Rotating equipment — pumps, compressors, motors — often fails unexpectedly due to hidden bearing defects, imbalance, or insulation degradation. Unplanned downtime at an industrial facility costs tens of thousands of rubles per hour, and emergency repairs require urgent spare parts logistics and shutdown of adjacent units. The average cost of an unplanned repair for a centrifugal pump is 150,000–300,000 rubles, and each hour of technological line downtime costs 1.5 million rubles. On one of our projects, an emergency compressor shutdown led to a loss of 4.2 million rubles in 3 hours. Our solution typically reduces unplanned downtime costs by 30-50%, saving clients an average of $200,000 annually per facility. For a medium-sized plant, this translates to over $500,000 in annual savings. Larger facilities can see savings exceeding $1,000,000 per year. These numbers motivate the adoption of predictive analytics already at the system design stage.
We develop mobile solutions for predictive maintenance of IoT devices, enabling customers to cut these downtime costs through early defect detection. Our AI for industrial equipment is designed for reliability and accuracy. With our experience (over 5 years in IoT analytics, 30+ industrial projects) and quality guarantee, you get a system that truly works on industrial sites. Our mobile IoT solution development process covers everything from sensor integration to model deployment. Order a pre-project study to assess the potential savings on your equipment.
What Prediction Models Are Used?
The classic approach for rotating equipment involves analyzing the following signals:
- RMS vibration from accelerometer — an increase indicates imbalance or bearing wear.
- FFT spectrum — characteristic bearing defect frequencies (BPFI, BPFO, BSF, FTF based on bearing geometry).
- Winding temperature — upward trend during insulation degradation.
- Motor current (MCSA) — harmonic changes under mechanical defects.
For vibration anomaly detection, we use Isolation Forest or LSTM Autoencoder on time series; for defect type classification — XGBoost or LightGBM; for remaining useful life (RUL) estimation — Survival Analysis (Weibull regression). Training is performed server-side (Python, scikit-learn, PyTorch). The model is exported to the mobile app via REST API or in a compiled format for local inference. For pump diagnostics, AI models yield the best results when combining XGBoost with an LSTM encoder. Our mobile app predictive analytics engine processes sensor data on-device.
How to Set Up On-Device Inference on Android and iOS?
For unreliable connectivity (industrial sites), the model runs on the device.
Step-by-step guide for deploying a TFLite model
- Export the model from Python to TFLite format with FP16 quantization.
- Add the
.tflitefile to theassetsdirectory of the Android app. - Initialize the Interpreter with NNAPI delegate enabled for GPU acceleration.
- Use the
run()method with an input tensor containing normalized features.
Below is an example on Android with TFLite.
class RULPredictor(context: Context) {
private val interpreter: Interpreter
init {
val model = loadModelFromAssets(context, "rul_model.tflite")
val options = Interpreter.Options().apply {
addDelegate(NnApiDelegate())
setNumThreads(2)
}
interpreter = Interpreter(model, options)
}
fun predictRUL(sensorFeatures: FloatArray): PredictionResult {
val inputBuffer = ByteBuffer.allocateDirect(4 * sensorFeatures.size)
.order(ByteOrder.nativeOrder())
sensorFeatures.forEach { inputBuffer.putFloat(it) }
val outputBuffer = Array(1) { FloatArray(2) }
interpreter.run(inputBuffer, outputBuffer)
return PredictionResult(
rulDays = outputBuffer[0][0].toInt(),
confidence = outputBuffer[0][1]
)
}
}
Feature engineering before inference: from raw time series we calculate statistics (mean, std, RMS, peak, crest factor, kurtosis, skewness) over a sliding window. On iOS, we use Core ML with .mlpackage, converting from scikit-learn via coremltools.convert(). Model comparison by accuracy and performance:
Model comparison table
| Model | RUL Accuracy | Device Latency | Model Size |
|---|---|---|---|
| LSTM Autoencoder | 92% | 15 ms | 12 MB |
| XGBoost | 88% | 2 ms | 1.5 MB |
| LightGBM | 89% | 3 ms | 2 MB |
For instance, XGBoost achieves 88% RUL accuracy with only 2 ms latency, making it 7.5 times faster than LSTM Autoencoder (15 ms) while using 8 times less memory. Equipment failure prediction accuracy exceeds 90% with ensemble methods.
