Overhead Crane AI Visual Inspection Cloud Platform Launched
Kelude Launches AI Visual Inspection Cloud Platform for Overhead Cranes — Five Core Detection Capabilities Now Available as SaaS. AI visual inspection for overhead cranes is shifting from on-premise deployment to cloud-based services.
AI visual inspection technology for overhead cranes is evolving from standalone on-premise systems to cloud-native services. Kelude Heavy Industry has officially launched its AI Visual Inspection Cloud Platform for overhead cranes, packaging five core detection capabilities — safety monitoring, load positioning, wire rope inspection, weld inspection, and visual SLAM mapping — as a SaaS offering. Users no longer need to purchase edge computing hardware or set up inference servers; they simply install industrial cameras and a communication terminal on the crane, and real-time detection results are delivered through the cloud-based AI inference engine. During beta testing, the platform was deployed and validated at 12 customer sites, processing over 500,000 frames of inspection imagery with an average AI inference response time under 200 ms.
This article provides a systematic overview of the technical solution behind the platform, covering four key aspects: system architecture, technical implementation of the five detection functions, SaaS deployment model, and engineering deployment results.
Cloud Platform Architecture Overview
The AI Visual Inspection Cloud Platform for overhead cranes adopts a three-layer "edge-device-cloud" architecture. AI inference is centralized in the cloud, while the device side handles only image acquisition and transmission, significantly reducing on-crane hardware costs and operational complexity.
Device Layer (Terminal Acquisition): Each crane is equipped with 1–4 industrial AI cameras (optional based on the inspection scenario; resolution from 5 to 20 megapixels, global shutter CMOS), connected via Gigabit Ethernet to an on-crane communication terminal (5G industrial gateway or 4G DTU). The terminal compresses raw images using JPEG encoding (compression ratio of approximately 15:1 with negligible quality loss), adds timestamps and device ID watermarks, then uploads the data to the cloud inference platform over 5G/4G networks. A single camera typically uploads at 2–4 Mbps (1080P@15fps compressed), and up to 8 cameras can stream simultaneously over 5G.
Cloud Layer (AI Inference): Deployed on Alibaba Cloud GPU clusters (NVIDIA A10 GPUs, delivering approximately 31 TFLOPS FP16 per card), the platform runs YOLOv8s, YOLOv8s-weld, EfficientNet-B4, and ORB-SLAM3 inference engines. The cloud inference engine employs model parallelism and dynamic batching (maximum batch size of 32), enabling a single GPU to handle real-time detection for four concurrent video streams, with per-frame inference latency of 8–15 ms on the GPU (including image decompression and preprocessing). Inference results are pushed in JSON format to user endpoints (web portal, mobile app, or customer MES system) via MQTT, with a median push latency of 45 ms.
Application Layer (User Interaction & Management): Provides a web-based admin console (compatible with Chrome/Edge/Firefox) and a mobile app (iOS/Android). Key features include: real-time detection dashboard (detection overlays on live video streams, historical alarm timeline, detection statistics charts), device management (crane registration, camera configuration, detection area calibration), alarm policy configuration (threshold settings, alarm channel selection — WeChat/SMS/email), detection data export (CSV/PDF, filterable by device and time range), and API management (API token generation for MES/ERP system integration).
The data flow across the three-layer architecture is: device cameras capture → terminal compresses and encodes → upload over 5G/4G → cloud GPU inference → results pushed via MQTT → user display. End-to-end latency (from image capture to detection results visible on the user side) typically ranges from 300–600 ms over 4G and 120–250 ms over 5G. All data transmission is encrypted with TLS 1.3, and cloud data storage complies with the Level 2 protection requirements of GB/T 22239-2019, the Chinese standard for cybersecurity classified protection.
| Tier | Component | Core Hardware/Software | CriticalIndicator |
|---|---|---|---|
| Edge Side | image acquisitionand Transmission | industrial camera+5G/4GCommunication Terminal | Compression Ratio15:1,Upload Bitrate2~4Mbps/Channel |
| Cloud | AIinference engine | Alibaba CloudA10 GPU + YOLOv8Series | SingleGPUProcessing4Video Channel,Inference8~15ms/Frame |
| application layer | User Interaction and Management | WebBackend+MobileAPP+MQTTPush Notification | Push NotificationlatencyMedian45ms,End-to-End<600ms |
Five Core Detection Modules
The cloud platform integrates five core AI visual inspection modules for overhead cranes. Users can subscribe to modules individually based on their specific needs, avoiding costs for features they don't require:
Module 1: Safety Monitoring (Personnel Intrusion Detection). Personnel detection in crane lifting zones based on the YOLOv8s model, with three cameras providing redundant coverage of both end carriage sides and the area beneath the trolley. A three-tier alarm strategy is implemented—a yellow caution zone (5–8 m) triggers an audible and visual alarm, an orange warning zone (3–5 m) triggers deceleration, and a red danger zone (<3 m) triggers STO (Safe Torque Off). Under normal lighting conditions, mAP ≥ 95%; under low-light conditions with infrared illumination, mAP ≥ 90%. Compliant with GB/T 28264-2017 Safety Monitoring and Management System for Cranes. Monthly subscription fee: approximately ¥800 per crane.
