An interpretable hybrid adaptive framework combining empirical mode decomposition and deep learning for real-time nonstationary signal analysis in smart manufacturing

The increasing adoption of smart manufacturing systems under the Industry 4.0 paradigm has transformed modern production environments into complex cyber-physical ecosystems. These systems are equipped with a multitude of sensors that continuously monitor processes to ensure quality, efficiency, and predictive maintenance. However, the vast amount of data generated in such settings presents challenges for traditional signal processing methods, especially when dealing with nonstationary and nonlinear characteristics inherent in real-world industrial signals. Variations in sensor readings due to tool wear, machine drift, thermal effects, and load changes lead to time-varying patterns that require adaptive analytical techniques capable of capturing evolving dynamics in real time.

Conventional data analysis frameworks, such as those based on fixed-basis transforms like Fourier or Wavelet, rely on assumptions of linearity and stationarity. Although effective in controlled environments, they often fail to generalize amid signal transients, operational mode shifts, and contextual dependencies. Furthermore, as manufacturing processes grow more automated and interconnected, there is a pressing demand for systems that not only analyze signals adaptively but also provide explainable outputs to facilitate decision-making in safety-critical applications.

To address these limitations, this study presents a hybrid framework combining Empirical Mode Decomposition (EMD) with deep learning techniques to enable real-time analysis of complex signals. EMD, as an adaptive signal decomposition method, provides a basis-free mechanism to extract intrinsic oscillatory modes from raw sensor data, making it particularly suitable for handling nonstationary signals (Adali and Haykin, 2010). Coupled with a hybrid deep learning pipeline comprising Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and an attention layer, the proposed framework captures spatiotemporal features while preserving interpretability and flexibility.

Despite the increasing deployment of adaptive algorithms in industry, a unified approach that combines time-domain signal decomposition, deep feature extraction, and interpretability is lacking. Many existing methods either rely on static preprocessing (e.g. FFT, STFT) that imposes limiting assumptions on the signal structure or employ deep learning models that function as “black boxes,” making them unsuitable for real-time and safety-critical deployment.

Moreover, nonstationary signals in manufacturing are often influenced by a combination of deterministic and stochastic factors—thermal expansion, material fatigue, machine vibrations, among others—which makes modeling and fault detection especially challenging. The inability to extract meaningful temporal features in real time from such signals compromises the efficacy of predictive maintenance and anomaly detection systems.

In contrast to pure end-to-end deep learning methods, which process raw signals directly and often struggle with noise robustness (e.g. leading to 5–10% accuracy drops in high-noise industrial scenarios due to overfitting to transients), our hybrid approach leverages EMD as an adaptive pre-filtering step to decompose signals into noise-separated IMFs. This enhances robustness by isolating intrinsic modes, improving feature extraction under variable conditions and reducing sensitivity to stochastic noise by up to 15% in our experiments.

There is thus a need for a signal processing framework that can:

1. Adapt to dynamically changing signals without reliance on fixed basis functions,

2. Extract hierarchical features from sensor data streams,

3. Provide transparent and interpretable results to engineers and operators.

The main goal of this study is to develop and validate an interpretable hybrid adaptive framework for real-time signal analysis in smart manufacturing. Specifically, the research aims to:

1. Propose a novel combination of EMD and deep learning (CNN + LSTM + Attention) tailored for nonstationary signal streams.

2. Design an end-to-end architecture that supports real-time analysis on multivariate sensor inputs.

3. Evaluate the model on real-world smart manufacturing datasets to benchmark its accuracy, latency, and interpretability.

4. Demonstrate the system’s adaptability to shifts in operating conditions and signal characteristics over time.

The significance of this work lies in its ability to bridge the gap between advanced signal decomposition techniques and modern AI-based analysis, with a focus on practical deployment in industrial systems. The hybrid architecture ensures:

1. Adaptability: Through EMD, the system can naturally handle signal variation without requiring reconfiguration or retraining.

