Non-intrusive appliance health monitoring and remaining useful life prediction using a DRN-transformer model

3.1. System overview

This section introduces the technical architecture and methodologies employed in the proposed smart metering system, including the hardware design based on STM32H7, signal preprocessing techniques, and the DRN-Transformer deep learning model. Fig. 1 System architecture diagram.

Fig. 1System architecture diagram

3.2. Hardware architecture

The proposed system is built around the STM32H7 microcontroller unit (MCU), chosen for its dual-core 480 MHz processor, integrated high-speed ADCs, and low-power consumption. The MCU interfaces with wideband current and voltage sensors capable of capturing transient signals and harmonic distortions commonly associated with appliance wear and malfunction. The analog signals are sampled at 1 MHz, providing the temporal resolution needed to identify high-frequency signal events, such as motor start-up currents or rapid switching harmonics [21].

Fig. 2 shows the hardware architecture of the system, illustrating the key components and their interconnections, including the power module, voltage and current acquisition modules, and the main control module.

Additional components include an SPI buffer for real-time data transfer, SD card storage for local data logging, and a low-noise power module to reduce electrical interference. The embedded firmware is optimized for synchronous multi-channel sampling and supports direct memory access (DMA) for efficient data handling. This setup ensures minimal latency in signal acquisition while maintaining long-term operational stability.

Fig. 2Hardware circuit diagram

The hardware system design fully leverages the high performance and rich interfaces of the STM32H7, combined with high-precision sensors and filter circuits, to achieve efficient and accurate acquisition of home appliance load signals. Multi-layer PCB design and protection circuits ensure the stability and safety of the system, providing a solid hardware foundation for subsequent health status assessment and lifespan prediction based on the DRN-Transformer model. Environmental parameters (temperature, humidity) are collected via independent sensor modules, with data synchronously acquired alongside electrical parameters. These serve as auxiliary features for the model, enhancing its adaptability to environmental influences. Fault records are annotated by professional technicians in conjunction with equipment maintenance records and user feedback, ensuring the accuracy of data labels.

Although the overall system architecture is software-centric, this work explicitly considers hardware-driven constraints for microcontroller deployment. Figure X illustrates the complete hardware architecture, highlighting the sensor front-end, ADC, STM32H7 MCU, on-chip memory, and communication modules. The STM32H7 is selected due to its Cortex-M7 core operating up to 480 MHz, large on-chip SRAM, TCM support, and high ADC throughput, which collectively enable real-time signal acquisition and edge intelligence under strict latency and power constraints.

To adapt the DRN-Transformer model to STM32H7 constraints, a co-design strategy combining lightweight architecture and quantization optimization is adopted. The number of DRN residual blocks is reduced from eight to four, and the Transformer attention heads are limited to two, reducing computational nodes by 37.5 %. Consequently, inference complexity decreases from 1.2×10⁶ FLOPs to 3.5×10⁵ FLOPs. Furthermore, model parameters are quantized from 32-bit floating point to 16-bit fixed-point representation, reducing storage from 1.8 MB to 450 KB, enabling full deployment within the STM32H7 on-chip SRAM without external memory access latency. In addition, the embedded system is designed with energy efficiency considerations. The STM32H7 operates under optimized clock scheduling and DMA-based data transfer, which reduces CPU overhead during signal acquisition and preprocessing. Experimental measurements show that the overall system maintains low power consumption during continuous monitoring, demonstrating its suitability for long-term smart metering deployment.

Fig. 3Hardware Physical Diagram. Photo by Xihuan Su, in the electrical engineering training room of Southern Vocational College in Guangdong, China, on June 29, 2025

Fig. 3 illustrates the physical implementation of the proposed hardware system. The system was independently designed, developed, and assembled by the author. The figure clearly shows the layout and interconnection of the power supply module, voltage and current acquisition module, voltage and current conversion circuit, main control module, and capacitive touch display screen. This figure intuitively demonstrates the compact embedded system based on the STM32H7 microcontroller, verifying the feasibility and stability of the designed acquisition and display system.

The materials and methods employed in this study fully integrate high-performance hardware platforms, advanced software tools, and extensive real-world data to construct a comprehensive non-invasive load monitoring system, providing a robust technical foundation for health status assessment and lifespan prediction of home appliances.

