Mehrdad Heydarzadeh

A Two-Stage Fault Detection and Isolation Platform for Industrial Systems Using Residual Evaluation

Fault detection and isolation (FDI) is an important part of modern industrial systems, and plays a vital role in maintainability, safety, and reliability of process. We propose a novel FDI architecture based on a predictive model for fault-free process. We use the least-square support vector machine for identifying a nonlinear system and detecting its faults that may occur. Wavelet analysis on residual is used for fault isolation. An average accuracy of 95% has been achieved for all the abrupt faults of the well-known development and application of methods for actuator diagnosis in industrial control systems benchmark.

Aging Precursor Identification and Lifetime Estimation for Thermally Aged Discrete Package Silicon Power Switches

Thermal/power cycles are widely acknowledged methods to accelerate the package related failures. Many studies have focused on one particular aging precursor at a time and continuously monitored it using custom-built circuits. Due to the difficulties in taking sensitive measurements, the reported findings are more on the quantities requiring less sensitive measurements. In this paper, two custom-designed testbeds are used to age a number of power MOSFETs and insulated gate bipolar transistors. An automated curve tracer is utilized to capture parametric variations in I–V curves, parasitic capacitances, and gate charges at certain time intervals. The results suggest that the only viable aging precursors are the on-state voltage drop/on-state resistance, body diode voltage drop, parasitic capacitances, and gate threshold voltage for die attach solder and gate-oxide degradation mechanisms. Based on the experimental results, gate threshold voltage variation is empirically modeled to estimate the remaining useful lifetime of the switches experiencing gate oxide degradation. The model parameters are found by the least squares method applied to inliers determined by the random sample and consensus outlier removal algorithm.

Remaining Useful Lifetime Estimation for Power MOSFETs Under Thermal Stress With RANSAC Outlier Removal

Reliability of power converters is crucial for mission critical systems. Among the components that are susceptible to failure, power semiconductor devices are one of the major causes of the power converter failures. This paper focuses on the remaining useful lifetime (RUL) estimation of degraded power MOSFETs, which are stressed by thermal cycling. The relative change in on-state resistance is identified as the fault signature. A data-driven RUL estimation algorithm based on a linear approximation model is proposed. The empirical coefficients are estimated by the classical least squares, where the outliers are removed by random sample consensus (RANSAC) algorithm. A sliding window approach is used to track the nonlinearities. The window size, threshold value, and number of samples used by RANSAC are optimized with the genetic algorithm. The accuracy of the proposed RUL estimation tool is verified on a number of thermally aged discrete power MOSFET data.

Gearbox Fault Diagnosis Using Power Spectral Analysis

Gearbox faults are the most important reason for failure of mechanical systems. In this paper, we propose a novel method for gearbox fault diagnosis based on vibration signal reported by accelerometers. We propose parametric power spectral analysis and support vector machine for feature extraction and classification, respectively. The proposed feature extraction technique reduces the dimensionality of the vibration signal and captures frequency characteristics simultaneously. We apply our technique to a well-known bearing benchmark dataset. The cross-validation indicates excellent performance and accuracy.

Gear fault diagnosis using discrete wavelet transform and deep neural networks

Automatic fault diagnosis is an inseparable part of todays electromechanical systems. Advanced signal processing and machine learning techniques are required to address variabilities and uncertainties associated with the monitoring signals. In this paper, deep neural networks are employed to diagnose five classes of gearbox faults applied to three common monitoring signals, i.e. vibration, acoustic and torque. Discrete wavelet transform is used to provide the initial features as the inputs of the network. A test-rig based on a 250W three-phase squirrel cage induction machine shaft connected to a single stage helical gear is built for validation of the proposed method. The experimental results indicate accurate fault diagnosis in various conditions such as different modalities, signal variabilities, and load conditions.

Non-invasive gearbox fault diagnosis using scattering transform of acoustic emission

Monitoring acoustic emission from mechanical systems is an effective non-invasive way of diagnosing both system performance as well as short-term/long-term system failures. A difficulty however in fault detection in such systems is inter-class variability caused by non-uniform or unknown load conditions which decrease the classification accuracy. In this paper, a scattering transform is employed to diagnose gearbox faults using acoustic emission analysis. The results analysis and solution shows that a two layer scattering transform can diagnose four gearbox faults with an average accuracy of 97% even if the system is not exposed to data of all loads in the training phase.

In-bed posture classification using deep autoencoders

Pressure ulcers are high prevalence complications among bed-bound patients which are not only extremely painful and difficult to treat, but also impose a great burden in our health-care system. We target automatic posture detection which is a key module in all pressure ulcer monitoring platforms. Using data collected from a commercially-available pressure mapping system, we applied deep neural networks to automatically classify in-bed posture using features extracted from the histogram of gradient technique. High accuracy of up to 98% was achieved in classifying five different in-bed postures for more than 60,000 pressure images.

A robust remaining useful lifetime estimation method for discrete power MOSFETs

This paper focuses on the remaining useful lifetime (RUL) estimation of power MOSFETs, which are stressed by thermal cycling. The relative change in on-state resistance is identified as the aging precursor for die attach solder degradation after exhaustive experiments. A data-driven RUL estimation algorithm based on a linear approximation model is proposed. The empirical coefficients are estimated by the classical least squares method. However, the initial part of the data contains a significant number of outliers, which decreases the estimation accuracy. In order to remove the outlier effect and make the estimation robust, least-squares method is applied to only inliers that are determined by random sample consensus (RANSAC) algorithm. With the exclusion of outliers, the RUL estimation is improved for the ones that contain outliers. The accuracy of the proposed RUL estimation tool is verified on a number of thermally aged discrete power MOSFET data.

Model-Free Testing of Analog Circuits

Analog circuits contribute to as much test cost as digital circuits. Traditional model-based testing (e.g. catastrophic, range-based, regression models) had limited success due to diversity of circuits and contributing metrics. We propose a model-free analog circuit testing using a combination of advanced signal processing and machine learning techniques with low computation and memory requirements. These unique features make the proposed test methodology scalable, reusable and suitable for both off-chip and on-chip test possibilities.

Bayesian remaining useful lifetime prediction of thermally aged power MOSFETs

The demand for more reliable power conversion is ever increasing. This necessitates smart gate drivers or smart system controllers that monitor the components that are susceptible to failure. Together with the electrolytic capacitors, power semiconductor devices are among the weakest components in a power converter. In an effort to predict power MOSFET aging, this paper proposes an remaining useful lifetime (RUL) estimation algorithm for degraded power MOSFETs, which are exposed to high amplitude thermal cycles. The relative change in on-state resistance is identified as the fault signature. A data-driven RUL estimation algorithm based on a linear model approximation is proposed. The outliers present particularly in the beginning part of the data decrease the accuracy of the estimation with classical least-squares method. In this paper, a Bayesian Interference estimator is proposed to improve the accuracy through incorporating prior knowledge to estimation. The accuracy of the proposed RUL estimation tool is verified on the collected experimental data of thermally aged discrete power MOSFETs.