2. Predictive Performance: Deep neural components enhance the ability to learn complex temporal-spatial patterns.

3. Explainability: Attention mechanisms offer traceable decision paths, a crucial feature for safety-critical environments like aerospace manufacturing or high-speed automation.

4. Scalability: The framework can be extended to edge-AI deployment scenarios due to its modular and lightweight design.

This research aligns with the current industrial need for intelligent monitoring solutions that are not only accurate but also interpretable and robust to operational variability. Additionally, it contributes methodologically by integrating techniques from adaptive signal processing, deep learning, and explainable AI—fields that are often studied in isolation.

The scope of this research is centered on the development and evaluation of the proposed adaptive hybrid framework in the context of smart manufacturing, particularly focusing on multivariate sensor signals such as vibration, current, acoustic emissions, and thermal readings. The model is evaluated using publicly available and proprietary datasets collected from CNC machinery operating under variable load and speed conditions.

However, the current study has certain limitations:

1. The framework assumes sensor synchronization and does not explicitly address asynchronous data streams.

2. Interpretability is provided primarily through attention visualization; deeper causal interpretability remains an open challenge.

3. The study does not extend to wireless sensor networks or IoT-based streaming frameworks but sets the foundation for such extensions.

Smart manufacturing environments are characterized by high-frequency, real-time data acquisition systems that capture complex phenomena such as tool wear, vibration, temperature fluctuation, and acoustic emissions. These signals frequently exhibit nonlinear and nonstationary characteristics arising from varying operational modes, material heterogeneity, and environmental noise. Traditional techniques such as Fourier Transform (FT) and Short-Time Fourier Transform (STFT) offer limited temporal resolution and are often unsuitable for accurately capturing transient events (Adali and Haykin, 2010). Wavelet Transform (WT) provides some improvement but still relies on predefined basis functions, which limit adaptability to diverse signal patterns.

To address these limitations, adaptive signal analysis methods have gained prominence. One such approach, Empirical Mode Decomposition (EMD), was proposed by Huang et al. (1998) and has since become a foundational technique in nonlinear signal processing. EMD adaptively decomposes a signal into a finite number of Intrinsic Mode Functions (IMFs) without assuming any a priori basis. This feature makes EMD particularly suitable for analyzing real-world industrial signals, as demonstrated in fault detection of rotating machinery (Chen and Zhao, 2021), gear wear detection (Musbah et al., 2023), and tool condition monitoring (Dixit and Jain, 2023).

In smart manufacturing, applications of EMD have been enhanced by hybrid models that integrate signal decomposition with statistical or machine learning classifiers. However, the integration of EMD with deep learning—particularly in an interpretable and real-time framework—remains an underexplored area.

Over the years, several variants of EMD have been developed to address issues like mode mixing and sensitivity to noise. Ensemble Empirical Mode Decomposition (EEMD), for instance, adds white noise to the signal before decomposition and averages the results over multiple trials, improving robustness (Wu and Huang, 2009). More recently, Improved Ensemble Noise-reconstructed EMD (IENEMD) has been introduced to enhance signal-to-noise ratio in complex environments, such as bearing fault diagnosis in strong noise conditions (Yin et al., 2025).

Despite these advancements, most EMD-based methods are used in tandem with conventional classifiers like SVM, decision trees, or feedforward neural networks, with limited adaptability to evolving data streams. Moreover, they rarely account for the spatial-temporal structure in the decomposed IMFs, which deep learning models are well suited to exploit.

Deep learning techniques, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) such as Long Short-Term Memory (LSTM), have revolutionized time-series and signal analysis tasks (Sa-Ngasoongsong et al., 2015). CNNs excel at extracting local spatial features, while LSTMs are adept at modeling temporal dependencies, making them ideal for applications like fault detection, vibration analysis, and predictive maintenance (Chen et al., 2020; Ur Rehman and Mandic, 2018). Hybrid CNN-LSTM architectures have shown promise in multi-sensor environments, allowing models to learn representations that combine feature localization with temporal context (Cheng and Li, 2021).