3.3. Signal preprocessing

Raw voltage and current signals acquired from the sensors are first interpolated to handle minor gaps caused by transient signal disruptions. Outlier detection is performed using a dynamic threshold-based filter to eliminate spikes and sensor noise. The cleaned signal is then normalized to a common scale to facilitate model convergence.

Fig. 4 shows the signal preprocessing steps and the transformations applied to the raw signals. To extract meaningful features, both Fast Fourier Transform (FFT) and Short-Time Fourier Transform (STFT) are applied. FFT captures the global frequency characteristics of the signal, while STFT enables localized spectral analysis over sliding windows. These transformations convert the time-domain waveforms into spectrogram-like inputs that highlight key characteristics, such as harmonic energy distribution and transient amplitude modulations, which are indicative of appliance health status.

Fig. 4Signal preprocessing

3.4. DRN-transformer model

The DRN-Transformer model proposed in this paper is a lightweight hybrid deep learning model co-designed with the hardware constraints of the STM32H7 embedded micro-system, and the model design takes the limited on-chip memory (1 MB SRAM) and computing power (480 MHz dual-core) as the core constraints, with the goal of realizing real-time inference on the embedded platform while ensuring the prediction performance of health classification and RUL prediction. The deep learning model used in this work is a hybrid of Deep Residual Networks (DRN) and Transformer architecture. The DRN component is composed of multiple stacked residual blocks that focus on extracting robust local features from the time-frequency input data. These layers help in preserving the gradient flow and improve learning stability, especially in deeper networks.

Following the DRN layers, the Transformer module is introduced to capture long-range dependencies across temporal sequences. It utilizes a self-attention mechanism to weigh the importance of different time steps, enabling the model to focus on crucial degradation patterns that evolve over time. Positional encoding is included to maintain the temporal order of input signals.

The DRN-Transformer model cleverly combines the Deep Residual Network (DRN) with the Transformer's self-attention mechanism, combining deep feature extraction capabilities with the ability to capture long-range temporal dependencies. Its structural design effectively alleviates the vanishing and exploding gradient problems common in deep network training, enabling the model to construct deeper feature representations. The Transformer module, through its multi-head self-attention mechanism, dynamically calculates the correlation between each time step in the input sequence, greatly enhancing the model's ability to perceive complex load changes.

The deep residual network (DRN) part of the proposed hybrid model focuses on extracting local spatial-temporal features from the time–frequency spectrograms. Each residual block learns a residual mapping rather than a direct transformation. Formally, the residual learning process can be expressed as Eq. (1):

where $X_l$ and $W_l$ denote the input and output of the $l$-th residual block, respectively, and $F\left(X_l,W_l\right)$ represents the nonlinear transformation composed of convolution, batch normalization, and ReLU activation. The skip connection $X_l$ ensures that gradient information can directly propagate to shallow layers, effectively mitigating vanishing gradient and degradation problems in deep architectures.

Fig. 5 illustrates the internal structure of a typical residual block in the DRN. The residual path (left branch) performs the nonlinear transformation $F\left(x\right)$ ,while the identity path (right branch) directly passes the input $x$ to the output. The addition operation $F\left(x\right)+x$ enables efficient information flow and stabilizes gradient propagation.

Fig. 5Structure of a typical residual block in the deep residual network (DRN)

Fig. 6Structure of the transformer module in the DRN-transformer model

Fig. 6 illustrates the structure of the Transformer module used in the proposed DRN-Transformer model. It consists of an encoder-decoder architecture where each layer employs multi-head attention, feed-forward networks, and residual normalization. The positional encoding preserves the temporal order of the time-series input, while the self-attention mechanism enables the model to capture long-range dependencies among signal segments. The decoder outputs are processed through a linear and softmax layer for appliance health classification and remaining useful life (RUL) prediction.

The Transformer module, following the DRN, captures long-range dependencies within the temporal sequences via a self-attention mechanism. The scaled dot-product attention is defined as Eq. (2):

where $Q$, $K$, and $V$ are the query, key, and value matrices derived from the input sequence, and $d_k$ is the key dimension used for normalization. This formulation allows the model to dynamically assign higher attention weights to signal segments that exhibit stronger degradation patterns.