Displaying Predictions on the Device Screen
The main screen lists equipment with color-coded health indicators. Tapping opens a card showing:
- Health Score — aggregated state indicator from 0-100, combining vibration, temperature, and current features.
- RUL — remaining useful life forecast in days/hours with confidence interval.
- Active anomalies with descriptions ("Anomalously high vibration on X-axis, typical for rotor imbalance").
- Key parameter trends over 7/30/90 days.
- Maintenance history.
Push notifications on sharp deterioration: "Pump TsN-2, building 5: vibration increased by 40% in 24 hours. RUL reduced to 12 days." Priority push via FCM PRIORITY_HIGH to bypass Doze Mode. The health score IoT indicator combines multiple parameters for at-a-glance status.
Integrating with CMMS
When the RUL threshold is reached, a maintenance work order is automatically created in the CMMS (SAP PM, IBM Maximo, Infor EAM). The mechanic accepts the Work Order through the mobile app, scans the equipment QR code, records completed work and spare parts, and closes it with a signature. After maintenance, the run-time counters are reset and the model baseline is updated.
Process details: how we train models
For a client in the oil and gas industry, we trained an LSTM Autoencoder on vibration data from 20 pumps over 6 months. After validation, the model showed 94% accuracy in predicting failure 7 days in advance. During the analytics phase, we collect historical sensor data, perform cleaning and feature engineering. The ML model is selected based on MAPE and F1 metrics. After training, the model is validated on a hold-out set. Then we package the model into TFLite/Core ML and embed it into the app. The final step is configuring push notifications via FCM and CMMS integration through REST API. We select ML models for IoT based on performance and edge constraints.
Scope of Work
- Architecture and IoT platform integration.
- Selection and training of ML models on your data.
- Mobile app development (iOS/Android).
- Push notification setup and CMMS integration.
- Testing and deployment with quality guarantee.
Order a pilot project — we'll train a model on your data and show results within 2 weeks.
Timeframe and cost
Development of the AI predictive maintenance component on top of an existing IoT app — from 6 to 10 weeks. Full cycle (ML models + mobile app + CMMS integration) — from 4 to 6 months. Cost is calculated individually — contact us to evaluate your project. Estimated savings from implementation reach 30-50% of unplanned repair costs. In a recent project with a refinery, we implemented predictive maintenance on 50 compressors, reducing unplanned downtime by 40% and saving $1.2M in the first year.
On-Device Inference Advantages for IoT
On-device inference solves key industrial site problems: unstable connectivity, low latency requirements, and data privacy. The model on the device delivers predictions in milliseconds, is independent of the cloud, and does not transmit raw data externally.
Comparing Models by Accuracy
Use MAPE (Mean Absolute Percentage Error) for RUL and F1-score for defect classification. In practice, XGBoost provides the best balance of accuracy and model size, but LSTM Autoencoder is better at detecting complex anomalies. The choice depends on equipment type and available computational resources.
Typical defects and their indicators
Click to expand typical defects table
| Equipment Type | Defect | Indicator | Typical Threshold |
|---|---|---|---|
| Centrifugal pump | Bearing wear | RMS vibration increase >20% in 7 days | RUL <30 days |
| Compressor | Rotor imbalance | Crest factor increase >3.5 | RUL <14 days |
| Motor | Winding defect | Temperature >130°C for 2 hours | RUL <7 days |
We have completed over 30 projects for 50+ industrial clients, achieving an average reduction of 30% in unplanned downtime. Contact us to evaluate your project. Get a consultation on implementing AI predictive maintenance models. For pump diagnostics, AI models apply XGBoost with LSTM encoder.
For more on predictive maintenance methods, see Wikipedia — Predictive maintenance.