However, most deep learning architectures are trained end-to-end on raw or statically preprocessed signals. This approach can be data-hungry, computationally intensive, and difficult to interpret. Moreover, they often assume stationarity within training batches, which may not reflect operational variability in manufacturing.

Recent research has begun to integrate signal preprocessing techniques like EMD with deep learning models. For instance, combining EMD with CNNs for fault detection tasks has been explored with success in enhancing model accuracy (Zhao and Chen, 2023). Yet these integrations often treat EMD as a static front-end rather than a dynamic, adaptive component of the pipeline, and rarely incorporate interpretability mechanisms.

Interpretability is critical in industrial AI systems where decisions must be transparent and justifiable, especially in safety-critical domains like aerospace or medical device manufacturing. Traditional deep models lack this capability due to their opaque, high-dimensional internal representations. To address this, attention mechanisms have been proposed as a means of visualizing model focus during inference (Montavon et al., 2018). These mechanisms assign weights to different parts of the input signal or feature maps, enabling a form of post hoc interpretability.

Recent advancements in self-attention, transformer models, and explainable AI (XAI) tools have been applied in domains such as healthcare (Holzinger et al., 2019), finance (Du et al., 2022), and fault diagnosis (Li and Zheng, 2021). However, their integration into real-time manufacturing signal analysis—especially within hybrid EMD-deep learning frameworks—has received little attention.

Some studies have attempted to visualize attention weights in multiscale CNN models after applying noise-resistant EMD, showing improved diagnostic reliability (Yin et al., 2025). Still, comprehensive integration of adaptive decomposition, interpretable learning, and real-time performance remain absent in the current literature.

Recent advancements in Graph Neural Networks (GNNs) have shown promise for modeling multi-sensor topologies in industrial signal analysis, particularly for handling asynchronous data via mechanisms like Graph Attention (GAT). For instance, DEST-GNN (Yang et al., 2025) introduces a double-explored spatio-temporal GNN for multi-site intra-hour PV power forecasting, leveraging graph structures to capture spatial dependencies and temporal dynamics in nonstationary time-series. Similarly, SDR-GNN (Fu et al., 2025) employs spectral domain reconstruction in GNNs for incomplete multimodal learning in conversational emotion recognition, addressing data asynchrony through graph-based fusion. Additionally, parallel multi-scale dynamic GNNs (Hou et al., 2025) have been proposed for multivariate time-series forecasting, using multi-layer graphs to model varying scales and adapt to nonstationary patterns. While these approaches excel in topology-aware processing, they often require predefined graph structures and may not inherently decompose nonstationary signals like EMD, highlighting a complementary role for hybrid EMD-GNN frameworks in future work.

The reviewed literature reveals three main gaps:

1. Lack of end-to-end frameworks that combine adaptive signal decomposition with deep learning in a way that preserves both performance and interpretability.

2. Limited exploration of real-time applications where models must adapt online to changes in signal distributions caused by tool wear, environmental variation, or workload shift.

3. Sparse research on applying attention-based interpretability within signal processing pipelines in industrial manufacturing settings.

This study addresses these gaps by proposing a novel hybrid adaptive framework that integrates EMD with a CNN-LSTM-attention model. It also evaluates this framework on real-world datasets collected from smart manufacturing environments, contributing to both methodological innovation and industrial applicability.

The proposed method introduces a novel hybrid framework that couples Empirical Mode Decomposition (EMD) with a deep learning architecture composed of Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) units, and an attention mechanism. The goal is to extract both time-localized and context-aware features from nonlinear, nonstationary signal data in smart manufacturing environments.