To enhance feature diversity and model expressiveness, multiple attention heads are employed to compute attention in parallel, defined as Eq. (3):

with each attention head given by Eq. (4):

where $hea{d}_i$ is the number of heads, and $W_i^Q$, $W_i^K$, $W_i^V$, $W^o$ are learnable projection matrices.

The final DRN-Transformer output integrates local representations $F\left(X_l\right)$ from the residual network and global dependencies $Attention(Q,K,V)$ from the Transformer through feature fusion. This hybrid structure enables simultaneous modeling of short-term transients and long-term degradation trends in appliance load signals, which is critical for accurate health classification and remaining useful life (RUL) prediction.

The model supports a multi-task learning paradigm, with two output heads: one for health state classification (e.g., healthy, degraded, fault) using a softmax layer, and the other for remaining useful life (RUL) prediction via a linear regression head. The combined loss function includes categorical cross-entropy for classification and mean squared error (MSE) for RUL estimation, encouraging the model to learn shared representations that benefit both tasks.

4.1. Dataset description

The dataset used in this study comes from a non-intrusive load monitoring experimental platform. This platform continuously collects operating data from four common household appliances: a refrigerator, a washing machine, an air conditioner and a microwave oven. Each appliance was monitored over a long period of time under both normal and degraded operating conditions to reflect the energy consumption characteristics of the device under different health conditions. The experimental system uses an STM32H7 microcontroller as the data acquisition core of the smart meter system. High-precision voltage and current sensors are used to sample electrical signals at a frequency of 1 MHz to capture subtle transient fluctuations. All four appliance types are included in the dataset and share the same labeling and evaluation criteria. Although the complete dataset contains four appliance types, some comparative experiments report results on four representative appliances for consistency with baseline methods and clarity of presentation. It should be noted that this study focuses on four household appliances – refrigerator, washing machine, air conditioner, and microwave oven.

The acquisition system transmits signals to a local storage module via the SPI bus. During the acquisition process, filtering and anti-interference processing are performed to ensure signal stability and data quality. Experimental data is calibrated and annotated weekly by laboratory technicians, based on equipment maintenance records, sensor calibration results, and operation logs.

To verify the performance of the STM32 data acquisition module, this article systematically evaluated five key aspects: sampling accuracy, sampling frequency, data stability, power consumption, and interference immunity. The STM32 integrated 12-bit ADC offers 4096 levels of resolution. Experimental measurements show that current measurement error is within ±0.5 %, and voltage measurement error is within ±0.3 %. Table 1 shows the measurement error statistics under different load conditions. As shown in Table 1, the STM32-based data acquisition module maintains current measurement errors within ±0.5 % and voltage measurement errors within ±0.3 % across different load conditions. The low standard deviation values indicate stable and repeatable measurement performance, which is sufficient for high-frequency NILM-based health assessment and RUL prediction.

Each sample consists of a synchronized voltage-current signal window, which is converted into a spectrogram using FFT and STFT techniques to extract time-frequency features. Health labels are defined based on equipment operating status monitoring and expert evaluation criteria. They are manually labeled by analyzing characteristic parameters such as current waveform distortion, harmonic energy distribution, and power factor, combined with equipment maintenance records and operation logs.

Table 2 presents the system module processing time statistics. The remaining useful life (RUL) value is calculated using a degradation trend analysis method. Specifically, the device's performance degradation rate (such as effective power drop and starting current increase) over multiple operating cycles, combined with the accumulated operating time, is used to estimate the remaining time until failure or significant performance degradation. The final labeling results are cross-verified by two independent technicians to ensure label consistency and reliability. All data preprocessing and feature extraction steps were implemented on the STM32H7 MCU, and the processing time of each module of the embedded micro-system was tested in actual operation, as shown in Table 2, which fully reflects the real-time performance of the system at the hardware level.

4.2. Evaluation metrics

To comprehensively evaluate the performance of the system, this study uses the following quantitative indicators to analyze the system from two aspects: health status classification and remaining useful life (RUL) prediction:

Classification Accuracy (Eq. (5)). This indicates the percentage of appliances whose health status is correctly identified by the model. It is a basic indicator for measuring overall classification performance:

Among them, TP and TN represent the number of correctly classified healthy and non-healthy samples, respectively.

