Optical Cable Trunk Intelligent Inspection

Optical-cable trunk lines are critical infrastructure for the stable operation of carrier networks, spanning urban roads, rural areas, mountain roads, manholes, and pole routes. Traditional trunk inspection relies on manual patrol, vehicle inspection, and on-site photo records, which suffer from wide coverage, heavy workload, reliance on individual experience to spot problems, slow hazard response, and difficulty accumulating inspection data.

In real operations, trunk lines face many risks — third-party construction damage, missing or damaged marker stones, abnormal manhole covers, sagging aerial cables, leaning poles, construction machinery operating near the line, and collisions from over-height vehicles. If a hazard is not detected and handled in time, it can affect network stability or even trigger large-scale outages.

Trunk inspection is also highly on-site, dispersed, and real-time. Sending all images, video, and location data back to a central platform consumes bandwidth and slows recognition and on-site response. The project therefore introduced edge computing, deploying lightweight AI models on edge devices for fast on-site recognition, preliminary judgment, and result upload.

Around the telecom optical-cable trunk inspection scenario, Wuhan Luxijie Technology Co., Ltd. built an intelligent inspection system based on AI image recognition, edge computing, GIS mapping, inspection task management, and closed-loop hazard handling, helping the client upgrade from manual inspection to a digital operations model of "intelligent recognition, edge analysis, cloud management, and closed-loop handling."

Luxijie adopted an overall solution of "cloud-edge collaboration + AI vision + closed-loop inspection management," integrating front-end inspection devices, edge nodes, the central platform, and business systems into a complete solution covering collection, recognition, risk alerting, hazard handling, and data analysis.

1. Multi-source inspection data collection

The system ingests data from manual photos, vehicle-mounted inspection, drones, mobile terminals, and edge cameras, adapting to different terrains, lines, and working modes. On site, devices capture images, video, location, time, and records, associating data with the corresponding line, point, and task to form standardized inspection data assets.

2. Edge computing deployment

Given the large data volume, complex networks, and high response requirements of trunk inspection, the project deploys lightweight AI recognition models on edge devices or nodes. Edge devices perform preliminary recognition of images or video streams on site, quickly judging whether key targets or potential hazards exist, for example:

• whether a cable marker stone is detected;
• whether the marker is obscured, damaged, or missing;
• whether manholes, poles, or aerial cables are present;
• whether construction machinery or excavation equipment poses a third-party-damage risk;
• whether the abnormal result should be reported to the central platform.

Through local inference at the edge, the system avoids uploading large volumes of raw images and video, prioritizing recognition results, abnormal clips, risk screenshots, and structured data — reducing bandwidth pressure and improving on-site analysis efficiency.

3. Lightweight AI model deployment

The recognition models were made lightweight to fit the compute, storage, and power limits of edge devices. The edge handles high-frequency, fundamental, real-time tasks such as object detection, anomaly pre-screening, and risk-point recognition; the central platform handles more complex task management, result review, data aggregation, model optimization, and global analysis — achieving "fast judgment at the edge, unified management in the cloud."

4. Computer-vision-based target recognition

Combining edge and central analysis, the platform automatically detects key targets in inspection images, including:

• cable marker stones;
• communication manhole covers;
• power and communication poles;
• aerial cables;
• direct-buried cable markers;
• the China Telecom logo;
• construction machinery, excavators, and drilling rigs;
• over-height vehicles and other potential third-party-damage objects.

Through object detection, image classification, and anomaly recognition, the system rapidly extracts useful information from large volumes of data, reducing the need to review images one by one.

5. Hazard recognition and risk assessment

Beyond detecting key targets, the system analyzes abnormal states — damaged, missing, or weed-obscured marker stones; blurred or missing marker text; abnormal manhole covers; sagging, dragging, or damaged cables; machinery near the line; and third-party-damage risk nearby. For suspected hazards detected at the edge, the system automatically generates risk tags, hazard types, risk levels, and handling suggestions, syncing them to the central platform for staff to confirm and handle.

6. Cloud-edge collaborative management

The cloud-edge architecture unifies edge devices and the central platform. The edge handles on-site collection, inference, anomaly pre-screening, and reporting; the central platform handles task and line management, hazard review, ticket flow, statistical analysis, model versioning, and algorithm optimization — meeting real-time on-site needs while keeping unified central oversight.

7. GIS map visualization

With GIS capabilities, the platform visualizes cable lines, inspection tracks, edge-device locations, risk points, hazard images, and handling status. Managers can see line distribution, inspection completion, abnormal points, high-risk areas, and handling progress directly on the map, enabling unified supervision and dispatch.

8. Closed-loop inspection and hazard management

The system manages the full process — planning, task dispatch, on-site collection, edge recognition, platform review, hazard confirmation, ticket assignment, rectification feedback, and archival review. Beyond finding problems, it tracks whether, by whom, when, and how they are handled, improving standardization, traceability, and closed-loop capability.

Phase 1: Business research and scenario mapping

The team conducted in-depth research into trunk-inspection operations, mapping inspection targets, workflows, hazard types, data-collection methods, and on-site pain points, and analyzed the data characteristics of manual, vehicle, drone, mobile-terminal, and edge collection. The key problems to solve:

• wide coverage and heavy manual workload;
• large image/video volumes and low manual processing efficiency;
• unstable on-site networks and data-upload pressure;
• hazard recognition dependent on individual experience, with inconsistent standards;
• a lack of a complete handling loop;
• inspection results hard to aggregate and review.

Based on these, the project chose an "edge pre-screening + unified cloud management" technical route.