The pipeline consists of the following main components:

1. Preprocessing and normalization of raw signal data;

2. Adaptive signal decomposition using EMD to extract Intrinsic Mode Functions (IMFs);

3. Feature extraction from each IMF using a multiscale CNN;

4. Temporal modeling using an LSTM network;

5. Attention-based fusion to improve interpretability and decision-making;

6. Output layer for classification or regression.

Each of these stages is described in detail below with mathematical formulation and architectural integration.

EMD is a nonlinear, data-driven decomposition technique that extracts Intrinsic Mode Functions (IMFs) from a nonstationary time series. Each IMF represents a simple oscillatory mode embedded in the signal. EMD is particularly powerful because it does not require a fixed basis like Fourier or wavelet methods.

Let $x(t)$ denote a continuous-time signal. The decomposition into $n$ IMFs ${ci(t)}i=1n$ and a residual $rn(t)$ is given by:

Each IMF must satisfy two conditions:

1. The number of extrema and the number of zero crossings must either be equal or differ by at most one.

2. At any point, the mean value of the upper and lower envelopes (formed by local maxima and minima) is zero.

The sifting algorithm for extracting a single IMF operates as follows:

1. Identify all local extremes of the signal $x(t)$⁠.

2. Interpolate upper envelope $emax(t)$ and lower envelope $emin(t)$ using cubic splines.

3. Compute the mean envelope:

1. Subtract the mean from the signal:

1. Repeat the process on $h1(t)$ until it satisfies the IMF conditions (typically 5–10 iterations).

2. Subtract the resulting IMF from the original signal and repeat to extract further IMFs.

This process is repeated until the residual $rn(t)$ is either a monotonic function or contains only one extremum.

In the discrete domain, for a sampled signal $x[k]$⁠, the above procedures are implemented iteratively. The algorithm terminates based on a threshold $ϵ$ controlling the envelope mean energy:

In this study, $ϵ=10−4$ is used, and a maximum IMF count of 8 is set to prevent over-decomposition. The threshold value of $ϵ=10−4$ was selected after empirical testing on sample signals from the datasets, as it balances convergence speed and decomposition accuracy; lower values (e.g. $10−5$⁠) increased computation time by 20–30% without significant improvements in IMF quality, while higher values (e.g. $10−3$⁠) led to under-decomposition and mode mixing in noisy segments. Similarly, the maximum IMF count of 8 was chosen to avoid over-decomposition artifacts observed in preliminary trials with higher limits (e.g. 10 IMFs), which introduced spurious low-amplitude modes and reduced model accuracy by 2–4% on validation sets; fewer IMFs (e.g. 6) missed subtle high-frequency features, degrading F1-scores by up to 5%.

After decomposing the signal into IMFs ${c1[k],…,cn[k]}$⁠, each IMF is processed by a CNN to extract spatial features.

For a given IMF $ci[k],$ let $Xi∈RT×1$ represent the time series segment. The CNN applies multiple 1D convolutional filters:

Where:

1. $W(1)∈Rf×1$ is a 1D convolutional filter of size $f$⁠,

2. $*$ denotes convolution

3. $b(1)$ is the bias term

4. ReLU is the Rectified Linear Unit activation function.

Multiple layers are stacked to form hierarchical features:

Max pooling is applied between layers to reduce dimensionality while preserving features.

Let the final CNN output for each IMF be denoted as $Fi∈Rd$⁠, a feature vector of dimension $d$⁠. The CNNs for all IMFs are run in parallel, and their outputs are concatenated:

This stacked representation captures multi-resolution spatial characteristics of the original signal.

The vector FFF is passed to an LSTM layer that models temporal dependencies across time windows or across successive signal events.

At each time step $t$⁠, the LSTM cell computes:

Where:

1. $σ$ is the sigmoid activation function,

2. $ʘ$ is the Hadamard (element-wise) product,

3. $Ct$ is the cell state,

4. $ht$ is the hidden state (passed to next layers).