F1-Score (Eq. (6)). It is the harmonic average of precision and recall, and is used to measure the classification robustness of the model under class imbalance:

Mean Absolute Error (MAE) (Eq. (7)) used to evaluate the accuracy of RUL prediction, reflecting the average deviation between the predicted value and the actual value:

Latency, the end-to-end time delay from signal acquisition to final prediction output, is a key metric for measuring a system’s real-time performance. In this study, the overall system response time was kept within 30 milliseconds, meeting the requirements of embedded real-time monitoring.

Power Consumption: This value represents the average power consumption of the system during continuous operation, measured in watts (W). This value is measured by the onboard power monitoring module. Low power consumption demonstrates the system's energy-saving advantages in smart home scenarios.

In summary, these indicators comprehensively reflect the performance of the system from four aspects: classification accuracy, prediction accuracy, real-time performance, and energy consumption, providing a reference for model optimization and deployment.

Distributed deployment of the STM32H7-based DRN-Transformer smart metering system outperforms the centralized architecture with an 8-node setup reducing average end-to-end latency by 33.6 % to 14.8 ms and single-node CPU load by 43.6 % to 25.4 %, while improving health classification accuracy to 94.3 % and lowering RUL prediction MAE to 2.6 months; the moderate power consumption increase is a reasonable trade-off for the enhanced real-time performance, prediction accuracy and scalability, making the system well-suited for large-scale smart home and industrial micro-equipment monitoring scenarios, and the modular DRN-Transformer model shows high compatibility with distributed embedded parallel inference.

4.3. Baseline models

To validate the effectiveness of the proposed DRN-Transformer model, comparisons were made against several baseline models:

– Long Short-Term Memory (LSTM).

– Bidirectional LSTM (BiLSTM).

– Autoregressive Integrated Moving Average (ARIMA).

– Support Vector Machine (SVM) with handcrafted features.

Each baseline was trained using the same preprocessed input features to ensure fair comparison.

4.4. Results and analysis

The baseline models include SVM, RF, LSTM, Transformer, and LSTM-VMD. All baseline models are trained and evaluated using the same dataset and preprocessing pipeline to ensure fair comparison. To further validate the effectiveness of the proposed DRN-Transformer model for appliance life prediction, a comparative experiment was conducted using a baseline model as a reference. The comparison targets included classic machine learning models (SVM, RF), deep learning models (LSTM, Transformer, LSTM-VMD), and their improved models (DRN, DRN-Transformer) Fig. 4Comparison of Model Accuracy Across Different Epochs.

As shown in Fig. 7, the DRN-Transformer model maintains the highest prediction accuracy across all lifespan stages, demonstrating its significant advantage in capturing both short-term fluctuations and long-term degradation trends. Theoretically, this performance improvement is attributed to the complementary nature of the model structure.

Fig. 7Comparison of model accuracy across different epochs

DRN (Deep Residual Network) effectively alleviates the gradient vanishing problem through the residual connection mechanism (Residual Connections), enhances the local feature learning ability, and enables the model to more stably extract fine-grained changes in voltage and current signals.

The transformer module captures long-range temporal dependencies through a lightweight self-attention mechanism, enabling efficient feature aggregation while maintaining low computational complexity suitable for embedded deployment.

The combination of the two, with the DRN responsible for local feature enhancement and the Transformer responsible for global temporal modeling, forms a hierarchical feature fusion architecture. This design enables the model to demonstrate greater robustness and generalization capabilities when dealing with nonlinear and multi-scale degraded features.

In contrast, models using only Transformers or DRNs, while able to capture some features, struggle to simultaneously account for both local details and global trends. Traditional models (SVM and RF) lack the ability to model temporal dependencies, making them particularly underperformers in early prediction stages. While LSTMs have some advantages in temporal modeling, they are prone to gradient decay when learning long sequences. The baseline model performed the worst across all metrics.

Overall, the DRN-Transformer model outperforms other methods in terms of accuracy, convergence speed, and stability, proving its theoretical rationality and application potential in health management and lifespan prediction of smart home appliances.

