Low-Parameter Critic-Based Multivariate WGAN Model for Clogging Detection in Drives

The role of environmental factors that impact operation of variable frequency drives (VFDs) has recently sparked a vibrant discussion in industry on extending the functionalities of their existing preventive maintenance systems (Al-Naseem and El-Sayed, 2013). This is due to the fact that frequent warranty interventions generate additional exploitation overheads both at the user and the producer/service ends (Yellamati et al., 2013). Among them, the influence of environmental factors, such as airborne pollution, humidity and ambient temperature, among others, were recently shortlisted common factors underlying most of their failures (Kennedy, 2021). In particular, the clogging of the cooling system, i.e., inlet/outlet filters, by airborne particles, seems to be particularly troublesome (Avor and Chang, 2019). It can lead to accumulation of metallic particles in boards, likely causing short circuits by contamination failure and increasing the risk of harmful corrosion (Avor and Chang, 2019). Therefore, detecting the requirement of filter replacement can be considered as a valuable maintenance action, so that anomaly/failure detection can be seen particularly beneficial to users to assure reduced exploitation costs (Avor and Chang, 2019). In the working environments with a high concentration of particulate matter, especially when combined with moisture and oil, the fan draws in all the debris, so that the layer of dust can be formed on the electronics, leading to the overheating of single elements in a VFD (Avor and Chang, 2019). Heat sinks are particularly vulnerable targets to dust accumulation. Their contamination results in reduced dissipation efficiency, so that the excessive heat is no longer efficiently drawn away to the exterior (Peterson, 2022). Excessive heat generation along with possible short circuits in a VFD can facilitate ageing of the power electronics system (Kang et al., 2023). For example, insulated gate bipolar transistors (IGBTs) can get damaged by increased temperatures resulting from overcurrent and overvoltage in the chips (Otsuki et al., 2003). One critical issue is that under prolonged thermal stress, the solder layer in IGBTs tends to develop cracks and voids, thus leading to delamination and subsequent failures (Ciappa and Fichtner, 2000). Inefficient operation of the cooling system in a VFD can also impact operation of the DC link circuit (Avor and Chang, 2019). This causes excessive capacitor wear, which can trigger transients, resulting in failures of input diodes and the capacitor itself. The latter can be due to an increased equivalent series resistance (ESR) value at high temperatures, gradually leading to DC link failure and time-consuming repairs in field conditions (Avor and Chang, 2019).

For both monitoring the conditions and screening failures in VFDs, anomaly detection algorithms are of increasing popularity in the power electronics of the drives (Zhang et al., 2023). Anomaly detection, in general, is a method that is aimed at detecting patterns that deviate from expected normal operation in the data and is typically used for outlier and novelty analysis (Chandola et al., 2009). There are two major types of anomalies: (a) point (abnormal parcels in the data) and (b) contextual anomalies (unusual temporal-wise patterns) (Chandola et al., 2009). Concerning time series data, often used in monitoring of VFDs using sensors, the latter type is more relevant, so that temporal correlations can be unravelled between consecutive timestamps (León-López et al., 2021). Model-based anomaly detection methods were recently found particularly reliable and flexible approaches for detecting faults in power electronics in VFDs, due to their flexibility, modularity and hence possibility for further using in embedded systems (Beattie et al., 2022). In particular, the usage of models utilising temperature signals was recently advised as it can yield models with better accuracy and training speed (Gómez et al., 2024). To this end, the use of machine/deep learning is often efficient (Surówka et al., 2023). We have recently demonstrated that anomaly detection systems using complex models, such as deep autoencoders (using either linear or long-short-term memory (LSTM) units), can be successfully applied to detect point anomalies in temperature signals from drives in wind turbine and simulated data (Surówka et al., 2024a,b). More recently, we showed that anomaly detection using hidden Markov models (HMM) with Gaussian mixture model initialisation is an effective method in detecting clogging in the cooling system of the drive: inlet, heatsink and outlet (Surówka et al., 2024a,b).

