How Can Smart Algorithms Make Elevators Safer?

Elevators are a cornerstone of urban life, but their safety hinges on tiny details like the bolts that secure landing doors. A loose bolt can cause vibrations, annoying noises, and, in worst cases, dangerous failures. That's why engineers are turning to smart computer‑vision algorithms that can spot these flaws in real time.

The new DS‑EMA model, built on YOLOv8, merges depthwise‑separable convolutions with an Efficient Multi‑scale Attention (EMA) module. Compared to the baseline, it lifts mean average precision (mAP) by 2.8 percentage points, trims parameters to 2.8 M, and still runs at 126 FPS, making it ready for live deployment.

For anti‑loosening lines, DeepLabv3+ has been upgraded with a MobileViT backbone, a Global Attention Mechanism (GAM), and an optimized ASPP dilation rate. The result is a mean IoU of 85.8 %, a 5.4‑point gain, while cutting parameters down to 38.5 M.

Benchmarks against lightweight rivals such as YOLOv5n, YOLOv6n and YOLOv7‑tiny show that DS‑EMA and the revamped DeepLabv3+ deliver the best mix of accuracy, speed and efficiency, ensuring more reliable inspections even under oily or dim lighting conditions.

The DS‑EMA model builds on a YOLOv8 "nano" variant (YOLOv8n), designed for devices with limited resources. Its main innovation is the integration of depthwise‑separable convolutions, which split a traditional filter into a channel‑wise (depthwise) filtering phase and a pointwise combination phase. This dramatically cuts the number of multiply‑add operations while preserving the ability to learn complex spatial features.

The EMA (Efficient Multi‑scale Attention) module adds the ability to focus on details of various sizes. It computes attention maps at multiple scales and merges them through weighted fusion. In practice, the model can give more weight to tiny structures (like bolts) without losing the global view of the door.

For anti‑loosening line segmentation, DeepLabv3+ is equipped with MobileViT, a hybrid network that blends local convolutions with global transformers. MobileViT captures long‑range pixel relationships, essential for distinguishing thin lines amid visual noise. The Global Attention Mechanism (GAM) adds an attention channel that weights the entire image, improving prediction consistency in uneven lighting.

The ASPP (Atrous Spatial Pyramid Pooling) was optimized by adjusting dilation rates to balance receptive‑field coverage and detail resolution. Reducing parameters from 57.6 M to 38.5 M makes the model suitable for embedded systems, such as the microcontrollers found in elevator control panels.

Evaluation metrics include mAP (mean Average Precision) for bolt detection and IoU (Intersection over Union) for line segmentation. A 2.8 % mAP gain and a 5.4 % IoU increase indicate a higher ability to spot critical defects, lowering false negatives, which are especially dangerous in safety contexts. Moreover, the higher recall compared to YOLOv9‑tiny and YOLOv11n ensures that almost all loose bolts are flagged, minimizing omission risk during automated inspections.

Real‑time performance (126 FPS) enables integration into production lines or mobile inspection robots, where every frame matters. The blend of lightweight design, precision, and speed makes DS‑EMA and the upgraded DeepLabv3+ practical solutions for predictive elevator maintenance, helping cut downtime costs and boost user safety.