From Passive Alarm to Predictive Maintenance:
On the hot rolling and cold rolling production lines of the steel industry, the multi-row rolling bearings of rolling mills are the core components determining production continuity. These bearings endure long-term combined axial and radial loads, making the rolling elements and cages highly susceptible to hidden damage. The traditional maintenance model relying on manual inspections and post-failure shutdowns has long been unable to meet the efficient production demands of modern steel enterprises.
In the past, most steel mills employed a passive alarm mechanism — the system would only issue a shutdown command when bearing faults had already developed to the point of causing significant vibration exceeding standards and triggering protection thresholds. By that time, roll damage and surface quality defects in the strip steel had often already occurred. A single unplanned shutdown could bring direct economic losses of hundreds of thousands of yuan, not to mention the chain reaction caused by disrupting the entire production schedule.
For a long time, rolling mill bearing fault monitoring has faced three difficult-to-break-through bottlenecks:
Severe Signal Distortion: Rolling mill sites are filled with mechanical noise from various motors and gears. Signals captured by ordinary vibration acquisition equipment are often drowned in a large amount of irrelevant noise. The early, subtle vibration characteristics of bearing damage are completely masked, making it very difficult for maintenance personnel to identify the signals of incipient faults from the massive data.
Feature Confusion in Multi-Row Bearings: Multiple rows of rolling mill bearings jointly bear radial forces. Vertical impact signals caused by faults are significantly weakened. Traditional single-channel vibration analysis methods struggle to capture the rich fault characteristics contained in axial vibrations, often leading to missed and false alarms.
Data Silos: On-site, a large number of vibration sensors and temperature sensors on older rolling mills mostly output data via RS485 serial ports. Traditional serial to ethernet devices could only perform simple data pass-through, unable to complete real-time analysis at the edge. All raw vibration data had to be sent back to the cloud platform for processing, not only occupying significant bandwidth but also missing the optimal fault warning window due to data transmission delays.
These pain points directly led to the rolling mill bearing fault warning accuracy in traditional modes hovering around 60% for a long time. A large number of early hidden dangers could not be detected promptly, and predictive maintenance remained at the conceptual level, difficult to truly implement on the production floor.
With the maturation of Industrial IoT technology, serial to ethernet devices and edge data acquisition gateways equipped with edge computing capabilities are fundamentally changing this situation. Taking the USR-N720 edge data acquisition gateway supporting Ethernet access as an example, this industrial-grade device, certified by 3C, CE, ROHS, and WEEE, uses a RISC-V core with a main frequency of up to 600MHz, integrates 2 RS485 interfaces and 1 network port, perfectly adapting to the serial port access needs of various legacy sensors on rolling mill sites, breaking the data barriers of traditional equipment.
Unlike ordinary serial to ethernet devices that can only perform data forwarding, the USR-N720 can directly complete real-time acquisition, preprocessing, and edge computing of vibration data for 1000 points at the edge, without needing to transmit all raw vibration data back to the cloud.
Relying on the built-in Adaptive Multivariate Variational Mode Decomposition (AMVMD) algorithm, the gateway can perform synchronous decoupling processing of the axial and vertical multi-channel vibration signals from rolling mill bearings: Using the mean of weighted permutation entropy as the fitness factor, it automatically completes the optimal selection of modal decomposition parameters K and α through a genetic algorithm, while introducing iterative operators to accelerate the optimization process. It accurately separates the characteristic modes corresponding to different faults from the complex on-site noise, completely extracting the originally masked early, subtle vibration characteristics of damage.
After completing signal preprocessing, the gateway can integrate time-domain features, frequency-domain features, and operating condition information into a unified feature representation. It then completes real-time inference through a locally deployed lightweight convolutional neural network model, directly outputting the bearing health score and fault warning results at the edge side. This "local analysis at the edge" model avoids the bandwidth pressure caused by uploading massive amounts of vibration data and compresses the fault identification response delay to the millisecond level, truly achieving real-time perception of vibration anomalies.
This vibration analysis solution based on edge data acquisition gateways has undergone industrial validation at rolling mill sites of several domestic steel enterprises. Compared to the traditional passive alarm mode, the fault warning accuracy rate has achieved a significant 90% improvement. In actual operation, the system can identify hidden faults in rolling mill bearings, such as early pitting and slight cage deformation, up to 72 hours in advance. Maintenance personnel can completely use planned shutdown windows during production gaps to complete spare part replacements, thoroughly bidding farewell to the passive situation of "emergency repairs after faults occur."
More crucially, addressing the common problems of scarce fault samples and imbalanced datasets in industrial settings, this solution also incorporates a data augmentation mechanism based on a deep convolutional generative adversarial network. By generating high-quality simulated fault samples to supplement the dataset, the diagnostic model can maintain extremely high recognition accuracy even in small-sample scenarios. Actual measurement data from a hot rolling production line shows that after the retrofit, the unplanned downtime of rolling mill bearings decreased by 85%, the average service life of bearings extended by 22%, and annual maintenance cost savings for a single production line exceeded one million yuan.
From the past passive response of "alarming only after a fault occurs" to the current predictive maintenance of "warning as soon as a hidden danger sprouts," serial to ethernet and edge data acquisition gateways are becoming key enablers for the intelligent equipment upgrade of steel enterprises. It does not require large-scale modification of existing rolling mill equipment. By simply connecting to legacy sensors via serial ports, it can achieve a leapfrog upgrade in vibration analysis capability, realizing a qualitative leap in equipment operation and maintenance efficiency with extremely low retrofit costs, and solidifying the technological foundation for the efficient and safe production of the steel industry.
