

[0010] Furthermore, in step S5, when optimizing the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method, the traffic splitting strategy is optimized using the traffic splitting optimization method, which further includes: Step D03: Compare the real-time network congestion level value E with the preset network congestion level value E0. Based on the comparison result, determine the network congestion situation and optimize the traffic distribution accordingly. The preset network congestion level value E0 includes mild congestion value Eq, moderate congestion value Ez, and severe congestion value Ed, with Eq=0.3, Ez=0.6, and Ed=0.8; When E < Eq, the network congestion is considered to be smooth, and no optimization is performed on the offloading transmission; When Eq≤E<Ez, the network congestion is determined to be mild. The offloading transmission is then optimized by extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain compressed secondary feature data, and uploading this compressed secondary feature data to the central platform every two hours. Simultaneously, the tertiary transmission strategy is optimized by pausing the upload of tertiary data and storing it in a local storage buffer for a preset period C. When Ez≤E<Ed, the network congestion is determined to be moderate. The offloading transmission is then optimized by extracting time-frequency domain features from the second-level data, compressing and encapsulating these features to obtain second-level compressed feature data, and uploading this compressed feature data to the central platform every four hours. Simultaneously, the third-level transmission strategy is optimized by pausing the uploading of third-level data and storing it in a local storage buffer for a second preset period C2. When Ed≤E, the network congestion is determined to be severe. The offloading transmission is then optimized by caching the secondary transmission strategy in the offloading transmission in a local storage buffer. When the network congestion is moderate, time-frequency domain features are extracted from the secondary data and compressed and encapsulated to obtain secondary compressed feature data. This compressed feature data is then uploaded to the central platform every four hours. At the same time, the tertiary transmission strategy is optimized by minimizing the caching of tertiary data in the local storage buffer and overwriting historical tertiary data.

[0011] Further, in step S6, when adjusting the optimization process of the traffic splitting strategy based on the network congestion changes after optimization, the optimized network congestion level value Eyh is obtained through steps D01-D02 in step S5. The network congestion change value ΔE is calculated based on Eyh and E, and ΔE = Eyh - E is set. The network congestion change value ΔE is compared with a preset effective network congestion change value ΔE0. Based on the comparison result, the network congestion change after optimization is judged, and the optimization process of the traffic splitting strategy is adjusted based on the judgment result. Wherein: ΔE0=0.3; When ΔE≥ΔE0, the network congestion change after the offloading optimization is considered to be an effective mitigation of network congestion, and the optimization process of the offloading strategy is not adjusted. When ΔE < ΔE0, the network congestion change after the offloading and transmission optimization is determined to be an ineffective mitigation of network congestion. The optimization process of the offloading and transmission strategy is then adjusted. The adjustment method is to change the network congestion situation when Eq ≤ E < Ez and Ez ≤ E < Ed to severe congestion. The method of optimizing offloading and transmission for severe congestion is then adjusted to stop sending attempts of the secondary and tertiary transmission strategies until the network congestion change is an effective mitigation of network congestion.

[0012] Furthermore, in step S7, when the preprocessed monitoring data transmitted to the central platform is aggregated and analyzed, alarm messages are stored in the real-time database and a first-level early warning strategy is triggered. Second-level compressed feature data is stored in the time series database, and the second-level data is aggregated and analyzed to generate a health assessment report. The health assessment report is then transmitted to the administrator port, and third-level data samples are stored in the object storage database. The anomaly confidence model is then corrected based on the third-level data samples.

[0013] Further, in step S7, when correcting the anomaly confidence model based on the three-level data samples, the three-level data samples are input into the trained teacher model. The trained teacher model outputs the anomaly confidence verification value Zjy corresponding to the three-level data samples. The anomaly confidence verification value Zjy is compared with the second preset reliability score Zmin in the preset reliability score Z0. Based on the comparison result, the performance of the anomaly confidence model is judged to meet the standard, and the anomaly confidence model is corrected based on the judgment result. When Zjy≤Zmin, the performance of the anomaly confidence model is deemed to be up to standard, and no correction is made to the anomaly confidence model. When Zjy > Zmin, the performance of the anomaly confidence model is deemed unsatisfactory, and the anomaly confidence model is corrected.