Module 2: Load Positioning (Vision-Guided Alignment). The YOLOv8s model detects four types of loads—steel coils, steel plates, equipment, and dies. A coordinate transformation module maps pixel coordinates to millimeter-level ground-plane coordinates, achieving positioning accuracy of ±5 mm. Supports integration with Siemens S7-1200/1500 PLCs via PROFINET gateway, automatically generating motion commands for the crane bridge, trolley, and hoisting mechanisms. Monthly subscription fee: approximately ¥1,200 per crane.
Module 3: AI Wire Rope Inspection. Multiple cameras work in coordination (top-down, side-angle, and bottom-up views). The YOLOv8s model detects three defect categories—wire breaks, wear, and corrosion—while an EfficientNet-B4 model performs secondary fine-grained classification into 11 sub-types, achieving a combined mAP ≥ 0.93. Alerts are linked to the PHM (Prognostics and Health Management) platform, with an average response time of no more than 30 seconds from detection to work-order generation. Monthly subscription fee: approximately ¥600 per crane.
Module 4: AI Weld Inspection. The YOLOv8s-weld model (enhanced with a small-object detection head) detects five defect types—cracks, incomplete penetration, porosity, slag inclusion, and undercut—achieving a combined mAP ≥ 0.94. The evaluation layer automatically aligns with the GB/T 3323-2020 quality grading standard, outputting ratings from Grade I to Grade IV. Supports three deployment configurations: AGV inspection vehicles, handheld detectors, and fixed workstations. Monthly subscription fee: approximately ¥500 per crane.
Module 5: Visual SLAM Mapping and Positioning. Fusing the ORB-SLAM3 and VINS-Mono frameworks, this module uses monocular/stereo cameras paired with a 6-axis IMU as sensors to deliver six-degree-of-freedom pose estimation and sparse map construction in factory environments. Positioning accuracy is ±10 cm, and tightly coupled IMU pre-integration reduces tracking loss under vibration from 8% to 0.5%. Monthly subscription fee: approximately ¥1,000 per crane.
All five modules can be subscribed to independently or as a bundle. The complete five-module package is priced at ¥2,800 per crane per month—approximately 32% savings compared to the combined individual subscription total of ¥4,100. Both monthly and annual billing options are available, with annual subscriptions receiving a 10% discount.
SaaS Deployment Model and Technical Advantages
Compared to traditional on-premises deployment, the SaaS-based AI visual inspection cloud platform offers significant advantages across three dimensions: deployment speed, compute elasticity, and operational cost.
Deployment Speed: On-premises deployment requires purchasing Jetson Orin NX/AGX edge computing devices (hardware cost of approximately ¥5,000–15,000 per unit), installing the inference environment (TensorRT, CUDA, model export, etc., typically taking 1–2 business days), and maintaining inference services (software updates, model upgrades, hardware troubleshooting). In the SaaS model, the edge side only needs an industrial camera and a communication terminal (existing surveillance cameras can be reused as long as they support the RTSP protocol). Once powered on and connected to the network, the communication terminal automatically registers with the cloud platform. Camera calibration and detection zone configuration are completed through a web-based dashboard, and detection can begin immediately. The typical deployment cycle is reduced from 3–5 days for on-premises to 2–4 hours.
Compute Elasticity: The Jetson Orin NX used in on-premises deployments has a fixed compute capacity of 100 TOPS, which cannot be dynamically adjusted based on actual workload. During idle detection periods (e.g., overnight shutdowns), compute resources are wasted; during peak detection times (e.g., multiple cranes operating simultaneously), compute capacity can become a bottleneck. The SaaS cloud GPU cluster supports on-demand elastic scaling—GPU instances automatically scale up when concurrent detection requests increase (from 4 video streams/GPU to 8 streams/GPU, with a scaling time of approximately 2 minutes) and scale down to release resources when requests decrease. Field measurements show that the cloud inference throughput for 20 cranes running simultaneously (60 cameras total) reaches 900 frames per second—a 4.5× improvement over the on-premises Jetson solution (estimated at approximately 200 frames per second for the same scale).
Operational Cost: On-premises deployment of AI visual inspection systems requires 1–2 Jetson modules per crane. For a five-crane setup, total hardware investment is approximately ¥30,000–100,000. Additionally, AI algorithm engineers must be hired for model maintenance, upgrades, and troubleshooting, adding approximately ¥150,000–250,000 in annual labor costs. The SaaS model requires no upfront hardware investment—pay-as-you-go pricing applies. Model upgrades (e.g., upgrading from YOLOv8s to YOLOv11 or adding new detection categories) are handled centrally by the cloud platform, requiring no action from the user. The annual subscription for all five modules across five cranes is approximately ¥168,000 (including cloud compute, storage, and model maintenance), delivering approximately 40–55% savings compared to the three-year TCO of an on-premises solution (hardware depreciation + labor + maintenance).