To enhance model transparency and focus on salient features, we introduce an attention mechanism on top of the LSTM outputs.

Let $H=[h1,h2,…,hT]$ be the sequence of LSTM hidden states. Attention weights $αt$ are computed as:

The context vector $c$ is a weighted sum of hidden states:

This context vector is passed to a fully connected layer for classification.

Let the output layer produce a score vector $y∈RK$ for $K$ classes (e.g. fault types, tool wear levels). A softmax function is applied:

The model is trained using cross-entropy loss:

Where $yk$ is the one-hot encoded true label.

The network is trained using Adam optimizer with learning rate $η=10−4$⁠, and batch size of 64. Dropout with probability $p=0.3$ is applied after the LSTM layer to mitigate overfitting.

The gradient updates for parameters $θ$ are:

Regularization is further enforced through $L2$ weight decay:

Assuming signal length $T$⁠, IMF count $n$⁠, and CNN filter size $f$⁠, the time complexity is:

1. EMD: $O(nTlog⁡T)$

2. CNN: $O(ndfT)$

3. LSTM: $O(d2T)$

4. Attention: $O(Td)$

Total complexity is dominated by the EMD and LSTM stages. The model is implemented in PyTorch, with GPU acceleration for deep learning components. The EMD module is implemented using the EMD-signal library optimized with NumPy.

The experimental validation of the proposed hybrid adaptive framework was conducted using two benchmark datasets and one proprietary dataset collected from smart manufacturing environments. These datasets include multi-channel sensor data such as vibration signals, thermal gradients, and acoustic emissions recorded under varying operating conditions.

The system was implemented in PyTorch 2.0 and trained on an NVIDIA RTX A6000 GPU with 48 GB of VRAM. For comparison, all baseline models were developed under identical data preprocessing and training pipelines to ensure fair evaluation.

1. Case 1: CNC Milling Dataset (Public)

This dataset contains vibration and acoustic signals collected during tool wear experiments on CNC milling machines. It includes 3-axis accelerometer data sampled at 20 kHz over 12 different wear states.

1. Case 2: Bearing Fault Dataset (CWRU)

From Case Western Reserve University, this dataset includes inner race, outer race, and ball bearing faults at various loads and speeds. Signals were sampled at 12 kHz from accelerometers placed on the housing of the bearing assembly.

1. Case 3: Industrial Robotic Arm (Proprietary)

Collected from an automated robotic welding cell, this dataset contains synchronized thermal, vibration, and motor current data. Anomalies include torch drift, wire feed failure, and overheat conditions. The dataset comprises 15,000 samples collected over a 72-h continuous operation period in a controlled industrial setting, with data sampled at 10 kHz across three synchronized channels (thermal, vibration, and motor current). It includes 1,200 anomaly instances, distributed as follows: 450 torch drift events, 500 wire feed failures, and 250 overheat conditions, labeled based on expert annotations from process logs.

Each dataset was preprocessed via normalization and segmented into sliding windows of 2 s with 50% overlap.

To assess performance, we adopted multiple metrics:

1. Accuracy (Acc): Proportion of correctly predicted labels.

2. F1-score: Harmonic mean of precision and recall.

3. Inference Time: Average time per prediction batch (128 samples).

4. Memory Usage: GPU memory required during inference.

5. Interpretability Index: A novel scoring heuristic evaluating human-aligned attention weight distributions using Spearman correlation with signal energy. The Interpretability Index is computed as follows: For each test sequence, we first calculate the signal energy profile across time steps as the squared amplitude summed over IMFs, normalized to [0,1]. Attention weights from the model are similarly normalized. The Spearman rank correlation coefficient ρ is then computed between the ranked attention weights and ranked energy values using the formula $ρ=1−6∑di2n(n2−1)$⁠, where $di$ is the difference in ranks for each time step i, and n is the sequence length. The index is the average ρ across all test samples, with higher values (closer to 1) indicating stronger alignment between model focus and physically meaningful signal regions.

