The experiment was conducted to evaluate the effectiveness of combining TinyML and CTGAN for intelligent load classification in Industry 5.0. The primary goal of the study is to classify the electrical appliances based on their power usage patterns48. The CTGAN is used to enhance the accuracy of classification, while TinyML is utilized to deploy the trained model onto a microcontroller49.
Experimentation setup
The experimental setup was designed to build a robust smart load classification system capable of running microcontrollers using TinyML and CTGAN50. Table 2 contains the basic experimental setup for the proposed study.
Optimizer comparison
The proposed model is tested with different optimizers, which are fitted to produce the best tradeoff between high prediction accuracy and low training time. The paired t-test is performed between different optimizers, namely the Adam optimizer, the AdaGrad optimizer, and the RMSprop optimizer, under fixed settings of hyperparameters such as batch size (500) and epochs (10), with 5 different seeds. The results are expressed in Table 3, including the mean and standard deviation51.
Table 3 shows that the Adam optimizer’s performance on the paired test is below 0.01, compared to AdaGrad and RMSprop optimizers, which are 0.27 and 0.81, respectively. For this reason, the proposed model utilizes the Adam optimizer to produce effective results for different classes of load prediction on current and future CO2 emission functions, using a TinyML with CTGAN52.
Hyperparameter ablation study
The proposed model hyperparameter selection ablation test is conducted using a grid search with an estimator ε for different accuracy ranges, such as 50 to 500, depth as 4, 8, 12, and finally, batch size as 250, 500, and 1000. After the grid search process get over the proposed model classifier configuration for resource constrained devices run for 5 cross folds with different seed where final interface latency as 2.1ms for 100 estimators with max depth of 8, TinyML model size for micro controller as 128 kb with mean accuracy as 94.9% The results are expressed in Fig. 3, where the accuracy improvement after 100 trees as marginal towards 0.5% where the plateau arises after 200 trees for TinyML deployments.
Hyperparameter Ablation Study.
Results analysisConfusion matrix
The confusion matrix is used to evaluate the performance of the classification model by comparing the actual equipment classes and predicted classes51. The confusion matrix for this proposed work is plotted among light load, medium load, and heavy load, as illustrated in Fig. 4.
The figure is plotted based on the test dataset, which is evaluated after CTGAN and TinyML model quantization. The diagonal elements represent the correct classifications, whereas the off-diagonal elements are classified as incorrect classifications. The incorrect classifications are relatively low compared to the proper classifications.
Correlation heat map
The correlation heat map is used to visualize the linear relationship between different features in the dataset which are used for smart load classification. This will categorize the dataset into two categories namely strongly correlated features and weakly correlated features. This will be much helpful during feature selection and dimensionality reductions. The Fig. 5, depicts the correlation heat map of the proposed model.
The Fig. 5 illustrates the correlation heat map of the proposed methodology. This has been plotted among the important key attributes of the dataset like usage load, CO2, NSM, current, month, lagging and leading power factors. In the diagram, which is displayed as red are strongly correlated whereas which are indicated blue in colour are weakly correlated. The power usage with lagging current (reactive), power usage with CO2, lagging current (reactive) with CO2 and NSM with hour are strongly correlated. Likewise, leading current (reactive) with normal leading current is weakly correlated.
Performance metrics
The performance of the intelligent energy load classification system, utilizing TinyML and CTGAN, has been tailored for Industry 5.0, which can be attributed to three primary reasons. They are human-machine collaboration, decentralization and sustainability. The final performance of the model depends on the accuracy of the classifications. The accuracy of the classifications is given as.
$$\:Accuracy=\:\frac{Number\:of\:correct\:predictions}{Total\:number\:of\:samples}$$
(15)
Usually, for any classifications the correct prediction and incorrect predictions of the samples will be termed as true positive (TP), true negative (TN), false positive (FP) and false negative (FN). Using these parameters, the performance metrics like precision, recall and F1-score will be calculated.
$$\:Precision\:\left(c\right)=\:\frac{{TP}_{c}}{{TP}_{c}+{FP}_{c}}$$
(16)
$$\:Recall\:\left(c\right)=\:\frac{{TP}_{c}}{{TP}_{c}+{FN}_{c}}$$
(17)
$$\:F1\left(c\right)=2\:\times\:\:\frac{Precision\:\left(c\right)\times\:Recall\:\left(c\right)}{Precision\:\left(c\right)+Recall\:\left(c\right)}$$
(18)
From the precision, recall and F1-score values, the weighted average values can be calculated as follows.
