Power electronic converters play a vital role across modern electrical infrastructures, facilitating efficient energy conversion in renewable generation, electric vehicles, flexible alternating current transmission systems, high-voltage direct current lines, and industrial motor drives. Within these converters, fast-switching semiconductor devices known as insulated gate bipolar transistors endure severe electrical, thermal, and environmental stresses that gradually deteriorate their physical structure over prolonged operational service. Component reliability surveys covering eighty commercial companies and two hundred industrial products show that power semiconductor devices cause twenty-one percent of all converter failures, second only to capacitors at thirty percent. When critical switching components fail unexpectedly in service, the resulting breakdowns cause extensive system downtime, substantial financial loss, and severe safety hazards. To address these vulnerabilities, researchers prioritize component-level prognostics and health management frameworks capable of estimating remaining useful life before catastrophic failure terminates device operation.
Component-level lifetime prediction strategies in power electronics traditionally fall into handbook-driven, physics-of-failure model-driven, and data-driven prognostic categories. Handbook-driven lifetime estimation offers operational simplicity and uses historical field data, but consistently suffers from inaccurate predictions across dynamic operating conditions. Physics-of-failure approaches evaluate specific physical degradation mechanisms, relying on mathematical models such as the Arrhenius equation and Coffin-Manson formulations to model thermal wear and fatigue. However, physics-of-failure methods encounter major obstacles due to the mathematical complexity required to model root failure mechanisms in packaged semiconductor devices under fluctuating stresses. In contrast, data-driven frameworks utilize machine learning algorithms to uncover degradation patterns directly from logged precursor measurements. These data-driven techniques require substantially less structural modeling complexity than physics-of-failure equations while delivering high predictive precision, provided the model architecture handles temporal sequences effectively without suffering optimization stagnation.
Earlier data-driven prognostic investigations utilizing accelerated aging datasets from the National Aeronautics and Space Administration Prognostics Center of Excellence encountered persistent methodological limitations. Studies employing feedforward neural networks relied solely on collector-emitter voltage as an individual precursor, yielding a fitness score of 0.74 and an average accuracy of 60.42 percent, which failed to reflect multivariate degradation dynamics. Other research examined non-recurrent machine learning algorithms, including random forest ensembles, extreme learning machines, and backpropagation neural networks, which produced high root mean squared errors between 560.1380 and 1103.0374 because their structures lacked attention mechanisms or recurrence to model sequential dependencies. Subsequent efforts introduced sequence models such as the Informer transformer and hybrid convolutional networks, but these implementations focused exclusively on transient turn-off voltage peaks or deployed full encoder-decoder structures. Full encoder-decoder designs introduced excessive architectural complexity and heightened risks of overfitting when trained on compact accelerated aging datasets.
To establish an empirical basis for transistor degradation analysis, researchers examined accelerated aging data from the National Aeronautics and Space Administration Prognostics Center of Excellence. The experiments evaluated IRG4BC30K insulated gate bipolar transistors, accelerating operational aging by maintaining package temperatures well above the rated junction temperature limit of 150 degrees Celsius. In the hardware configuration, the transistor emitter was grounded while the collector connected to the power supply in series with a resistor, with the gate driven by a square-wave signal boosted through a high-speed amplifier. Data acquisition utilized high-speed oscilloscopes and standard National Instruments hardware as supply voltage was increased in steps of 0.5 volts. Under these severe overstress conditions, testing continued until thermal runaway and latching failures occurred, producing operational lifetimes of 2501 seconds for device two, 2267 seconds for device three, 2460 seconds for device four, and 2360 seconds for device five.
Data preparation focused on steady-state measurements, bypassing transient recordings because switching speed limitations prevented consistent capture of transient signals during testing. The steady-state records comprised 10,978 sub-structures per device, tracking supply voltage, node voltages, package temperature, heatsink temperature, internal temperature, collector current, and ambient temperature. Preprocessing in Python with Pandas identified that ambient temperature was entirely null, prompting its removal alongside an initial dummy date row. Elapsed time in seconds was computed from the initial date, enabling remaining useful life to be engineered as total failure time minus elapsed operational seconds. Collector current, heatsink temperature, and internal temperature were removed due to discrepancies across test devices, while time epoch values were dropped prior to normalization to eliminate outliers. Correlation heatmaps showed that remaining useful life exhibited an exact negative correlation of minus one with elapsed time, while package temperature showed a negative correlation of minus 0.47 with remaining life.
To overcome the architectural shortcomings of recurrent networks and full sequence-to-sequence transformers, researchers developed a specialized encoder-only transformer for multivariate time-series forecasting. Rather than relying on natural language tokenization, the design uses a time-distributed dense layer that projects six selected features—supply voltage, two node voltages, package temperature, elapsed time, and remaining useful life—into a representation dimension of sixty-four. Sinusoidal positional encodings, calculated via alternating sine and cosine expressions, were added directly to input projections to preserve temporal ordering without recurrent connections. The encoder incorporates a multi-head attention layer configured with four attention heads, each possessing a key dimension of sixteen. Each head calculates scaled dot-product attention through separate projection matrices across sliding input windows of sixty timesteps, enabling the model to process distinct representation subspaces simultaneously and capture long-term temporal dependencies across the aging sequence.