The proposed method was benchmarked against:

1. CNN-LSTM (without EMD or attention)

2. LSTM Only

3. CNN Only

4. EMD + SVM

5. Transformer-based Classifier (small-scale)

Table 1 summarizes the average classification results across all three datasets.

As shown in Table 1, our proposed model significantly outperforms the others, achieving an F1-score of 0.938. This indicates robust classification even under complex, noisy, and time-varying conditions.

Figure 1 illustrates these findings visually, offering clear evidence of the performance gain from incorporating adaptive decomposition and attention mechanisms.

The attention mechanism plays a critical role in interpretability. Figure 2 presents a heatmap of attention weights over time steps, averaged across sequences from the robotic welding dataset.

Figure 2 shows that the model’s attention aligns with transient events – e.g. wire feed failures and thermal spikes. When superimposed on the energy profile of IMFs, we found a Spearman correlation of 0.81 between attention weights and high-energy localized signal features, indicating high semantic alignment.

Figure 3 compares inference time and memory usage. Despite having more components, the proposed model runs at 32 ms per batch, which is 45% faster than CNN-LSTM, largely due to efficient parallelization of IMF-based sub-models.

Table 2 shows the comparative inference performance of the proposed hybrid model against baseline architectures in terms of latency and memory usage.

As shown in Table 2, the proposed method achieves the lowest inference time (32 ms per batch) and reduced GPU memory consumption (210 MB), outperforming more conventional CNN-LSTM and EMD-based SVM models.

These results, also visualized in Figure 3, demonstrate the efficiency and scalability of the proposed system for edge deployment.

To further evaluate edge-deployability, we deployed the model on two representative edge devices: NVIDIA Jetson Nano and Intel Movidius Neural Compute Stick. Inference was tested on a subset of 1,000 samples from the CNC Milling Dataset, with results shown in Table 3.

These results confirm the model’s efficiency on resource-constrained hardware, with only minor accuracy drops due to quantization.

To evaluate the impact of EMD, we analyzed the distribution of signal energy across IMFs. Lower-index IMFs (1–3) typically captured high-frequency mechanical noise, while mid-range IMFs (4–6) contained the most discriminative features.

Table 4 summarizes the average energy contribution of IMFs grouped by their rank index across all experiments. As observed in Table 4, IMFs ranked 4 through 6 contribute over 64% of the total signal energy, aligning closely with the attention weights observed in Figure 2.

The attention scores from the model were highly concentrated on IMF 4–6 windows, confirming the decomposition’s utility in focusing model capacity on relevant features.

An ablation analysis was performed to isolate the contributions of each component. Table 5 outlines the classification accuracy with specific modules removed.

This confirms that EMD and Attention modules are crucial, each contributing more than a 3% gain in performance.

We performed a Wilcoxon signed-rank test to compare the proposed model’s predictions to each baseline on paired sample batches. The p-values obtained were below 0.01 in all cases, confirming statistical significance.

Additionally, a 95% confidence interval on the difference in F1-score between the proposed model and CNN-LSTM was [0.053, 0.069].

Figure 4 shows the ROC curves for the three datasets, where the area under the curve (AUC) for the proposed method consistently exceeds 0.93.

This affirms the model’s domain generalization ability, critical for manufacturing settings with frequent context shifts.

One of the central contributions of the proposed hybrid framework lies in its emphasis on interpretable signal analysis through the integration of Empirical Mode Decomposition (EMD) and attention-based deep learning. As presented in Figure 2, the attention mechanism effectively localized temporal regions of significance within the signal stream, offering a practical interpretive layer for engineering users.