Beta Validation and Measured Results
Since October 2025, the cloud platform has undergone beta testing at 12 customer sites across four industries: metallurgy (5 sites), shipbuilding (3 sites), heavy equipment manufacturing (3 sites), and new energy (1 site). The beta program used a free-trial-plus-feedback model, with each crane configured with 1–3 detection modules, accumulating 180 days of operation. Key measured data is presented below:
In safety monitoring module testing, 142 personnel intrusion events occurred across the 12 sites. The cloud platform's AI detection successfully triggered alarms for 136 events, with 6 missed detections (all in scenarios combining low light with severe occlusion), achieving a recall rate of 95.8% and a false alarm rate of 0.3 per crane per day. The wire rope inspection module inspected over 12,000 meters of wire rope across 3 metallurgical sites, identifying 38 weld defects at Grade 2 and above, with a 93.2% agreement rate compared to manual shutdown re-inspection results. The weld inspection module was validated at 1 shipbuilding site against UT (ultrasonic testing) on main girder butt welds, with a positioning deviation of less than 3 mm between AI-detected defects and UT re-inspection.
The customer satisfaction survey (NPS, Net Promoter Score) yielded a score of 62, well above the industry average NPS baseline for SaaS software (approximately 40). The three most highly rated features in customer feedback were: rapid deployment with zero hardware (average score 4.7/5), maintenance-free operation with automatic cloud model updates (4.5/5), and flexible per-module subscription billing (4.3/5). Key improvement suggestions centered on result caching and offline detection capabilities for unstable network environments—this feature has been scheduled for the Q3 2026 release.
FAQ
Q: Does the cloud platform's AI detection meet the real-time requirements for crane safety control?
Q: What is the end-to-end latency of the Kelude overhead crane AI visual inspection cloud platform?
A: The end-to-end latency of the Kelude Heavy Industry overhead crane AI visual inspection cloud platform is 120–250 ms over 5G networks and 300–600 ms over 4G networks. For safety monitoring applications, we recommend using detection results as supplementary safety information rather than direct safety circuit control signals. Per GB/T 28264-2017 Safety Monitoring and Management System for Lifting Appliances, signals that directly control the safety circuit must have latency below 100 ms. The current cloud platform solution is suitable for alarm notification and auxiliary monitoring; for direct safety interlock control, we recommend a local Jetson edge inference solution (end-to-end latency <100 ms). The cloud platform supports hybrid deployment with the edge solution—the edge tier performs real-time safety control while the cloud handles data storage and remote monitoring.
Q: Can the system be used in workshops with poor network conditions (no 5G coverage, unstable 4G signal)?
A: Yes. The communication terminal supports dual-SIM dual-module redundancy and local caching: when the network is healthy, images are uploaded in real time and inference results are received; when the network drops, the system automatically switches to the standby carrier (switchover time <15 seconds) and caches images captured during the outage to local terminal storage (maximum cache capacity 256 GB, storing approximately 72 hours of 1080P@15fps compressed video). Once the network is restored, cached images are automatically uploaded and re-processed, ensuring no detection data is lost. For workshops with persistently poor connectivity, we recommend a hybrid deployment—a lightweight model runs on the edge device for real-time detection, while the cloud handles model updates and data backup.
Q: Could cloud-stored detection data expose the customer's workshop production information?
A: The platform is designed around the principle of minimal data collection: the cloud stores only AI detection results (object class, bounding box coordinates, confidence scores, and timestamps) and key-frame images that triggered alarms—not the full raw video stream. Customers can configure the data retention period independently via the web console (7/30/90 days), and expired data is automatically purged. Data storage uses a multi-tenant physical isolation architecture (each customer gets a dedicated database instance), with encrypted transmission (TLS 1.3) and encrypted storage (AES-256). The platform has passed MLPS Level 2 assessment and holds ISO 27001 information security management system certification. Per GB/T 22239-2019, MLPS Level 2 applies to general industrial internet systems and meets the data security requirements of overhead crane remote operation and maintenance scenarios.
Q: Under the SaaS subscription model, how are model versions upgraded? Can customers use their own models?
A: Detection models are maintained and upgraded centrally by the cloud platform. The platform currently runs YOLOv8s-series models, with a model iteration released quarterly (incremental training on new training data followed by accuracy evaluation). New model versions undergo regression testing (accuracy must not degrade on the standard test set before release) and are then rolled out via canary deployment. For customers with custom detection needs (e.g., identifying special workpiece types), the platform allows customers to upload their own training datasets; platform engineers perform model fine-tuning and deploy the tuned model to a dedicated inference instance for that customer. Fine-tuning is billed per session (approximately $450–$1,200 per session), with a turnaround of about 5–10 business days. Model upgrades and fine-tuning do not affect other customers' normal operations.