Fig. 8Accuracy convergence curves of different household appliances

To evaluate the generalization capability of the proposed DRN-Transformer model across different appliances, four common device types were selected: refrigerator, washing machine, air conditioner and microwave oven. As shown in Fig. 8, all devices exhibit a consistent performance trend – rapid improvement followed by stability – indicating the model’s ability to capture diverse degradation patterns. Refrigerators and washing machines achieved the highest prediction accuracy, likely due to their stable operation and clearer fault modes. Air conditioners and microwaves performed slightly lower, possibly due to usage variability.

To fully verify the performance of the proposed STM32H7-based embedded micro-sensing system, the core hardware performance metrics (CPU load, memory usage, power consumption, end-to-end latency) were tested in actual continuous operation, and compared with the mainstream embedded NILM system baselines (STM32F4, Arduino, Raspberry Pi) under the same test conditions (1 MHz sampling rate, real-time preprocessing + model inference, continuous operation for 72 hours). The test results are shown in Table 4, which fully reflects the superiority of the proposed system in hardware performance.

During the model training phase, a variety of appliance load datasets were employed, covering typical appliances such as refrigerators, air conditioners, washing machines, and microwave ovens. The data sampling frequency was 1 kHz, with a data collection time span exceeding 30 days. During training, the model extracted local temporal features through multi-layer residual modules, then utilised the self-attention mechanism of the Transformer to capture global temporal dependencies, ultimately outputting classification results for appliance operational states and regression predictions for remaining lifespan. Model performance evaluation was conducted using four metrics: accuracy, root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination ($R^2$). The specific results are shown in Table 5. As can be seen from the data in the table, the DRN-Transformer model demonstrated high accuracy in home appliance health status classification, all exceeding 92 %, with the highest accuracy for washing machines reaching 97.8 %. In terms of lifespan prediction, the model’s RMSE and MAE remain at low levels, with $R^2$ values exceeding 0.8, indicating that the model has strong explanatory power and stability in predicting the lifespan of appliances. Table 5 presents the performance metrics of the DRN-Transformer model across different appliance devices.

The health assessment and lifespan prediction module is a core component of this system. The accuracy and stability of its performance directly impact the overall performance of the smart meter system. To verify the module’s effectiveness, a series of experiments were designed covering typical household appliances such as air conditioners, refrigerators, washing machines, and microwave ovens. The experiments collected operational data and performed health assessments and lifespan predictions on these appliances.

In theory, this module is built on degradation modeling theory and a deep feature learning mechanism. The current and voltage signals collected by the STM32 are preprocessed and then fed into the DRN-Transformer model. The deep residual network (DRN) within the model extracts local features at short time scales and mitigates the vanishing gradient problem, while the Transformer self-attention mechanism captures signal correlations over long time scales, enabling adaptive fusion of multi-scale degradation features. This structure theoretically allows the model to balance sensitivity to local fluctuations with robustness to global degradation trends.

The health status assessment module uses a feature fusion strategy to calculate key indicators such as load fluctuation rate, abnormal frequency, and energy efficiency factor. It also achieves a comprehensive assessment of the equipment health status by integrating the time-frequency distribution of the multi-dimensional feature space. The system divides the health status into five levels: excellent, good, fair, poor, and dangerous. The life prediction module is based on degradation trend analysis. By fitting the decay curve of performance indicators in the time series, it calculates the remaining useful life (RUL) based on the degradation rate:

where $T_{fail}$ is the time when the device is expected to reach the fault threshold, and $T_{current}$ is the current running time. Table 6 shows the accuracy and recall of different equipment health status assessments. The RMSE and MAPE indicators for different equipment life predictions demonstrate that the model performs well in life prediction, with an average RMSE of 3.2 months and an average MAPE of 8.5 %, meeting practical application requirements.

4.5. Ablation studies

To validate the role of each component in the DRN-Transformer model and its impact on overall performance, this study conducted ablation experiments. By removing the residual connection module (DRN), the Transformer self-attention module, and the FFT time-frequency feature input, the prediction performance of different model structures was compared.

The experimental results show that when the DRN module is removed, the model convergence speed during training is significantly reduced, and the health status classification accuracy drops by 3.8 %, indicating that the residual structure is crucial for alleviating gradient vanishing and improving local feature extraction. After removing the Transformer module, the model’s RMSE in the lifespan prediction task increases by approximately 15 %, demonstrating that the self-attention mechanism plays a key role in capturing the long-term temporal dependencies of device degradation. Removing the FFT feature input reduces the model’s ability to recognize high-frequency fluctuations, indicating the importance of time-frequency information for early anomaly detection.