[0014] Furthermore, when correcting the abnormal confidence model in step S7, step S3 is executed on the central platform to obtain the corrected abnormal confidence model. The corrected abnormal confidence model parameters are then transmitted to the intelligent edge computing gateway through a high-priority channel to correct the abnormal confidence model parameters in the intelligent edge computing gateway.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: the method acquires monitoring data in real time through step S1, laying the foundation for subsequent multimodal data fusion analysis. It also solves the problems of traditional monitoring data being singular, isolated, and having inconsistent time stamps, thus ensuring data completeness and reliability. Through waveletting, denoising, and normalization processing in step S2, the method effectively suppresses on-site interference, eliminates dimensional differences, and generates a standardized data stream that can be directly processed by the model. Simultaneously, the local circular buffer provides short-term data backtracking capability, enabling data preservation and post-analysis during network interruptions, thus enhancing the system. The method enhances the robustness of data at the edge, thereby improving data quality and standardizing specifications. Step S3, using knowledge distillation, quantization, and pruning techniques, compresses the large cloud model into a lightweight model suitable for edge deployment. This enables millisecond-level inference of multi-source data within resource-constrained gateways, generating anomaly confidence scores to improve the model's edge-based and real-time analysis capabilities. Step S4, based on confidence scores, implements a "value-first" transmission strategy, reducing network bandwidth consumption and cloud storage costs while minimizing the transmission latency of critical early warning information. The method monitors network performance in real time and quantifies congestion levels in step S5, thereby dynamically adjusting the transmission frequency, compression strategy, and local storage cycle of secondary and tertiary data to form a closed-loop feedback control mechanism. This allows the system to automatically find the optimal resource allocation scheme when facing network fluctuations, ensuring the absolute smoothness of the primary alarm channel and improving stability and resource utilization efficiency in complex industrial network environments, thus achieving adaptive environmental response capabilities. In step S6, the central platform aggregates global data for in-depth analysis, model retraining, and strategy optimization, and distributes the optimized model and strategy to the edge, thereby continuously evolving the model of the edge nodes to adapt to new fault modes and environmental changes, avoiding the problem of model performance decay over time, and ensuring the accuracy and advancement of long-term operation. In step S7, primary data triggers real-time early warning to ensure safety. Secondary data is aggregated and analyzed to generate health assessments for condition-based maintenance. Tertiary data samples are used for "post-audit" and performance verification of the edge model, thereby driving model iterative updates. This transforms massive amounts of low-value data into valuable resources for continuous system self-optimization and self-verification, further improving reliability and credibility. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring in this embodiment. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Please see Figure 1 The diagram shown is a flowchart illustrating the edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring in this example. The method includes: Step S1: Real-time acquisition of monitoring data to obtain raw monitoring data; Step S2: Preprocess the raw monitoring data to obtain preprocessed monitoring data; Step S3: Construct an anomaly confidence model; Step S4: Generate anomaly confidence scores for the preprocessed monitoring data based on the anomaly confidence model, and transmit the preprocessed monitoring data to the central platform according to the anomaly confidence scores of the preprocessed monitoring data using a diversion transmission strategy. Step S5: Optimize the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method; Step S6: Adjust the optimization process of the off-transmission strategy according to the changes in network congestion after the off-transmission strategy optimization; Step S7: Perform aggregated analysis on the preprocessed monitoring data transmitted to the central platform and correct the anomaly confidence model.

[0022] Specifically, the method is applied to the overhead contact system of electrified railways. The method forms a high-quality data source through steps S1 and S2, outputs anomaly confidence scores through step S3, implements "value-first" transmission through step S4, dynamically optimizes the transmission strategy based on network conditions in step S5 to ensure system self-adaptation, and drives continuous evolution through steps S6 and S7. This achieves a paradigm shift from "full data transmission" to "value transmission" and constructs an intelligent monitoring process with self-sensing and self-optimizing capabilities, thereby improving the economy, real-time performance, and reliability of overhead contact system monitoring. In particular, the method collects monitoring data in real time through step S1, laying the foundation for subsequent multimodal data fusion analysis, while simultaneously solving the problems of traditional... To address the issues of isolated, fragmented, and inconsistent time-stamped monitoring data, ensuring data completeness and reliability, the method employs steps S2 (wavelengthening, denoising, and normalization) to effectively suppress on-site interference, eliminate dimensional differences, and generate a standardized data stream directly processable by the model. Simultaneously, a local circular buffer provides short-term data backtracking capabilities, enabling data preservation and post-event analysis during network outages, enhancing system robustness. This allows for data quality improvement and standardization at the edge. Step S3, knowledge distillation, quantization, and pruning techniques, compresses the large cloud-based model into a lightweight model suitable for edge deployment, enabling millisecond-level inference on multi-source data within resource-constrained gateways and generating anomaly confidence scores to improve model performance. The method, with its edge-based and real-time analysis capabilities, implements a "value-first" transmission strategy through a three-level traffic splitting based on confidence scores in step S4. This reduces network bandwidth consumption and cloud storage costs while minimizing the transmission latency of critical early warning information. Step S5 involves real-time monitoring of network performance and quantification of congestion levels, dynamically adjusting the transmission frequency, compression strategy, and local storage cycle of secondary and tertiary data to form a closed-loop feedback control mechanism. This allows the method to automatically find the optimal resource allocation solution when facing network fluctuations, ensuring the absolute unobstructed flow of the primary alarm channel and improving stability and resource utilization efficiency in complex industrial network environments, thereby achieving adaptive environmental response capabilities. The S6 central platform aggregates global data for in-depth analysis, model retraining, and strategy optimization. The optimized models and strategies are then distributed to the edge, enabling continuous evolution of edge node models to adapt to new failure modes and environmental changes. This avoids performance degradation over time and ensures long-term accuracy and advanced capabilities. The method involves triggering real-time early warnings through first-level data in step S7 to ensure safety; generating health assessments from aggregated second-level data for condition-based maintenance; and using third-level data samples for "post-audit" and performance verification of the edge models, driving iterative model updates. This transforms massive amounts of low-value data into valuable resources for continuous system self-optimization and self-verification, further enhancing reliability and trustworthiness.

[0023] Specifically, in step S1, when the monitoring data is collected in real time, the monitoring data is collected in real time through an intelligent edge computing gateway to obtain raw monitoring data. The raw monitoring data includes triaxial raw acceleration time-series signals, with units of m / s². 2 The original measurements of lateral displacement (mm), biaxial tilt angle (°), tension (kN), visible light image and video (RGB pixel matrix format), thermal infrared image (temperature pixel matrix format), ambient temperature (°C), ambient humidity (%RH), wind speed (m / s), wind direction (°C), sensor battery voltage (V), sensor internal temperature (°C), and signal strength (dBm).