Phase 2: Edge-computing architecture design

The team designed an architecture in which the edge and central platform work together: the edge handles on-site collection, image preprocessing, lightweight inference, and abnormal-result reporting; the central platform handles task management, hazard review, ticket flow, GIS display, statistical analysis, and model management. This avoids uploading all raw data centrally and improves on-site availability and response efficiency.

Phase 3: Sample data preparation and model training

Around core targets such as marker stones, manholes, poles, aerial cables, and construction machinery, the team organized samples, annotated data, trained models, and validated results. To handle lighting changes, angle deviations, occlusion, blur, and complex backgrounds, the models went through multiple optimization rounds to improve real-environment stability, and were made lightweight and inference-optimized to run stably on edge devices.

Phase 4: Edge deployment and integration testing

Once the models had basic recognition capability, the team deployed them to edge devices and integration-tested with collection devices, communication links, and the central platform, verifying:

• whether edge devices run the models stably;
• whether images and video are preliminarily recognized on site;
• whether abnormal results upload in time;
• whether the central platform receives and displays edge results;
• whether model versions can be managed and updated centrally;
• whether data can be cached and re-sent under network failures.

Edge deployment gave the system on-site rapid recognition and anomaly pre-screening, laying a foundation for scaled application.

Phase 5: Inspection platform development

On top of the algorithm and edge capabilities, the team built the inspection platform, combining AI recognition, edge results, and inspection workflows. Key modules include:

• inspection task management;
• image and video upload;
• edge recognition-result ingestion;
• automatic AI recognition and analysis;
• hazard-type classification;
• risk-point map display;
• inspection-track management;
• hazard ticket flow;
• statistical reports and a visualization dashboard;
• inspection-result query and archival;
• model-version and edge-device management.

Through this platform, AI and edge capabilities became part of a complete inspection workflow serving front-line work and operations management.

Phase 6: Field pilot and continuous optimization

After the pilot launch, the team continuously optimized models, edge-inference strategies, business rules, and platform interactions using real images, edge runtime data, and field feedback. For complex cases — weed occlusion, long-distance shots, low-resolution images, varied weather and angles — it kept adding training samples to improve adaptability, and refined task lists, hazard confirmation, map positioning, image viewing, edge-device monitoring, and reporting to improve usability.

Phase 7: Capability consolidation and productization

Through the project, Luxijie consolidated telecom-oriented intelligent-inspection capabilities — edge AI deployment, lightweight vision-model optimization, cloud-edge architecture, GIS visualization, closed-loop inspection and hazard management, multi-source data ingestion, and inspection analytics. These apply not only to optical-cable trunk inspection but also to pipeline, power-line, campus, municipal, and transportation-facility inspection, offering strong cross-industry replication value.

1. A cloud-edge collaborative inspection system

The project formed an "edge recognition + central management" inspection system, combining fast on-site analysis with unified central oversight. Edge devices complete basic recognition and anomaly pre-screening; the central platform unifies task management, hazard review, ticket closure, and data analysis.

2. Higher on-site recognition and response efficiency

By deploying lightweight models at the edge, the system performs preliminary recognition and risk assessment on site, reducing reliance on real-time upload and centralized computation. For obvious anomalies and high-risk situations, the edge prioritizes uploading structured results and key screenshots, helping managers spot problems faster.

3. Lower data-transmission pressure

With edge computing, the system performs target recognition and anomaly screening on site and uploads only key results, abnormal images, risk clips, and structured data, reducing ineffective data transmission.

4. Trunk-inspection intelligent-recognition capability

The project built AI image-recognition capability for trunk-inspection scenarios, automatically recognizing marker stones, manholes, poles, cables, machinery, and more, improving inspection-data processing efficiency.

5. More timely hazard detection

The system automatically recognizes risk objects and abnormal states around cable lines, helping staff detect potential hazards earlier and reducing omissions and delayed responses.

6. Visualized inspection management

Through GIS maps and a data dashboard, managers can monitor inspection progress, line coverage, edge-device status, risk-point distribution, and handling status in real time, improving transparency and management efficiency.

7. A closed loop for problem handling

The platform links hazard-recognition results with ticket workflows, achieving closed-loop management from detection, confirmation, dispatch, and rectification to review, raising the standardization of operations.

Value for front-line inspectors

The system reduces repetitive image review, manual record-keeping, and manual hazard judgment, letting inspectors focus on on-site verification and problem handling. Edge recognition completes preliminary analysis on site, giving inspectors faster anomaly alerts and improving field efficiency and inspection quality.

Value for operations management

The platform gives management a full view of inspection status, hazard distribution, edge-device operation, and handling progress, replacing manual aggregation, spreadsheets, and after-the-fact tracking. Through map-based, visual, data-driven management, the team can find high-risk areas faster, optimize inspection resources, and improve trunk-network operations quality.

Value for enterprise IT

The project integrates AI vision, edge computing, GIS, inspection management, ticket closure, and data analytics in one platform, providing foundational capability for an intelligent operations system. The cloud-edge architecture ensures on-site real-time performance while enabling unified central management and continuous optimization, offering a reusable technical base for multi-scenario inspection.

Value for cost reduction and efficiency

Through edge auto-recognition, anomaly pre-screening, and platform-based closed-loop management, the system reduces repetitive manual recognition and ineffective data processing and raises inspection efficiency. Detecting hazards in time helps lower failure probability and subsequent repair costs, improving the stability and safety of communications infrastructure.

Value for Luxijie

The project demonstrates Luxijie's combined strengths in computer vision, edge computing, cloud-edge architecture, lightweight AI deployment, GIS visualization, industry system integration, and business-process digitalization. It is both a representative telecom intelligent-inspection case and a flagship project in the direction of AI + edge computing + industry applications, offering replicable experience for expanding into carrier, power, municipal, campus, and transportation clients.