Recently, there has been growing interest in applying generative AI tools for anomaly detection due to their enormous potential to detect the most subtle and complex patterns in data, similar to a human expert (Tuli et al., 2022; Xu et al., 2022). Specifically, there has been an increasing number of applications of generative adversarial networks (GAN) for anomaly detection in time series data (Lüer and Böhm, 2021). The core idea behind GANs involves adversarial training of two models: a generator and a discriminator (or critic). The generator transforms random noise into data similar to the training set, while the discriminator aims to distinguish between real and generated data (Chollet, 2018). This adversarial process guides the generator to produce more realistic data while simultaneously improving the discriminator’s ability to detect data that differ from the training examples (anomalies) (Yin and Kaynak, 2020). However, the training process of GANs is very difficult to converge due to the instability of the adversarial training dynamics and the sensitivity to hyperparameter settings (Xia et al., 2023). To address this issue, Wasserstein Generative Adversarial Networks (WGAN) networks were proposed to improve upon traditional GANs by providing more stable training and better quality of generated data. This is achieved through the utilisation of the Wasserstein loss function Eq. (1) (Chen et al., 2022): (1) L=Ex∼PrealDx−Ez∼PzDGz L = {E_{x \sim {{P} _{{\rm{real}}}}}}\left[ {D\left( x \right)} \right] - {E_{z \sim {P_z}}}\left[ {D\left( {G\left( z \right)} \right)} \right] where D (x) denotes the critic score for real data x, (D (G(z)), is the critic score for generated data G (z), P_real is the distribution of real data, P_z is the distribution of the noise input to the generator and z represents the noise data. This equation encapsulates the zero-sum game nature of the adversarial training in WGANs. The generator model tries to minimise the loss L by generating data from noise (G(z)) that resembles the training data, while the discriminator (critic D) aims to maximise the loss by accurately distinguishing real data from generated data (Gulrajani et al., 2017). The noise data z are thus used to explore and adjust the latent space of the training data (Chollet, 2018).

WGAN networks are currently proposed as promising solutions for anomaly detection due to their enormous potential for detecting subtle and complex anomalies in the time domain (Qi et al., 2023). This capability stems from their unique adversarial training process, which mimics human-like learning by continuously improving through feedback between the generator and the discriminator (Goodfellow et al., 2016). For example, recently, Dai et al. (2024) proposed an unsupervised anomaly detection model called Wave-GANomaly, which uses multi-feature fusion and selection, and was proved successful, as benchmarked against CIFAR-10 and MNIST datasets.

Unfortunately, WGAN models often pose a challenge in productization due to generalisation/training convergence issues caused by limited data availability, especially that of high accuracy and diversity (Tang et al., 2025). Not to mention that in case of huge model architectures, there are complex problems with their deployment in embedded systems, with limited memory, thus constraining their operation to Clouds (Qi et al., 2023). To address these issues, herein, we propose to use a WGAN model with a low parameter critic (for anomaly detection) and a substantially complex generator (for training only).