Following multi-head self-attention, the representations pass through layer normalization with an epsilon of one millionth, incorporating residual connections that sum positional encoding inputs with attention outputs. The representations then enter a position-wise feedforward network comprising two linear transformations with rectified linear unit activation, configured with an inner hidden dimension of 128 and input-output dimensions of 64. Dropout regularization of 0.2 was applied after both attention and feedforward layers to mitigate overfitting on compact experimental data. Rather than using autoregressive decoders or extracting only the final timestep, global average pooling aggregates the sequence output before a final dense layer produces predictions. Experiments were conducted in TensorFlow Keras within a Jupyter notebook on an AMD processor with 7.46 gigabytes of usable memory under Windows 10, using the Adam optimizer, mean squared error loss, batch sizes of 64, maximum limits of 100 epochs, and early stopping patience of 10 epochs.
Training evaluations on device two demonstrated significant performance advantages for the encoder-only transformer over standard recurrent baselines, including simple recurrent neural networks, gated recurrent units, and long short-term memory networks sharing identical dropout rates and input windows. The encoder-only transformer attained an initial mean squared error loss of 0.0618, lower than the simple recurrent network at 0.2426, the gated recurrent unit at 0.1050, and the long short-term memory model at 0.0750 due to parallelized multi-head sequence processing. The transformer converged to a minimal final training loss of 0.000304 at epoch 40, showing rapid optimization and robust pattern capture. In comparison, the gated recurrent unit reached a final training loss of 0.00154 at epoch 69, while the long short-term memory model plateaued after fifteen epochs, finishing at a suboptimal loss of 0.004924. The simple recurrent network required 36 epochs to achieve a final loss of 0.002325, hindered by vanishing gradients that impaired convergence.
Cross-validation on unseen devices three, four, and five established the superior generalization accuracy of the encoder-only transformer over recurrent alternatives and ablation models. When predicting remaining useful life in seconds, the proposed transformer recorded root mean squared errors of 53.81 on device three, 14.90 on device four, and 37.76 on device five, with corresponding mean absolute errors of 51.00, 11.33, and 34.45, and high determination coefficients of 0.9932, 0.9996, and 0.9969. Conversely, the simple recurrent network produced root mean squared errors of 156.14, 105.14, and 62.34; the gated recurrent unit yielded 166.32, 78.84, and 111.69; and the long short-term memory network recorded 139.89, 92.78, and 114.92. Ablation analysis confirmed that altering key components degraded accuracy: utilizing the final timestep output yielded a device three mean squared error of 10,679.90, removing positional encoding increased final training loss to 0.000445, stacking two encoder layers raised device three error to 16,059.07, expanding to eight heads produced 12,129.65 error, and enlarging embedding dimensions to 128 caused overfitting with 5,396.17 test error.
Although prediction curves verified that the encoder-only transformer closely tracked actual component degradation without late-stage instability, the authors highlighted clear operational limitations in the present study. Accelerated aging was performed in laboratory test chambers under continuous thermal and electrical overstress, conditions that do not capture the fluctuating environmental and mechanical stresses found in commercial operations. Additionally, individual device lifespans ranged between 2000 and 2500 seconds, representing relatively short operational histories compared to industrial monitoring datasets. Future research plans include pairing the transformer encoder with physics-of-failure equations such as Arrhenius and Coffin-Manson models to form hybrid prognostic frameworks, while using transfer learning across larger datasets. Further efforts aim to evaluate the framework against real-world operational records from solar inverters, wind turbines, and electric vehicles, exploring model quantization and pruning for edge deployment and extending health monitoring to power capacitors and field-effect transistors.
Subject of Research: Remaining useful life prediction of insulated gate bipolar transistors using transformer neural networks
Article Title: Development of encoder-multivariate time-series transformer for remaining useful life prediction of IGBTs
Article References: Annan, K. A., Effah, F. B., & Frimpong, E. A. (2026). Development of encoder-multivariate time-series transformer for remaining useful life prediction of IGBTs. Journal of Electrical Systems and Information Technology, 13(1), Article 68. https://doi.org/10.1186/s43067-026-00368-4
Image Credits: AI Generated
DOI: 10.1186/s43067-026-00368-4
Keywords: insulated gate bipolar transistors, remaining useful life, encoder-only transformer, power electronic converters, multivariate time series, steady-state precursors, accelerated aging, multi-head attention, predictive maintenance, prognostics and health management, thermal overstress, root mean squared error, global average pooling, recurrent neural networks, package temperature, sinusoidal positional encoding, NASA PCoE dataset
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Encoder-Transformer Predicts Power Transistor Useful Life Using Steady-State Precursor Data. Scienmag. https://scienmag.com/encoder-transformer-predicts-power-transistor-useful-life-using-steady-state-precursor-data/
Cassandra Pierce. "Encoder-Transformer Predicts Power Transistor Useful Life Using Steady-State Precursor Data." Scienmag, 21 November 2026, https://scienmag.com/encoder-transformer-predicts-power-transistor-useful-life-using-steady-state-precursor-data/. Accessed 21 November 2026.
Cassandra Pierce. "Encoder-Transformer Predicts Power Transistor Useful Life Using Steady-State Precursor Data." Scienmag. November 21, 2026. https://scienmag.com/encoder-transformer-predicts-power-transistor-useful-life-using-steady-state-precursor-data/