[0024] Specifically, the intelligent edge computing gateway refers to an industrial IoT gateway device deployed at the contact wire monitoring site to realize sensor data access, edge computing, and hierarchical transmission control. The triaxial raw acceleration time-series signal refers to unprocessed continuous acceleration change waveform data acquired by a triaxial MEMS along the three orthogonal directions of the spatial rectangular coordinate system X, Y, and Z. The raw lateral displacement measurement value refers to the instantaneous offset distance of the contact wire relative to the center reference position, measured perpendicular to the contact wire direction by a laser displacement sensor or microwave radar sensor. The raw dual-axis tilt angle measurement value refers to the instantaneous tilt angle of the contact wire relative to the horizontal reference in the direction along and perpendicular to the line, measured by a dual-axis tilt angle sensor. The raw tension measurement signal refers to the instantaneous tension value borne by the contact wire and the catenary cable, directly measured by a strain gauge tension sensor. The raw visible light image and video frame refers to uncompressed static images acquired by an industrial high-definition visible light camera. The data consists of a single frame of a dynamic image or video sequence. The original frame of the thermal infrared image refers to a single frame of image data that reflects the surface temperature distribution of a target object, acquired by an uncooled infrared thermal imager. The temperature pixel matrix refers to a two-dimensional array in which the value of each element corresponds to the measured temperature value of the corresponding spatial location in the thermal infrared image. The original ambient temperature value refers to the instantaneous value of the atmospheric temperature measured near the monitoring point by a digital temperature sensor. The original ambient humidity value refers to the instantaneous value of the atmospheric relative humidity measured near the monitoring point by a capacitive or resistive humidity sensor. The original wind speed value refers to the instantaneous value of the airflow speed measured near the monitoring point by an ultrasonic or mechanical anemometer. The original wind direction value refers to the instantaneous value of the airflow direction measured near the monitoring point by a wind vane and an ultrasonic anemometer. The sensor battery voltage refers to the instantaneous value of the current terminal voltage of the battery that powers the wireless sensor node. The sensor internal temperature refers to the temperature measurement value of the sensor's own circuit board.

[0025] Specifically, step S1 involves real-time acquisition of monitoring data to lay the foundation for subsequent multimodal data fusion analysis, while addressing the problems of traditional monitoring data being singular, isolated, and having inconsistent time stamps, thereby ensuring the completeness and reliability of the data.

[0026] Specifically, when the original monitoring data is preprocessed in step S2, the preprocessing includes filtering, denoising, normalization, and temporary storage. The temporary storage process involves temporarily storing the preprocessed monitoring data obtained after normalization in a local storage buffer.

[0027] Specifically, the filtering process refers to selecting and applying appropriate digital filters to remove and suppress specific frequency components in the signal based on the sensor data type and subsequent analysis requirements. The denoising process refers to using signal processing algorithms on the basis of filtering to separate and suppress random noise and sudden interference mixed in the effective signal. The normalization process refers to mapping raw data collected by different sensors with different physical meanings and dimensions to a unified, dimensionless standard numerical range through mathematical transformation. The local storage buffer refers to a non-volatile storage space configured within the intelligent edge computing gateway using a cyclic overlay strategy. This embodiment does not limit the temporary storage time of the preprocessed monitoring data obtained after normalization in the local storage buffer. Those skilled in the art can set it freely according to the actual situation, as long as the requirement of temporarily storing the preprocessed monitoring data obtained after normalization in the local storage buffer is met. For example, the preprocessed monitoring data obtained after normalization can be temporarily stored in the local storage buffer. The temporary storage time in the local storage buffer is set to 7 days. This embodiment does not limit the specific method of filtering; those skilled in the art can freely set it according to the actual situation, as long as it meets the requirement of preprocessing the original monitoring data. For example, the filtering process can be set to use a digital filter to remove high-frequency noise from the original monitoring data to obtain filtered data. This embodiment does not limit the specific method of denoising; those skilled in the art can freely set it according to the actual situation, as long as it meets the requirement of preprocessing the original monitoring data. For example, the denoising process can be set to use a wavelet threshold denoising algorithm to retain the effective signal features in the filtered data to obtain denoised data. This embodiment does not limit the specific method of normalization; those skilled in the art can freely set it according to the actual situation, as long as it meets the requirement of preprocessing the original monitoring data. For example, the normalization process can be set to use a maximum-minimum value normalization method to scale the denoised data to the [0,1] range to obtain preprocessed monitoring data. Specifically, step S2 effectively suppresses on-site interference and eliminates dimensional differences through filtering, denoising, and normalization, generating a standardized data stream that can be directly processed by the model. At the same time, the local circular buffer provides short-term data backtracking capability, making it possible to preserve data and perform post-event analysis when the network is interrupted, thereby enhancing the robustness of the system and completing the quality improvement and standardization of data at the edge.