Figure 2a shows the stacked, augmented, ambient-subtracted training data (no clogging) corresponding to ambient temperature changes between 25°C and 45°C. It is important to note that we originally had only three profiles for 25°C, 35°C and 45°C. However, the training of the WGANs with such limited data was inefficient with frequent gradient explosions. To somewhat smooth the cost hyperspace, the new data were interpolated from the original profiles, thus introducing data augmentation. By analysing these data, we can see that the net peak temperature values (upon ambient subtraction) were growing until approximately 35 C and then started to slightly decrease (<5%), which was due to the cooling efficiency of the ventilation system in the drive. Figures 2b–d show the applied testing data that were a blend of healthy profiles (25°C, 35°C and 45°C) and those for the 50% and 75% clogging for the inlet (Figure 2b), outlet (Figure 2c), heatsink (Figure 2d) and fan blockage (Figure 2e). Compared with healthy data, the most dramatic changes in the profiles were observed for the heatsink clogging case (Figure 2d); whereas, the least changes were noted for the outlet clogging case (Figure 2c), including both startup/cooling and active operation cycles. Therefore, we could conclude that air flow restrictions in the outlet caused no substantial harm to the performance of the ventilation system up to the clogging limit of 75% (cf. Figure 2c). Above this point, pressure losses and air resistance could become more significant. This could make the air flow turbulent, reducing overall cooling efficiency by the fan (Manaserh et al., 2021; Zmrhal and Boháč, 2021). At the same time, clogging of the inlet restricts the system from cool air with more dramatic impact. The data from Figure 2a data were then used for training and testing of WGAN networks. First of all, we capitalised on developing a WGAN architecture with a small critic model. The goal was to produce the critic model with the smallest number of trainable parameters possible using the 1D convolutional layers, which helped to save on parameter counts. As shown in Table 1, the M1 architecture with a one-filter critic turned out to be the smallest possible. We also tested the M4 and M8 architectures to compare their performance with the M1 model, so that we could assess if higher information capacity of a model brings better performance. By comparing final implementations of these models, their size was < 10 kb, which seems to be sufficient to be handled in low memory systems. With the M1 architecture in place, the major challenge in this study was to perform stable and convergent training of the M1 model while generating data that resembled the training dataset. Figure 4a shows the training curve for the WGAN model of the M1 architecture. These data indicate that the training was convergent for both the generator and the critic, with neither significant vanishing gradients nor gradient explosions. The generator’s quality was assessed by generating example data using input normal noise, which were then visualised using the t-distributed stochastic neighbor embedding (t-SNE) transform with two components, as shown in Figure 4b. The distributions of the generated and real training data were found to overlap, although the spread of the artificially generated data was smaller than that of the original training data. This suggests that the WGAN captured only the most essential part of the variance in the training data, implying that the M1 model was not overfitted. Figure 5 shows the outcome of the critic-based anomaly detection procedure for the testing data for: the heatsink (cf. Figure 5a), inlet (cf. Figure 5b), outlet (cf. Figure 5c) and fan blockage (cf. Figure 5d) cases. In these figures, given a specific clogging case, critic scores were computed for the data shown in Figures 5b–e and then subjected to the asymmetric OBAS transform (see Eq. 2). The parameters for the OBAS transform were derived from the critic scores computed for the training data (see Figure 2a). Figure 5 also presents the outcomes for more complex architectures, M4 and M8, to evaluate whether adding more complexity enhances performance. These data indicate that the outcomes for the M4 and M8 models were nearly four times weaker than those for M1. This seems to be due to their excessive information capacity and the relatively low volume and complexity of the data, which could lead to undertraining. Undertraining is challenging to conclude definitively, as it depends on the available testing data and can be unmasked by any new datasets. If any anomaly detection model cannot distinguish between healthy and failure (unseen) data (similar anomaly scores assigned to healthy and failure data), likewise presented herein, it indicates a failure to capture essential features. This could be caused by, among others, insufficient training or a too complex structure of a specific model, making generalisation too difficult to achieve, given available data. Despite further attempts at parameter tuning and further training, the M4 and M8 models faced vanishing gradient issues due to their higher number of parameters, making training difficult. The M1 model, with lower information capacity, performed well, showing convergent training and appropriate responses to anomalous data. In contrast, the M4 and M8 models, with 7× and 20× more parameters respectively, struggled due to likely numerical instability, despite attempting to train them at the limits of regularisation. Conversely, the OBAS values for the M1 model were generally higher for clogging data compared to healthy cases, regardless of ambient temperature. The critic-based OBAS scores generally reflected the patterns observed in the data in Figure 2b–e, with the highest OBAS scores for inlet clogging and the lowest for outlet clogging. The highest peak values for OBAS were observed during the start-up and cooling operation phases, while responses during active operation were just slightly above the threshold. Figure 5 shows that the M1 model generally remained silent (OBAS < 1) for no-clogging data, regardless of ambient temperature. However, for inlet clogging, as presented in Figure 5a, the anomaly indicator far exceeded the threshold starting from the 75% clogging data acquired at 25°C during both the startup and active operation phases, although the responses were borderline for the latter. As shown in Figure 5b, the M1 model indicated weak/boundary (OBAS ∼1) outlet clogging anomalies starting from the 75% clogging case at 25°C, with maximum indications for the 75% clogging case at 35°C. Figure 5c demonstrates that for the heatsink clogging case, the M1 model responded to all clogging cases, with the strongest response during the startup/early cooling operation phases. Again, the responses to anomalous cases were just above the threshold during active operation (OBAS in the range [1, 2]). The results for fan blockage, as shown in Figure 5d, were generally similar to those for inlet and heatsink clogging scenarios, with OBAS values increasing for the 