Overall, the ablation experiments validate the rationale of the model design from both theoretical and experimental perspectives. The DRN module strengthens local feature learning, while the Transformer module implements global dependency modeling. The combination of the two enables multi-scale feature fusion, providing a solid theoretical and practical basis for health prediction of smart home appliances.

During model training, a multi-task joint loss function is used to optimize both health status classification and lifespan prediction performance. The overall loss is defined as (Eq. (9)):

where $L_{cls}$ and $L_{RUL}$ denote classification and RUL regression losses, respectively, and $\alpha$, $\beta$ balance the two objectives, which evaluates the difference between the predicted lifespan and the actual value. Parameters $\alpha$ and $\beta$ are used to balance the importance of the two tasks. The total loss is defined as $L_{total}=\alpha L_{cls}+\beta L_{RUL}$ where $\alpha =$0.6 and $\beta =$ 0.4 are determined via cross-validation to balance classification and regression objectives. The classification task provides prior structural information for RUL estimation (e.g., appliances in a dangerous state typically correspond to near-zero RUL). The additive formulation ensures simultaneous minimization of both tasks and follows the shared-representation optimization principle of multi-task learning. Ablation results show that increasing $\beta$ beyond 0.6 causes classification accuracy to drop by more than 5 %, validating the rationality of the adopted weighting strategy.

As shown in the figure, the training loss decreases rapidly and stabilizes within the 200 epochs, indicating that the model training has converged well. The smooth decline of the loss function demonstrates the stability of the parameter optimization process and verifies the model's effectiveness in multi-task learning. This loss function design theoretically ensures the coordinated optimization of the classification and prediction tasks, helping to improve the model's generalization performance and robustness.

Fig. 9 illustrates the DRN-Transformer’s training loss over 200 epochs. Specifically, the loss becomes relatively stable within the first 50 epochs, while full convergence is achieved around epoch 120, as shown in Fig. 9. The model shows a rapid decrease in loss during the first 50 epochs and reaches convergence around epoch 120, with no sign of overfitting or instability. This indicates strong trainability and stable learning behavior in time series modeling. Structural optimizations – such as tuning the number of residual blocks in DRN and adjusting multi-head attention in the Transformer – led to improved feature extraction. As shown in Table 4, these adjustments significantly enhanced the model’s prediction accuracy and robustness for appliance lifetime forecasting.

Fig. 9Training loss curve of the DRN-transformer model

Fig. 10Accuracy curves of different models under various learning rates (Lr= 0.1, 0.001, 0.05, 0.01)

a)$Lr=$0.1

b)$Lr=$0.001

c)$Lr=$0.05

d)$Lr=$0.01

During preprocessing, adaptive noise filtering and anomaly detection were applied to purify input signals, while sliding window statistics and wavelet denoising were used to suppress environmental noise and transient outliers. As shown in Fig. 4, the fluctuation rate of the processed data significantly decreased – from 7-11 % to a stable range of 2-3 % – enhancing signal stability and continuity. Moreover, learning rate tuning revealed that a value of 0.01 yields the best model performance, ensuring optimal training convergence and prediction accuracy.

The system was tested on appliances from different brands not seen during training. Using a transfer learning approach with minimal retraining, the DRN-Transformer maintained an error margin below 7 %, compared to over 37 % in non-adapted models. This confirms the model’s robustness and adaptability to varied appliance types and manufacturers. Additional robustness evaluations were conducted by introducing noise perturbations, varying learning rates, and testing cross-brand appliance data. The results indicate stable performance and strong robustness under realistic operating conditions. Overall, the experimental findings demonstrate the practicality and effectiveness of the proposed NILM system for real-time appliance monitoring, predictive maintenance, and cross-domain deployment.

Compared with previously reported embedded NILM implementations, the proposed system demonstrates improved real-time performance and model efficiency. While many embedded solutions rely on simplified statistical models due to hardware limitations, the proposed DRN-Transformer framework maintains high prediction accuracy while remaining deployable on a microcontroller platform. This highlights the effectiveness of the proposed hardware-algorithm co-design strategy for intelligent edge sensing applications.