Cloud-Edge Predictive Maintenance & Remote O&M for Overhead Cranes

As critical assets in continuous production, overhead cranes can incur losses ranging from thousands to tens of thousands of yuan per hour during unplanned downtime. The traditional “scheduled maintenance + breakdown repair” model suffers from two critical flaws: scheduled maintenance replaces parts at fixed intervals regardless of actual condition, leading to waste; and fault shutdowns are completely unpredictable, potentially halting the entire production line. A cloud-edge collaborative predictive maintenance model continuously collects and analyzes operational data to anticipate equipment health trends before failures occur, enabling maintenance on demand. This article provides a complete engineering overview of overhead crane predictive maintenance, covering sensor configuration, edge feature extraction, cloud model training, and the remote operation and maintenance platform.

Sensor Configuration and Data Acquisition Strategy

Sensor deployment for overhead crane predictive maintenance covers three core areas: the hoisting motor (vibration + temperature), the gearbox (vibration + oil temperature), and the brake (temperature + wear detection). Each crane is recommended to be equipped with 4 vibration sensors (IEPE type, frequency range 0.5Hz~10kHz, sensitivity 100mV/g, sampling rate 25.6kHz), 6 temperature sensors (PT100, accuracy ±0.3°C), 3 current sensors (closed-loop Hall effect, accuracy ±1% F.S.), and 1 brake wear sensor (magnetostrictive displacement sensor, accuracy ±0.1mm). Sensor signals are fed into the crane PLC's AI module (SM1231 AI 8×13bit) or a dedicated edge Data Acquisition Module (ADS1256, 24-bit ADC, SPI interface to the edge computing box).

Acquisition Strategy: Vibration sensors sample continuously at 25.6kHz. The edge computing box performs an FFT (1024 points) every 10 seconds, extracts frequency-domain features, and discards the raw signal (not uploaded to the cloud to save bandwidth and storage). Temperature sensors sample every 1 second, with an upload period of 10 seconds (average value). Current sensors sample every 100ms (synchronized with the PLC scan cycle), with a 1-second upload period. Total upstream bandwidth for data acquisition is approximately 2.5KB per crane per second (including vibration features, temperature, current, and status). For 20 cranes, this is about 50KB/s, well within 4G network capacity.

Edge Feature Extraction and FFT Spectrum Analysis

The edge computing box (Jetson Orin NX or IoT2040 gateway) runs a real-time FFT analysis program, performing a 1024-point FFT on the 25.6kHz sampled data for each channel (frequency Resolution: 25Hz, spectrum range: 0~12.8kHz). The extracted frequency-domain features fall into four categories:

Baseline Establishment: The first 30 days after commissioning serve as the baseline period. The edge computes daily averages ±3σ for each feature as the normal range. Baselines are stored locally in an SQLite database on the edge and in the cloud time-series database. When a feature exceeds the baseline range (e.g., vibration RMS exceeds 2x the baseline), the edge immediately sends an alert message to the gateway (MQTT QoS 1). The gateway uploads it to the cloud, and a notification is pushed to the maintenance personnel's APP within 30 seconds. The edge also records raw waveform data of the anomaly (10 seconds before the fault) as supplementary input for cloud model analysis.

Cloud-Based 1D-CNN and LSTM Remaining Life Prediction

The cloud AI model operates in two stages: health state classification (1D-CNN) and remaining useful life regression (LSTM). The 1D-CNN classification model takes as input the 32-dimensional feature vector uploaded from the edge (4 vibration channels × 8 frequency-domain features) and outputs one of 5 health states: Normal (Label 0), Minor Degradation (Label 1), Moderate Degradation (Label 2), Severe Degradation (Label 3), and Pre-Failure (Label 4). Model architecture: Conv1D(64,128,256)+MaxPool+GlobalAvgPool+Dense(128,5), with approximately 86K parameters and a classification accuracy of 94.7% (validated against labeled after-sales maintenance data).

The LSTM remaining useful life prediction model uses a time series of the 32-dimensional feature vectors from the past 7 days as input (time steps = 7×24×6 = 1008, one feature vector every 10 minutes). Through two LSTM layers (128,64) + Dense(32,1), it outputs the RUL prediction in hours. Training data comes from Kelude's after-sales maintenance database (1,200 overhead cranes × an average of 3 years of operational data), totaling approximately 360,000 crane-years of acceleration data. With 5-fold cross-validation, the RUL prediction deviation (MAPE) is ≤15%. The model is incrementally retrained every 24 hours with new operational data (training time ~8 minutes per round).

Predictive Maintenance Work Order Flow and APP Alerts

The ultimate value of predictive maintenance lies in triggering the right maintenance action. Kelude's Cloud Platform routes work orders based on a health-level classification system:

Maintenance App Dashboard Features: Equipment overview (real-time health status of all overhead cranes displayed in red/yellow/green), vibration trend curves (RMS trend and FFT spectrum for any selected time period), historical alarm records (including raw waveform data at the time of the alarm), maintenance work order list (status, assigned technician, estimated completion time), and spare parts inventory lookup (integrated with the enterprise ERP spare parts module). The Kelude app supports both Android and iOS, with data secured via HTTPS and JWT Token authentication.

Remote Diagnostics and OTA Firmware Upgrades

Remote Operation & Maintenance capabilities allow the Kelude after-sales team to resolve the majority of maintenance issues without an on-site visit. Core functions fall into four categories:

Predictive vs. Traditional Maintenance: Key Differences

Sensor Configuration & Technical Parameters

Frequently Asked Questions