$$\:Weighted\:precision=\:\sum\:_{c=1}^{C}\frac{{n}_{c}}{N}\:\times\:precision\:\left(c\right)$$
(19)
Where nc is the support of class c and N is the total samples.
$$\:Weighted\:Recall=\sum\:_{c=1}^{C}\frac{{n}_{c}}{N}\:\times\:recall\:\left(c\right)$$
(20)
$$\:Weighted\:F1=\:\sum\:_{c=1}^{C}\frac{{n}_{c}}{N}\:\times\:F1\left(c\right)$$
(21)
Model evaluation
The model evaluation represents the process of measuring performance metrics, such as accuracy, precision, recall, and F1-score, of the classification task. Figure 6 illustrates the model evaluation of the proposed study with convergence curve.
Figure 6 clearly illustrates that all performance metrics, including precision, recall, F1-score, and accuracy values, are consistently high (> 92% to 98%). With an accuracy of 95%, it indicates that the overall percentage of correct predictions is high. The convergence curve is occurred with downward trend towards 10th to 12th epoch which defines the model effectively learns and stable with different epochs. From the above result the model evaluation states that TinyML with CTGAN capable of separating multiple load condition with balanced precision, recall, F1-score, and accuracy over low, medium, high load categories.
Feature usage of load prediction
The top feature of this research is NSM, and hour both the functions are directly related to human work shift which is directly link towards the production cycle and industrial equipment on and off in innovative industry environments, these features are capture and fine grained as temporal structure for class like heavy load, light load and medium load starting from timeline of 6 AM to 10 PM. After the NSM, the hours of usage are considered another critical feature for the whole analysis of industrial equipment operation states. The plots define an increase in usage, showing the production hours. Based on usage, CO2 emissions are simultaneously analysed, which directly relate to the output attained during demand time. Production operation schedules are mapped as a feature necessary for predictive maintenance to predict load.
Feature Usage of Load predictions.
The Fig. 7 illustrates the top 10 important features for load prediction namely NSM, hour, usage, lagging current, month, CO2, lagging current (reactive), leading current, leading current (reactive) and weak status whereas the NSM features contribution towards the prediction is high and weak status is low in load prediction. This can support in the electrical appliance’s states like different loads. By enriching the training dataset with high quality synthetic samples, the TinyML and CTGAN helps in building more resilient and accurate load prediction models.
Data load distribution
The data load prediction refers to the statistical properties of the energy consumption data, such as light load, medium load, and maximum load. The TinyML algorithm is resource-constrained and requires statistical patterns to learn. The CTGAN is used to model and augment imbalanced data distributions. It will understand the distribution of tabular data and generate synthetic samples to imbalance. Figure 8 shows the data load distribution.
From this figure, it’s clearly observed that the original data load distribution and synthetic data load predictions for the light, medium and maximum load are optimal and distributed wisely. This is based upon the time and electrical appliance type. It reflects the original and synthetic distribution compactly because its primary goal is to predict across all load levels and devices.
The energy usages
From Fig. 9, it is observed that the medium load optimization is processed with four batching processes, namely (a) staggering start, which minimizes the continuous medium load start by 6.2% of peak load lowered by 2 h of CTGAN. (b) Time shifting, which processes off-peak hours monitored by the proposed model, which issues no-critical batch shifts. (c) operator alerts for medium load, which is off the production schedule. (d) automated control, which alerts to emissions and maximum load as real-time constraints.
Energy usage differences 9b: Energy usage in histogram graph.
Figure 9a illustrates the energy usage on weekdays and weekends, while Fig. 9b shows the energy usage by load for both the original data and the synthetic data. This is useful for determining which device is consuming energy in real-time. Additionally, it identifies high-consumption devices for efficiency improvements. It generates more load samples to train a machine learning model. Additionally, it reduces the overuse of electrical devices in the industry and helps schedule loads efficiently.