75% clogging case as the ambient temperature increased. Overall, the results from Figure 5 demonstrate that the M1 model can generally detect all clogging cases; whereas, the responses for the M4 and M8 architectures were at most borderline. This indicates that increasing the number of parameters (filters) in the critic model does not improve performance. By comparing these outcomes with our previous study, we can conclude that the critic response to anomalies was weaker than that of HMM (Surówka et al., 2024a,b).The latter produced very strong responses up to OBAS∼10² for 75% clogging cases in the inlet and heatsink. However, HMMs fell short in detecting some clogging scenarios: the outlet clogging was almost undetectable, and they struggled to respond to 50% clogging cases (Surówka et al., 2024a,b). In contrast, the critic-based WGAN model was more balanced in its response: OBAS indicators were significantly smaller (OBAS < 4), but they responded to almost all clogging cases, including the outlet clogging and 50% clogging cases missed by HMMs (Surówka et al., 2024a,b). Additionally, the critic scores were more reflective of the mutual difference in temperature readouts between anomalous and non-anomalous data, ranging between 5% and 10%, compared to HMMs where the anomaly scores were exaggerated (Surówka et al., 2024a,b). This reflects the nature of the feedback-based training, so that the ‘critic’ level, as the name suggests, is proportional to the effect the same way a human expert will do. Another question we posed was the minimal detectable clogging/blockage level that could be identified using our proposed model architecture (M1). This turned out to be a difficult question due to limited data: we only had profiles acquired for 25°C, 35°C and 45°C including no clogging and clogging data. However, we found out that we can interpolate between these boundary profiles with three anchors: 0% clogging (just ambient data), 50% clogging and 75% clogging data, given a specific ambient temperature. The interpolation covered ranging from 0% to 75% with 5% increments. The M1 critic was then applied to these data, including the OBAS transform, as shown in Figure 6 for ambient temperature values of 25°C (cf. Figure 6a), 35°C (cf. Figure 6b) and 45°C (cf. Figure 6c). Generally, the response to anomalies at given clogging levels varied with different ambient temperatures. From Figure 6a, we can see that at 25°C, the anomaly threshold was exceeded at 25% fan clogging/blockage, 30% heatsink clogging and 55% inlet clogging. However, the outlet clogging was almost undetectable at 25°C, with weak anomalies appearing from 70% on. At 35°C (cf. Figure 6b), stronger responses to clogging were observed, and all clogging scenarios could be detected starting from the 5% level. However, caution is needed as the indications for 0% clogging just hit the anomaly threshold. At 45°C (cf. Figure 6c), fan blockage could be detected from 25%, heatsink clogging from 30%, inlet clogging from 45% and outlet clogging from 40% (with a borderline outcome). In total, the data in Figure 6 indicate that the critic model is reactive from roughly 20% to 30% clogging levels, with the response being influenced to some extent by ambient conditions. To further understand this behaviour, we quantitatively assessed the anomaly detection capabilities of our low-parameter critic-based WGAN model (M1) by comparing performance metrics, such as ACC, TPR, FNR, PPV, and F1, as shown in Table 2 (input data from Figure 5). The metrics were computed for each clogging scenario separately. As the non-anomalous data ranges and the proportion of positive and negative data were the same in these plots, therefore computing FPR would not indicate any meaningful differences between different scenarios and was therefore skipped. For inlet, heatsink and fan clogging cases, the best outcomes were noted for startup and cooling phases. These phases exhibited 10–20% higher ACC and significantly higher TPR compared to the active operation phase. It can also be drawn from these data that the performance of outlet clogging detection was low, although slightly higher than a random classifier with 50% accuracy. Additionally, there was good confidence in positive predictions, as reflected by PPV values ranging from 0.7 to 1.0 for all cases except outlet clogging. Starting with the inlet clogging prediction of the heatsink, as shown in Table 2, the models achieved good ACC of 76%, TPR between 82% and 84% and FPR of 10%–30% for both startup and cooling phases. Regarding the outlet clogging case, the performance figures indicated that the model had only a weak detection capability, reflected by low F1 scores between 0.2 and 0.6. For heatsink detection during startup and cooling phases, the performance was excellent, with an accuracy of nearly 90% and an F1 score above 0.9. Similarly, for fan blockage detection, the model achieved an F1 score of 0.87 for both startup and cooling phases. Overall, Table 2 shows that the developed clogging detection model had higher accuracy and TPR during startup and cooling phases; whereas, the active phase exhibited lower performance metrics, particularly for outlet clogging detection. These results align with our previous study on clogging detection in VFDs using HMMs, where higher performance figures were also concluded for the startup phase compared to active operation. The only difference is that HMMs were more efficient in detecting inlet clogging, while in this study, substantially better performance was reported for heatsink clogging detection. It shall also be mentioned that the previously applied 10–40 states HMMs had substantially higher number of parameters ranging from 100 to 1,500 (Surówka et al., 2024a,b). Concerning other studies, to our knowledge, there is little evidence of the application of critic-based WGAN models for condition monitoring/anomaly detection of power electronics in VFDs. Nevertheless, Zhang and Yang (2023) have recently presented a prototype Long Short-Term Memory-Based Variational Autoencoder Wasserstein Generation Adversarial Network for anomaly detection in wind turbines based on supervisory control and data acquisition (SCADA) data involving wind speed, output power and generator input/output shaft temperature among others. The authors achieved TPR above 90%, PPV between 50% and 70% and F1 score between 0.7 and 0.8, all of which comparable with the data presented in this study (Zhang and Yang, 2023). Recently, Dai et al. (2024) presented a novel Wave-GAN architecture that scored TPR, ACC and F1 of over 90% using open-source datasets. It should be noted that Zhang and Yang (2023) and Dai et al. (2024) used a rather large GAN architectures with tens of thousands of parameters, utilising 2D convolutional layers and intermediate transformation of data to images using, for example, the wavelet transform. In contrast, in this study, we managed to prepare a model with less than 30 parameters in the critic, without intermediate transformation of the time series data, and with memory-efficient 1D convolutional layers.