Unlike traditional deep neural networks, where decision pathways are opaque, our architecture assigns quantifiable weights to different time slices and signal modes (IMFs), which enables analysts to trace model predictions back to specific oscillatory behaviors. This interpretability is critical in domains such as predictive maintenance, where human operators must validate machine decisions under uncertainty. Moreover, the high correlation observed between attention weights and energy-rich IMF segments provides statistical evidence that the model’s decision focus aligns with physically relevant signal behavior.

This result suggests the model does more than merely learn statistical shortcuts—it aligns with underlying mechanical dynamics, supporting its deployment in mission-critical environments such as aerospace, nuclear tooling, and autonomous factory lines.

The experimental results in Section 4 offer multiple insights into the generalizability and robustness of the proposed approach. Across all datasets, the model consistently achieved superior classification performance—highlighted by its F1-score of 0.938 and accuracy of 94.5%. These results are a direct consequence of the synergistic interaction between EMD and deep neural layers.

1. CNN components extract low-level spatial features from decomposed IMFs, handling noise and localized variations effectively.

2. LSTM units capture temporal transitions across IMFs and sensor windows, improving the model’s sensitivity to gradually evolving faults such as tool wear or thermal drift.

3. Attention mechanisms direct computational focus to informative regions in the temporal sequence, optimizing memory and accuracy.

This integration forms a hierarchical, interpretable representation of the signal that adapts in real-time to complex, nonlinear dynamics. The inclusion of EMD as a front-end ensures data-driven decomposition without the assumptions typically required by Fourier or wavelet transforms. This allows the system to respond organically to changing operating conditions—a property crucial in manufacturing systems where variability is inevitable.

One of the unique strengths of the proposed architecture lies in its ability to deliver high classification accuracy at low computational cost. As shown in Table 2 and Figure 3, the inference time of 32 milliseconds and GPU memory usage of 210 MB make the system compatible with real-time deployment on embedded AI chips, such as NVIDIA Jetson or Intel Movidius.

This efficiency is achieved by three deliberate design decisions:

1. Parallel processing of IMFs: Each IMF is passed through independent CNN layers, allowing for GPU concurrency and pipelined execution.

2. Temporal compactness: Using fixed-length time windows and efficient LSTM implementations reduces the recurrent computation overhead.

3. Model modularity: The architecture supports pruning and quantization without significant performance degradation, facilitating deployment on edge devices in smart factory settings.

These traits make the model especially suitable for online monitoring applications, such as:

1. High-speed anomaly detection in industrial robotics

2. Real-time tool wear tracking in CNC machining

3. Autonomous inspection of material defects in production lines

The ablation study (Table 5) reveals how each component contributes to the overall performance. Removing EMD leads to the most severe drop in accuracy (from 94.5% to 87.4%), underscoring its role in adaptive signal decomposition and noise suppression. The model without attention still performed well (91.6%), but the loss of interpretability and slight performance drop indicates the importance of focusing mechanisms in dynamic signal environments.

Removing the CNN or LSTM components yielded results similar to each other (around 87–88%), demonstrating that both spatial and temporal modeling are crucial and must co-exist for maximal performance.

The study confirms that the design philosophy of hybrid integration—rather than stacking components arbitrarily—yields substantial gains in robustness, accuracy, and transparency.

Despite its successes, the current approach has several technical limitations that warrant acknowledgment.

1. Synchronous sensor assumption: Synchronous sensor assumption: The current framework assumes that multi-channel sensor data is time-synchronized, which was verified on our datasets through timestamp alignment checks showing <1 ms jitter. However, in broader industrial scenarios, sampling jitter (e.g. 5–10 ms delays) can occur due to hardware variations. A sensitivity analysis on simulated jitter (added Gaussian delays up to 10 ms to test data) showed only a 2–3% drop in F1-score, indicating moderate robustness; future work may incorporate dynamic time warping or sequence alignment to fully mitigate this.

2. Noise sensitivity in EMD: Although EMD performs well in nonstationary environments, it can still suffer from mode mixing or end effects, especially under high noise levels. Using advanced variants like CEEMDAN or noise-assisted EMD could further enhance decomposition robustness.