Energy vs CO2 by load
The energy usage metrics processed by TinyML with CTGAN present an evaluation pipeline where the input load data is pre-processed, with the load data split based on the equipment’s run, at a ratio of 70% for training and 30% for testing. After initial processing is completed, CTGAN augmentation is processed to balance the data classes. The proposed model is trained using hyperparameters and an optimizer for efficient estimation, with k-fold cross-validation performed over 30 runs. Then, the TinyML deployment is tested with quantized, pruned, and flashed into memory as an ESP32. After flashing the model into memory, the inference latency and RAM usage are analyzed on resource-constrained devices using the memory operator. Finally, energy load usage week weekdays are analyzed with low and high loads, where medium load dates need to be checked normal process is expressed in the figure below.
Energy Vs CO2 emissions by Load.
The TinyML is used to optimize the running of resource constrained devices and the CTGAN is used to generate the synthetic data which can be augment imbalanced datasets. The Fig. 10 depicts clearly the reduced energy wastages and emissions. Also, it has identified the high impact loads with inconsistent loads.
Average energy consumption
The combination of TinyML and CTGAN in the industry 5.0 gives energy efficiency and emission reduction. By using the average energy consumptions, the load types can be identified easily. The Fig. 11 illustrates the average energy consumption.
Average Energy Consumption.
The energy consumption are the mandatory parameters for improving the industry environments. It gives clear feedback to operators on emissions and consumption. It supports real-time tracking of environmental impact. It works offline and continues to monitor without external servers. It improves the model accuracy with synthetic data and runs it on edge devices.
Lagging power factor
The lagging power factor is defined as the condition in which the current waveform lags behind the voltage waveform in electrical systems. Inductive loads, such as motors, transformers, and HVAC systems, are responsible for this. Figure 12 illustrates the lagging power factor which are obtained during the implementation of the proposed work.
The lagging power factor is very important in the field of smart load classification. From Fig. 12, it is clearly classified as light, medium, and maximum load in both the original and synthetic data. The electrical devices will create a magnetic field as a part of their operations, which leads to lag voltage. This may incur penalties due to the low power factor. This may require power factor correction using the capacitors.
Hourly load type
The hourly load type prediction in the smart load classification is required for the time dimension of the proposed classification model. This is a perfect parameter because the load pattern may vary significantly by hour in industrial settings. Figure 13 illustrates the hourly load type of the proposed model.
Figure 13 illustrates the hourly load types of the original data and the synthetic data. It will categorise the type of electrical load by hour of the day, such as motor running hours, standby loads, and others. This may improve the temporal intelligence of the proposed model. Additionally, it detects when inefficient loads are active using time-aware insights. Using human-centric operations will help the operator understand machine behaviour hourly.
Hourly energy usage
This parameter adds time-based energy consumption analysis to the proposed model. This will be more supportive of load profiling, demand forecasting and operational optimisation. This enhances the decision-making process by taking into account energy usage.
Figure 14 defines the high usage of load and CO2 wastage periods. Here, the period of 10:00 to 13:00 h is identified to track peak consumption hours, where the daily load consumption reaches 28% and CO2 Emissions reach 30% of wastage. After tracking mean consumption, high wastage periods occur during non-production, appearing at 14:30 h, which contribute 6.7% wasted during standby. These wastages are correctly predicted by the proposed model, suggesting the mitigation, which includes the shutdown of unused machines during peak time, which minimizes the load and CO2 emissions.
Proposed model performance analysis
The performance analysis involves both evaluating the machine learning model and assessing the system-level impact. The machine learning model analysis can be evaluated by calculating accuracy, precision, recall, and F1-scores. System-level analysis can be conducted with the help of light loads, medium loads, and maximum loads in an industrial environment.
Key result three panel results of the proposed model.
Figure 15 depicts the results as (a) shows the per-class precision and recall with mean across 30 runs, (b) shows the overall accuracy of the baseline model with the proposed TinyML with CTGAN, and finally (c) defines the model size vs. latency trade-off for 200 estimator trees with error bars. It achieves a load type prediction accuracy of 95%, which is significantly high in smart industry environments. Like accuracy at 95%, the other features, such as NSM at 0.3, hour at 0.2, power usage at 1.4, lagging current at 0.07, and month values at 0.06, RBF kernel mean heuristic 0.038 are also predicted correctly for final validation.
Synthetic data validation
The proposed model utilizes the tabular condition (CTGAN) to compute the fidelity function, where the per-feature KL divergence for real and synthetic data is accessed with Frobenius normalization. The results are expressed in Fig. 16, where NSM is 0.021, Current usage is 0.037, CO2 emission is 0.042, lagging current is 0.033, and normalization is 0.11, computed to preserve inter-feature correlation and achieve class balancing.