[0028] Specifically, in step S3, when constructing the anomaly confidence model, the anomaly confidence model is constructed using an anomaly confidence model construction method, wherein the anomaly confidence model construction method includes: Step A01: Obtain model construction simulation data, which includes normal operating condition data and labeled abnormal data; Step A02, structuring the teacher model, including: The visual branch of the teacher model is set to use a deep residual network to process visible light and infrared images to extract appearance and thermal features; The time-series signal branch of the teacher model is set to use a one-dimensional deep convolutional network to process time-series data and extract time-frequency features; The feature fusion and decision layer of the teacher model is set to fuse the appearance and thermal features and time-frequency features extracted from the visual branch and the temporal signal branch of the teacher model, and the classification output layer outputs the probability distribution vector representing the possibility of various anomalies. Step A03: Transmit the model construction simulation data to the teacher model for supervised learning to obtain the trained teacher model; Step A04: Use knowledge distillation to transfer the parameters of the trained teacher model to the student model; Step A05: Perform low-bit quantization on the student model, converting the student model weight activation values ​​from 32-bit floating-point numbers to 8-bit fixed-point numbers to obtain the quantized student model. Step A06 involves using structured pruning to optimize the quantized model, obtaining an anomaly confidence model, and deploying it in the intelligent edge computing gateway.

[0029] Specifically, the model construction simulation data refers to a multimodal data set covering various operating states of the overhead contact system, used for training and optimizing the anomaly confidence model. This embodiment does not limit the method of acquiring the model construction simulation data; those skilled in the art can freely set it according to actual conditions, as long as the need for acquiring the model construction simulation data is met. For example, it can be set to acquire the model construction simulation data through big data. The normal operating condition data refers to monitoring data collected when the overhead contact system and its supporting structures are in a healthy and stable operating state. The labeled abnormal data refers to monitoring data segments whose fault type, location, and occurrence time have been clearly identified by domain experts or through historical fault records. The visual branch of the teacher model is... The term "deep residual network" refers to the component responsible for processing image modal inputs. The "deep residual network" refers to a deep convolutional neural network architecture containing residual connections. The "visible light and infrared images" refer to two-dimensional data acquired by corresponding imaging sensors deployed on-site. The "appearance and thermal features" refer to high-dimensional vectors extracted from visible light images by the deep residual network, representing visual attributes such as target shape, contour, and texture, and high-dimensional vectors extracted from infrared images, representing temperature distribution patterns and thermal anomaly patterns. The "temporal signal branch of the teacher model" refers to the component responsible for processing time-series modal inputs. The "one-dimensional deep convolutional network" refers to a deep neural network where the convolutional kernel slides only along the time dimension. The "time-series data" refers to data acquired in chronological order. The signal sequence is a collection of signals with inherent sequential correlation. The time-frequency feature refers to the abstract feature representation learned and output by a one-dimensional deep convolutional network from time-series data, which can simultaneously reflect the temporal evolution law and frequency component characteristics of the signal. The feature fusion and decision layer of the teacher model refers to the network layer that receives and integrates multimodal features extracted from the visual branch and the time-series signal branch, and finally outputs the classification decision. This embodiment does not limit the method of fusing appearance and thermal features and time-frequency features. Those skilled in the art can set it freely according to the actual situation, as long as the requirement of fusing appearance and thermal features and time-frequency features is met. For example, it can be set to use feature concatenation to connect feature vectors of different modalities into a joint feature vector. The output layer refers to the Softmax activation function layer following the fully connected layer. The probability distribution vector representing the likelihood of various anomalies is a multi-dimensional vector where each element represents the predicted probability of the input data belonging to the corresponding category, and each element's value is between 0 and 1, with the sum of all elements being 1. Supervised learning refers to a machine learning paradigm that uses labeled datasets to train the model. Knowledge distillation refers to model compression techniques. Low-bit quantization refers to model compression and acceleration techniques. The student model weight activation values ​​refer to the trainable parameters in the student neural network and the intermediate results generated by each layer during forward propagation. 32-bit floating-point numbers refer to the default high-precision numerical representation for deep learning training and inference.The 8-bit fixed-point number refers to a low-precision numerical representation that linearly maps a numerical range to 256 discrete integers; the structured pruning refers to a model compression method.

[0030] Specifically, step S3 uses knowledge distillation, quantization, and pruning techniques to compress the large cloud model into a lightweight model suitable for edge deployment, enabling it to perform millisecond-level inference on multi-source data within resource-constrained gateways and generate anomaly confidence scores, thereby improving the model's edge and real-time analysis capabilities.

[0031] Specifically, in step S4, when generating the anomaly confidence score of the preprocessed monitoring data according to the anomaly confidence model, the preprocessed monitoring data is input into the anomaly confidence model, and the anomaly confidence model outputs the anomaly confidence score Z. In step S4, when transmitting the preprocessed monitoring data to the central platform according to the anomaly confidence score of the preprocessed monitoring data and the diversion transmission strategy, the anomaly confidence score Z of the transmitted preprocessed monitoring data is compared with the preset confidence score Z0. The priority of the preprocessed monitoring data is determined based on the comparison result, and the preprocessed monitoring data is output based on the determination result. Wherein: The preset reliability score Z0 includes a first preset reliability score Zmax and a second preset reliability score Zmin, where Zmax = 0.8 and Zmin = 0.3. When Zmax < Z, the priority of the preprocessed monitoring data is determined to be level one, and the preprocessed monitoring data is output as level one data. When Zmin<Z≤Zmax, the priority of the preprocessed monitoring data is determined to be level two, and the preprocessed monitoring data is output as level two data; When Z≤Zmin, the priority of the preprocessed monitoring data is determined to be level three, and the preprocessed monitoring data is output as level three data; In step S4, when the preprocessed monitoring data is transmitted to the central platform according to the anomaly confidence score, the transmission strategy includes a primary transmission strategy, a secondary transmission strategy, and a tertiary transmission strategy, wherein: The primary transmission strategy is to generate alarm messages containing time, location, type, and confidence level summaries for the primary data, and then upload the alarm messages to the central platform in real time through a high-priority channel. The secondary transmission strategy involves extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain secondary compressed feature data, and uploading the secondary compressed feature data to the central platform at a frequency of once per hour. The three-level transmission strategy involves randomly selecting a preset number B of three-level data samples from the three-level data. Data that is not randomly selected is considered local data samples. The three-level data samples are uploaded to the central platform once a day, and the local data samples are stored in the local storage buffer for a preset period C.

