Control of Measurement Reliability in Technical Monitoring Instruments Using a Cascade of Autoencoders

The paper proposes a method for ensuring measurement reliability in technical monitoring systems – a cascaded model C-LPC-AE that combines informative spectral feature extraction based on Linear Predictive Coding (LPC) with a two-stage architecture of convolutional autoencoders. The method is designed for bearing condition diagnostics and simultaneous verification of sensor operational integrity, which is particularly relevant in digital industry environments requiring high autonomy and reliability of monitoring systems. The first stage of the cascade, trained on signals from a healthy bearing with properly mounted sensors, performs anomaly detection based on reconstruction error. The second stage, trained on data with a loosened accelerometer mount, analyzes the nature of the anomaly and enables differentiation between bearing faults and signal distortions caused by improper sensor installation. A key advantage of the approach is that it does not require data from actual equipment failures: training is performed exclusively on easily reproducible conditions — normal operation and simulated sensor mounting faults. Experiments were conducted using data from the SpectraQuest test rig, including bearings with an artificially introduced outer race defect. The results demonstrate high model sensitivity to actual bearing defects and sensor mounting issues. The use of LPC-based features ensures compact signal representation and reduces computational load, making the proposed approach promising for integration into real-time industrial diagnostic systems.