Synthetic data analysis of Tabular CTGAN for Class balancing.
Statistical testing K fold cross validation
The proposed model was evaluated using 10-fold cross-validation for 30 independent runs. During each fold of the proposed TinyML with CTGAN process, a Paired two-sided t-test is performed to compare the mean accuracy level at 95% confidence with that of other models. Similarly, models like XGBoost and LightGBM are also analyzing the predictive load in an Industry 5.0 standard. The proposed model achieves an accuracy of 95% with a mean of 0.8, and p-values of 0.05 are obtained using the Wilcoxon signed-rank test. The other state-of-the-art model, like XGBoost, achieves an accuracy of 94.3% with a mean of 1.1 and a p-value of 0.09. Similarly, LightGBM achieves an accuracy of 94.8 with a mean of 1.2 and a p-value of 0.21. Compared to all other models, the k-fold cross-validation is stable in achieving accuracy with a 95% confidence level for the proposed TinyML with CTGAN, which exhibits significant performance in the tabular setting for analyzing load in the smart industry, as expressed in Fig. 17.
Statistical Testing K Fold Cross Validation.
Ablation study of CTGAN
The results are analyzed with three different training of CTGAN considered as ablation, where initial process is processed with original data where accuracy as 93.8% with minority recall as 85.1%, next with original + CTGAN mixed achieve an accuracy as 95% with minority recall as 91.3%, finally with 100% synthetic test set archives 92% accuracy with minority recall as 88.4%. The mixed data improve minority recall by 6% and increase accuracy by 25% compared to the original training set, demonstrating that the proposed model, incorporating TinyML and CTGAN, is beneficial for predicting load and CO2 emission classification in class-imbalanced tabular data, as shown in Fig. 18.
The quantitative distribution defines the TinyML model fused with CTGAN fidelity evaluated augmentation methods like SMOTE, ADASYN, CVAE a real test set, where regular training with entirely original data accuracy as 93.8%, some original and synthetic add-ons achieved accuracy as 95%, fully synthetic seed data as 92.6%, and finally, minimum class recall increases from 85 to 91% for combining both synthetic and original data expressed in Fig. 19.CTGAN produce the best trade off with lowest MMD as 0.003 and correlation distance as 0.11 and highest mixed mean accuracy as 95% due to this comparative experiment the tabular industrial data which contain different categorical feature and multi model distribution desired to produce good comparative result than other augmentation comparative experiment models.
Quantitative distribution Augmentation methods.
Load type distribution prediction.
The Fig. 20 represents the load type prediction. The proposed work is correctly predicted the light load, medium load and maximum load. The time taken by the TinyML on the devices are very less, so the power used by the devices can be inferred easily. The Table 4 illustrates the comparison of the proposed model with existing methods.
The TinyML used in the research is ahead of other ML model baselines, such as quantized MLP and Bonsai, where the size of the MLB is 96KB and Bonsai is 72KB. Still, the accuracy attained by the models is 92.1% and 89.4%, which are relatively low compared to Random Forest, which achieves superior accuracy, a smaller model size, and a lower latency trade-off. The proposed model is very competitive and it gives accuracy of 95% which is significantly high when compared to other models. It is a lightweight model when compared to LSTM, CNN + LSTM and TabNet. The CTGAN is a futuristic approach as it balances the dataset for load prediction. When compared to XGBoost and lightBGM models, the proposed model is good in speed. When compared to the deep learning models, the proposed work achieves better accuracy without the support of massive computational power.
Practical deployment analysis
The practical deployment of this proposed model involves the concept drift detection, which is achieved using a 1-minute window monitor to classify the load probability. This detector helps analyze selective load data with a centralized training set. The window monitor runs continuously on the device, requiring minimal computational resources. After monitoring the load probability feature, the cost is analyzed using a rolling window function with a 5-minute timestep to calculate the load and CO2 emissions. This analysis incorporates exponential moving averages, with a 3.4ms inference latency, which includes sensor readings of load and emission features. Finally, the system latency is analyzed with an end-to-end sensor read and action display, where the mean latency lies around 3.4ms, based on input-output overhead. This real-time latency is more suitable for real-time implementations in industry operations, such as alerts and load classifications.