[0032] Specifically, the anomaly confidence score Z refers to the probability that the preprocessed monitoring data exhibits abnormal characteristics within the given time period. The preset confidence score Z0 refers to a grading threshold pre-set for comparison with the anomaly confidence score Z, used to achieve adaptive data grading. The first preset confidence score Zmax refers to the high confidence threshold in the grading thresholds, and the second preset confidence score Zmin refers to the low confidence threshold in the grading thresholds. This embodiment does not limit the method of generating alarm messages. For example, it can be set to automatically fill and serialize the timestamp, geographic coordinates, anomaly type determined by the model, and anomaly confidence score Z corresponding to the first-level data using a predefined Protobuf data template to generate a structured alarm message. This embodiment does not limit the method of extracting time-frequency domain features from the second-level data. For example, it can be set to calculate the spectrum of the vibration signal using fast Fourier transform, extract the frequency and amplitude of the first N main spectral peaks as frequency domain features, and simultaneously calculate... The time-domain statistical features of the computational signal together constitute the time-frequency domain features. This embodiment does not limit the method of compressing and encapsulating the time-frequency domain features. For example, it can be set to first use principal component analysis to reduce the dimensionality of the high-dimensional feature vector, then use a high-speed lossless compression algorithm to compress the dimensionality-reduced data, and finally add a data header to encapsulate it into a data packet. The preset quantity B refers to the number of data sample points or data fragments randomly selected when sampling and uploading the third-level data every day. For example, the preset quantity B is one percent of the total amount of third-level data generated on that day. The high-priority channel refers to the physical communication channel with deterministic service quality guarantee established between the edge gateway and the central platform, such as 5G network and industrial Ethernet. The central platform refers to a cloud server cluster deployed in a remote data center with storage and computing capabilities. The preset period C refers to the longest time that the local data samples in the third-level data are retained in the local storage buffer of the intelligent edge computing gateway. For example, the preset period C is seven days.

[0033] Specifically, step S4 implements a "value-first" transmission strategy through a three-level traffic splitting based on confidence scores, thereby reducing network bandwidth usage and cloud storage costs while ensuring that the transmission latency of critical early warning information is minimized.

[0034] Specifically, when optimizing the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method in step S5, the traffic splitting strategy is optimized using the traffic splitting optimization method, which includes: Step D01: Obtain end-to-end transmission delay, packet loss rate, uplink bandwidth utilization, and transmission jitter. Step D02: Calculate the real-time network congestion level value E based on end-to-end transmission delay E1, packet loss rate E2, uplink bandwidth utilization E3, and transmission jitter E4. Set E = w1×1 + w2×E2 + w3×E3 + w4×E4, and the range of the real-time network congestion level value E is [0,1]. Where w1 is the end-to-end transmission delay weight coefficient, w2 is the packet loss rate weight coefficient, w3 is the uplink bandwidth utilization weight coefficient, w4 is the transmission jitter weight coefficient, and w1 + w2 + w3 + w4 = 1. Step D03: Compare the real-time network congestion level value E with the preset network congestion level value E0. Based on the comparison result, determine the network congestion situation and optimize the traffic distribution accordingly. The preset network congestion level value E0 includes mild congestion value Eq, moderate congestion value Ez, and severe congestion value Ed, with Eq=0.3, Ez=0.6, and Ed=0.8; When E < Eq, the network congestion is considered to be smooth, and no optimization is performed on the offloading transmission; When Eq≤E<Ez, the network congestion is determined to be mild. The offloading transmission is then optimized by extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain compressed secondary feature data, and uploading this compressed secondary feature data to the central platform every two hours. Simultaneously, the tertiary transmission strategy is optimized by pausing the upload of tertiary data and storing it in a local storage buffer for a preset period C. When Ez≤E<Ed, the network congestion is determined to be moderate. The offloading transmission is then optimized by extracting time-frequency domain features from the second-level data, compressing and encapsulating these features to obtain second-level compressed feature data, and uploading this compressed feature data to the central platform every four hours. Simultaneously, the third-level transmission strategy is optimized by pausing the uploading of third-level data and storing it in a local storage buffer for a second preset period C2. When Ed≤E, the network congestion is determined to be severe. The offloading transmission is then optimized by caching the secondary transmission strategy in the offloading transmission in a local storage buffer. When the network congestion is moderate, time-frequency domain features are extracted from the secondary data and compressed and encapsulated to obtain secondary compressed feature data. This compressed feature data is then uploaded to the central platform every four hours. At the same time, the tertiary transmission strategy is optimized by minimizing the caching of tertiary data in the local storage buffer and overwriting historical tertiary data.