Figure 4.

(a) Training curve and (b) tSNE plot for the generated vs. real data for the M1 WGAN model.

Figure 5.

Values of critic-based OBAS for the VFD system affected by (a) heatsink; (b) outlet; (c) heatsink clogging and (d) fan blockage. The time periods during which the system operates at ambient temperatures of 25 C, 35 C and 45 C are shaded in green. Anomalous operations due to clogging are highlighted with magenta boxes. The letter ‘A’ indicates ambient temperature, while ‘C’ denotes the clogging/blockage level (50% vs. 75%). OBAS, out of bounds anomaly scores; VFD, variable frequency drive.

Figure 6.

Analysis of the minimal clogging/blockage level detectable in the analysed VFD under different ambient conditions: (a) 25 C, (b) 35 C, and (c) 45 C. The letter ‘C’ indicates the clogging/blockage level for which the data are presented. OBAS, out of bounds anomaly scores; VFD, variable frequency drive.

In this study, we demonstrated the feasibility of developing a 26-parameter critic-based model for clogging detection in drives. The key feature of this model is not its complexity, but the feedback-based training approach, which mimics a human-like learning and inference process. The model proved effective in detecting fan failure and heatsink/inlet clogging, although it was less successful in detecting outlet clogging. The performance metrics achieved were comparable to other developed solutions, with top scores obtained for both fan failure and heatsink clogging detection. Additionally, we showed that the minimum detectable clogging level ranged between 10% and 30%, depending on the scenario. It was also shown that clogging detection is successful during the startup or cooling phases of drive operations, with moderate outcomes noted during active operation. Overall, this study demonstrates that it is possible to develop a low-parameter WGAN model for clogging detection in drives and paves the way for generative AI models for anomaly detection in electric machines.