3. Model interpretability remains empirical: While the attention mechanism improves interpretability, the framework lacks a formal method to validate whether attention weights reflect causal signal mechanisms. Interpretability remains observational rather than guaranteed, which may be insufficient for regulatory or safety-critical applications.

4. Preprocessing latency: The EMD stage adds ∼12% overhead to total inference time. Compared to alternative decomposition methods, this 12% overhead is relatively efficient; for instance, Ensemble Empirical Mode Decomposition (EEMD) incurs 25–35% additional latency due to multiple noise-assisted trials, while Wavelet Transform (WT) adds only 5–8% but lacks EMD’s adaptive basis, leading to poorer handling of nonstationary signals in our preliminary tests (e.g. 3–5% lower F1-scores on noisy subsets). While negligible on modern GPUs, this may be critical for ultra-low-latency applications (e.g. sub-millisecond response systems). Potential solutions include EMD acceleration using parallel computing or real-time signal approximation techniques.

These limitations are not unique to our model but represent broader challenges in adaptive AI for real-world systems. Addressing them offers a rich opportunity for future research.

From a broader perspective, the proposed method contributes to the ongoing convergence of signal processing, machine learning, and edge computing. It serves as a template for:

1. Embedding domain knowledge (via signal decomposition) into deep models,

2. Building interpretable yet performant architectures, and

3. Designing systems that operate autonomously and transparently in uncontrolled environments.

This fusion of data-driven and physics-aligned modeling is essential as smart factories become increasingly autonomous. With the right deployment pipeline, the proposed framework could form the core of intelligent monitoring systems, self-diagnosing and responding to faults with little to no human intervention.

This study introduced a novel, interpretable hybrid framework for real-time analysis of nonlinear and nonstationary signals in smart manufacturing environments. By integrating Empirical Mode Decomposition (EMD) with a hybrid deep learning model composed of CNN, LSTM, and an attention mechanism, the proposed system achieves significant advancements in both predictive performance and model interpretability.

The application of EMD enabled adaptive signal decomposition without reliance on fixed basis functions, effectively isolating localized oscillatory modes relevant to fault conditions and process anomalies. These IMFs were then processed using CNNs to extract spatial features and LSTMs to capture temporal dependencies. The attention mechanism further enhanced the model’s diagnostic power by identifying critical regions in the signal where decision-relevant changes occurred, thus offering interpretability aligned with physical phenomena.

Experimental evaluations across three datasets—CNC milling, bearing fault detection, and industrial robotic arm monitoring—demonstrated the model’s superiority over standard baselines. The framework achieved a mean classification accuracy of 94.5% and an F1-score of 0.938, outperforming traditional CNN-LSTM architectures, EMD with classical classifiers, and other ablation variants. Importantly, the model preserved a low inference latency (32 ms per batch) and modest memory usage (210 MB), making it suitable for deployment on edge devices in production environments.

From a technical standpoint, this work makes three core contributions:

1. It provides a principled integration of signal decomposition and deep learning, ensuring adaptability to nonstationary signal characteristics.

2. It incorporates an attention mechanism that not only improves performance but also yields explainable inferences, satisfying operational demands for traceable AI in high-stakes settings.

3. It balances accuracy and computational efficiency, meeting the requirements of real-time manufacturing systems constrained by latency and resource budgets.

The findings reaffirm the utility of hybrid architectures in bridging the gap between signal theory and machine learning, particularly in industrial contexts where interpretability, adaptiveness, and low-latency operation are not just desired but essential. The proposed system offers a scalable, explainable, and robust solution to the challenges of signal variability, transient event detection, and decision traceability in modern smart factories.

In conclusion, the work presented herein lays a strong methodological foundation for future advancements in adaptive, interpretable, and real-time AI-driven monitoring systems tailored to complex engineering environments.