[0035] Specifically, the end-to-end transmission delay refers to the total time difference from when a data packet is sent from the application layer of the intelligent edge computing gateway to when it is successfully received and acknowledged by the application layer of the central platform. The packet loss rate refers to the proportion of data packets that are not successfully received by the central platform out of the total number of data packets sent by the intelligent edge computing gateway. The uplink bandwidth utilization rate refers to the percentage of the actual uplink data throughput of the intelligent edge computing gateway within a specific time window relative to its theoretical maximum available uplink bandwidth. The transmission jitter refers to the degree of fluctuation in the end-to-end transmission delay value of a series of continuously sent data packets. This embodiment does not limit the acquisition method. For example, it can be set to periodically acquire the end-to-end transmission delay, packet loss rate, uplink bandwidth utilization rate, and transmission jitter through network protocol stack statistics, actively sending probe messages, and using the acknowledgment mechanism of the transport layer protocol. The end-to-end transmission delay weighting coefficient is a parameter used to assign importance to the end-to-end transmission delay performance index in the weighted calculation formula used to construct the real-time network congestion level value E. The packet loss rate weighting coefficient is a parameter used to construct the real-time network congestion level value E. In the weighted calculation formula for the real-time network congestion level value E, the parameter assigning importance to the network performance indicator of packet loss rate is [parameter name missing]. The weight coefficient for cross-link bandwidth utilization is a parameter used in the weighted calculation formula for the real-time network congestion level value E to assign importance to the network performance indicator of cross-link bandwidth utilization. The weight coefficient for transmission jitter is a parameter used in the weighted calculation formula for the real-time network congestion level value E to assign importance to the network performance indicator of transmission jitter. This embodiment does not include the weight coefficients for end-to-end transmission delay, packet loss rate, and cross-link bandwidth. The specific values ​​of the bandwidth utilization weight coefficient and the transmission jitter weight coefficient are limited. For example, w1=w2=w3=w4=0.25 can be set. The second preset period C2 refers to the storage time set for the local storage buffer of the third-level data when the network is moderately congested. It is shorter than the normal preset period C. For example, the second preset period C2 is 3 days. The third-level data minimization cache in the local storage buffer refers to the storage strategy adopted for the third-level data when the network is heavily congested. The historical third-level data refers to the third-level data that has been stored in the local buffer before the current moment.

[0036] Specifically, step S5 dynamically adjusts the transmission frequency, compression strategy, and local storage cycle of secondary and tertiary data by monitoring network performance in real time and quantifying the degree of congestion. This forms a closed-loop feedback control mechanism, which can automatically find the optimal resource allocation scheme when facing network fluctuations, ensure the absolute smoothness of the primary alarm channel, and improve the stability and resource utilization efficiency in complex industrial network environments, thereby achieving adaptive environmental response capabilities.

[0037] Specifically, in step S6, when adjusting the optimization process of the traffic splitting strategy based on the network congestion changes after optimization, the optimized network congestion level value Eyh is obtained through steps D01-D02 in step S5. The network congestion change value ΔE is calculated based on Eyh and E, and ΔE = Eyh - E is set. The network congestion change value ΔE is compared with a preset effective network congestion change value ΔE0. Based on the comparison result, the network congestion change after optimization is judged, and the optimization process of the traffic splitting strategy is adjusted based on the judgment result. ΔE0=0.3; When ΔE≥ΔE0, the network congestion change after the offloading optimization is considered to be an effective mitigation of network congestion, and the optimization process of the offloading strategy is not adjusted. When ΔE < ΔE0, the network congestion change after the offloading and transmission optimization is determined to be an ineffective mitigation of network congestion. The optimization process of the offloading and transmission strategy is then adjusted. The adjustment method is to change the network congestion situation when Eq ≤ E < Ez and Ez ≤ E < Ed to severe congestion. The method of optimizing offloading and transmission for severe congestion is then adjusted to stop sending attempts of the secondary and tertiary transmission strategies until the network congestion change is an effective mitigation of network congestion.

[0038] Specifically, the preset effective network congestion change value ΔE0 refers to an empirical threshold used to determine whether the transmission optimization strategy executed in step S5 has a significant effect.

[0039] Specifically, step S6 involves aggregating global data through a central platform for in-depth analysis, model retraining, and strategy optimization. The optimized model and strategy are then distributed to the edge, enabling continuous evolution of the edge node model to adapt to new failure modes and environmental changes. This avoids the problem of model performance decaying over time and ensures long-term accuracy and advanced performance.

[0040] Specifically, in step S7, when the preprocessed monitoring data transmitted to the central platform is aggregated and analyzed, alarm messages are stored in the real-time database and a first-level early warning strategy is triggered. Second-level compressed feature data is stored in the time series database, and the second-level data is aggregated and analyzed to generate a health assessment report. The health assessment report is then transmitted to the administrator port, and third-level data samples are stored in the object storage database. The anomaly confidence model is then corrected based on the third-level data samples. In step S7, when correcting the anomaly confidence model based on the three-level data samples, the three-level data samples are input into the trained teacher model. The trained teacher model outputs the anomaly confidence verification value Zjy corresponding to the three-level data samples. The anomaly confidence verification value Zjy is compared with the second preset reliability score Zmin in the preset reliability score Z0. Based on the comparison result, the performance of the anomaly confidence model is judged to meet the standard, and the anomaly confidence model is corrected based on the judgment result. When Zjy≤Zmin, the performance of the anomaly confidence model is deemed to be up to standard, and no correction is made to the anomaly confidence model. When Zjy > Zmin, the performance of the anomaly confidence model is deemed unsatisfactory, and the anomaly confidence model is corrected. When the abnormal confidence model is corrected in step S7, step S3 is executed on the central platform to obtain the corrected abnormal confidence model. The corrected abnormal confidence model parameters are then transmitted to the intelligent edge computing gateway through a high-priority channel to correct the abnormal confidence model parameters in the intelligent edge computing gateway.

[0041] Specifically, the real-time database refers to a database system optimized for high-speed writing and massive real-time querying; the primary early warning strategy refers to the emergency handling process automatically triggered by the central platform after receiving an alarm message, such as issuing audible and visual alarms on the operation and maintenance monitoring screen and automatically generating work orders to be dispatched to relevant responsible persons; the time-series database refers to a database optimized for storing and querying time-series data; the aggregation analysis refers to the process of statistically calculating the secondary feature data stored in the time-series database according to equipment, time range, and anomaly type, and generating a health assessment report reflecting the health trend of the contact network line based on this; the administrator port refers to the human-computer interaction interface for system operation and maintenance personnel, such as a mobile APP; and the object storage database refers to a cloud storage service used to store unstructured data.

[0042] Specifically, step S7 triggers real-time early warnings through first-level data to ensure safety; second-level data is aggregated and analyzed to generate health assessments to serve condition-based maintenance; and third-level data samples are used to perform "post-audit" and performance verification on the edge model, thereby driving iterative updates of the model. This transforms massive amounts of low-value data into valuable resources for continuous self-optimization and self-verification of the system, thereby further improving reliability and credibility.

[0043] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for edge computing and adaptive hierarchical data transmission for overhead contact line monitoring, characterized in that, The method includes: Step S1: Real-time acquisition of monitoring data to obtain raw monitoring data; Step S2: Preprocess the raw monitoring data to obtain preprocessed monitoring data; Step S3: Construct an anomaly confidence model; Step S4: Generate anomaly confidence scores for the preprocessed monitoring data based on the anomaly confidence model, and transmit the preprocessed monitoring data to the central platform according to the anomaly confidence scores of the preprocessed monitoring data using a diversion transmission strategy. Step S5: Optimize the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method; Step S6: Adjust the optimization process of the off-transmission strategy according to the changes in network congestion after the off-transmission strategy optimization; Step S7: Perform aggregated analysis on the preprocessed monitoring data transmitted to the central platform and correct the anomaly confidence model.

2. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 1, characterized in that, In step S3, when constructing the anomaly confidence model, the anomaly confidence model is constructed using an anomaly confidence model construction method, wherein the anomaly confidence model construction method includes: Step A01: Obtain model construction simulation data, which includes normal operating condition data and labeled abnormal data; Step A02, structuring the teacher model, including: The visual branch of the teacher model is set to use a deep residual network to process visible light and infrared images to extract appearance and thermal features; The time-series signal branch of the teacher model is set to use a one-dimensional deep convolutional network to process time-series data and extract time-frequency features; The feature fusion and decision layer of the teacher model is set to fuse the appearance and thermal features and time-frequency features extracted from the visual branch and the temporal signal branch of the teacher model, and the classification output layer outputs the probability distribution vector representing the possibility of various anomalies. Step A03: Transmit the model construction simulation data to the teacher model for supervised learning to obtain the trained teacher model; Step A04: Use knowledge distillation to transfer the parameters of the trained teacher model to the student model; Step A05: Perform low-bit quantization on the student model, converting the student model weight activation values ​​from 32-bit floating-point numbers to 8-bit fixed-point numbers to obtain the quantized student model. Step A06 involves using structured pruning to optimize the quantized model, obtaining an anomaly confidence model, and deploying it in the intelligent edge computing gateway.

3. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 1, characterized in that, In step S4, when generating the anomaly confidence score of the preprocessed monitoring data according to the anomaly confidence model, the preprocessed monitoring data is input into the anomaly confidence model, and the anomaly confidence model outputs the anomaly confidence score Z. In step S4, when transmitting the preprocessed monitoring data to the central platform according to the anomaly confidence score based on the preprocessed monitoring data, the anomaly confidence score Z of the preprocessed monitoring data is compared with the preset confidence score Z0. The priority of the preprocessed monitoring data is determined based on the comparison result, and the preprocessed monitoring data is output based on the determination result. Wherein: The preset reliability score Z0 includes a first preset reliability score Zmax and a second preset reliability score Zmin, where Zmax = 0.8 and Zmin = 0.

3. When Zmax < Z, the priority of the preprocessed monitoring data is determined to be level one, and the preprocessed monitoring data is output as level one data. When Zmin<Z≤Zmax, the priority of the preprocessed monitoring data is determined to be level two, and the preprocessed monitoring data is output as level two data; When Z≤Zmin, the priority of the preprocessed monitoring data is determined to be level three, and the preprocessed monitoring data is output as level three data.

4. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 3, characterized in that, In step S4, when the preprocessed monitoring data is transmitted to the central platform according to the anomaly confidence score, the transmission strategy includes a primary transmission strategy, a secondary transmission strategy, and a tertiary transmission strategy, wherein: The primary transmission strategy is to generate alarm messages containing time, location, type, and confidence level summaries for the primary data, and then upload the alarm messages to the central platform in real time through a high-priority channel. The secondary transmission strategy involves extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain secondary compressed feature data, and uploading the secondary compressed feature data to the central platform at a frequency of once per hour. The three-level transmission strategy involves randomly selecting a preset number B of three-level data samples from the three-level data. Data that is not randomly selected is considered local data samples. The three-level data samples are uploaded to the central platform once a day, and the local data samples are stored in the local storage buffer for a preset period C.

5. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 1, characterized in that, When optimizing the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method in step S5, the traffic splitting strategy is optimized using the traffic splitting optimization method, which includes: Step D01: Obtain end-to-end transmission delay, packet loss rate, uplink bandwidth utilization, and transmission jitter. Step D02: Calculate the real-time network congestion level value E based on end-to-end transmission delay E1, packet loss rate E2, uplink bandwidth utilization E3, and transmission jitter E4. Set E = w1×1 + w2×E2 + w3×E3 + w4×E4, and the range of the real-time network congestion level value E is [0,1]. Here, w1 is the end-to-end transmission delay weight coefficient, w2 is the packet loss rate weight coefficient, w3 is the uplink bandwidth utilization weight coefficient, and w4 is the transmission jitter weight coefficient, and w1 + w2 + w3 + w4 = 1.

6. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 5, characterized in that, When optimizing the traffic splitting strategy based on the real-time network congestion level and the traffic splitting optimization method in step S5, the traffic splitting strategy is optimized using the traffic splitting optimization method, which further includes: Step D03: Compare the real-time network congestion level value E with the preset network congestion level value E0. Based on the comparison result, determine the network congestion situation and optimize the traffic distribution accordingly. The preset network congestion level value E0 includes mild congestion value Eq, moderate congestion value Ez, and severe congestion value Ed, with Eq=0.3, Ez=0.6, and Ed=0.8; When E < Eq, the network congestion is considered to be smooth, and no optimization is performed on the offloading transmission; When Eq≤E<Ez, the network congestion is determined to be mild. The offloading transmission is then optimized by extracting time-frequency domain features from the secondary data, compressing and encapsulating these features to obtain compressed secondary feature data, and uploading this compressed secondary feature data to the central platform every two hours. Simultaneously, the tertiary transmission strategy is optimized by pausing the upload of tertiary data and storing it in a local storage buffer for a preset period C. When Ez≤E<Ed, the network congestion is determined to be moderate. The offloading transmission is then optimized by extracting time-frequency domain features from the second-level data, compressing and encapsulating these features to obtain second-level compressed feature data, and uploading this compressed feature data to the central platform every four hours. Simultaneously, the third-level transmission strategy is optimized by pausing the uploading of third-level data and storing it in a local storage buffer for a second preset period C2. When Ed≤E, the network congestion is determined to be severe. The offloading transmission is then optimized by caching the secondary transmission strategy in the offloading transmission in a local storage buffer. When the network congestion is moderate, time-frequency domain features are extracted from the secondary data and compressed and encapsulated to obtain secondary compressed feature data. This compressed feature data is then uploaded to the central platform every four hours. At the same time, the tertiary transmission strategy is optimized by minimizing the caching of tertiary data in the local storage buffer and overwriting historical tertiary data.

7. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 5, characterized in that, When adjusting the optimization process of the traffic splitting strategy in step S6 based on the network congestion changes after optimization, the network congestion level value Eyh after traffic splitting optimization is obtained through steps D01-D02 in step S5. The network congestion change value ΔE is calculated based on the network congestion level value Eyh and the network congestion level value E, and ΔE = Eyh - E is set. The network congestion change value ΔE is compared with the preset effective network congestion change value ΔE0. Based on the comparison result, the network congestion change after traffic splitting optimization is judged, and the optimization process of the traffic splitting strategy is adjusted based on the judgment result. Wherein: ΔE0=0.3; When ΔE≥ΔE0, the network congestion change after the offloading optimization is considered to be an effective mitigation of network congestion, and the optimization process of the offloading strategy is not adjusted. When ΔE < ΔE0, the network congestion change after the offloading and transmission optimization is determined to be an ineffective mitigation of network congestion. The optimization process of the offloading and transmission strategy is then adjusted. The adjustment method is to change the network congestion situation when Eq ≤ E < Ez and Ez ≤ E < Ed to severe congestion. The method of optimizing offloading and transmission for severe congestion is then adjusted to stop sending attempts of the secondary and tertiary transmission strategies until the network congestion change is an effective mitigation of network congestion.

8. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 1, characterized in that, In step S7, when the preprocessed monitoring data transmitted to the central platform is aggregated and analyzed, alarm messages are stored in the real-time database and a first-level early warning strategy is triggered. Second-level compressed feature data is stored in the time series database, and aggregation analysis is performed on the second-level data to generate a health assessment report. The health assessment report is then transmitted to the administrator port, and third-level data samples are stored in the object storage database. The anomaly confidence model is then corrected based on the third-level data samples.

9. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 8, characterized in that, In step S7, when correcting the anomaly confidence model based on the three-level data samples, the three-level data samples are input into the trained teacher model. The trained teacher model outputs the anomaly confidence verification value Zjy corresponding to the three-level data samples. The anomaly confidence verification value Zjy is compared with the second preset reliability score Zmin in the preset reliability score Z0. Based on the comparison result, the performance of the anomaly confidence model is judged to meet the standard, and the anomaly confidence model is corrected based on the judgment result. When Zjy≤Zmin, the performance of the anomaly confidence model is deemed to be up to standard, and no correction is made to the anomaly confidence model. When Zjy > Zmin, the performance of the anomaly confidence model is deemed unsatisfactory, and the anomaly confidence model is corrected.

10. The edge computing and adaptive data hierarchical transmission method for overhead contact line monitoring according to claim 9, characterized in that, When the abnormal confidence model is corrected in step S7, step S3 is executed on the central platform to obtain the corrected abnormal confidence model. The corrected abnormal confidence model parameters are then transmitted to the intelligent edge computing gateway through a high-priority channel to correct the abnormal confidence model parameters in the intelligent edge computing gateway.

Images