Utilizing TGAN and ConSinGAN for Improved Tool Wear Prediction: A Comparative Study with ED-LSTM, GRU, and CNN Models

Abstract

Developing precise deep learning (DL) models for predicting tool wear is challenging, particularly due to the scarcity of experimental data. To address this issue, this paper introduces an innovative approach that leverages the capabilities of tabular generative adversarial networks (TGAN) and conditional single image GAN (ConSinGAN). These models are employed to generate synthetic data, thereby enriching the dataset and enhancing the robustness of the predictive models. The efficacy of this methodology was rigorously evaluated using publicly available milling datasets. The pre-processing of acoustic emission data involved the application of the Walsh-Hadamard transform, followed by the generation of spectrograms. These spectrograms were then used to extract statistical attributes, forming a comprehensive feature vector for model input. Three DL models—encoder-decoder long short-term memory (ED-LSTM), gated recurrent unit (GRU), and convolutional neural network (CNN)—were applied to assess their tool wear prediction capabilities. The application of 10-fold cross-validation across these models yielded exceptionally low RMSE and MAE values of 0.02 and 0.16, respectively, underscoring the effectiveness of this approach. The results not only highlight the potential of TGAN and ConSinGAN in mitigating data scarcity but also demonstrate significant improvements in the accuracy of tool wear predictions, paving the way for more reliable and precise predictive maintenance in manufacturing processes.

1. Introduction

Metal cutting, often known as machining, is the removal of undesired material from a piece of metal. This is done using a cutting tool and includes a range of techniques used to fabricate metal objects of various forms and dimensions [1]. Its significance spans across a spectrum of manufacturing activities, from small-scale operations to large industrial applications. Among these, milling processes are particularly vital, as they transform raw materials into precisely shaped components by utilizing rotary cutters to remove material from workpieces, thereby achieving the desired geometries and dimensions. These versatile processes are integral to industries such as aerospace and automotive manufacturing, where they play a critical role in producing components with precise shapes, sizes, and surface finishes [2]. The importance of metal cutting in manufacturing has led to extensive research, particularly in the area of tool wear prediction, which is crucial for maintaining product quality and operational efficiency. Tool wear prediction models and methodologies can be broadly categorized into two approaches: physics-based and sensor-based approaches. A physics-based approach involves the development of mathematical models that accurately represent the underlying physical phenomena occurring during the machining process. These approaches require a deep understanding of the mechanics and physics of cutting, allowing for precise predictions of tool wear based on theoretical principles [3,4,5]. Conversely, sensor-based approaches rely on the continuous acquisition of real-time data throughout the machining process to monitor and predict tool wear. These approaches utilize various sensing technologies, such as force sensors, temperature sensors, vibration sensors, and acoustic emission sensors. During machining operations, the sensors capture critical data related to temperature variations, vibrations, cutting forces, and acoustic emissions, which are then analyzed to assess and predict tool wear [6]. By integrating sensor data with advanced analytical techniques, the sensor-based approach offers a dynamic and responsive approach to tool wear monitoring, enabling real-time adjustments and improvements in manufacturing processes.

In order to identify patterns or trends connected to tool degradation, the sensor data is evaluated using statistical techniques, signal processing algorithms, and machine learning (ML) approaches [7,8]. Sensor-based solutions enable continuous tool monitoring, facilitating customizable machining operations and rapid diagnostics of tool wear. The effectiveness and efficiency of machining operations are heavily dependent on the regular assessment of tool condition. Frequent monitoring allows operators to promptly detect wear, damage, or anomalies in cutting tools, enabling them to make timely adjustments, such as optimizing machining parameters or replacing the tool as needed [9,10,11]. Important factors that function as indications of the state of the tool include signals obtained from cutting forces, vibrations, acoustic emissions, and temperature changes. Variations in these signals may serve as indications of abnormality in machining procedures, such as the deterioration of tools or breakage of tools [12,13,14]. ML algorithms play a crucial role in analyzing data collected during the machining process to monitor the state of tool wear. These algorithms excel at anomaly detection, making them highly effective in reliably identifying deviations from normal operational behavior [15,16]. Moreover, since ML models are always learning, they can improve with time. Since the algorithm has a greater understanding of the relationships between tool status and signal structures, the more data it collects, the more reliable its predictions become. Several studies have been carried out to examine the relationship among tool condition and machining variables [17,18]. The investigation of vibration signals collected from a milling machine was carried out by Kumar et al. [19]. The decision tree technique was used by the researchers in their study to identify important features and create a feature vector. Furthermore, based on the selected features, an artificial neural network (ANN) was used to predict tool wear. Zhou et al. [20] proposed a different approach to monitor the state of milling tools by utilizing a wireless vibration sensing tool holder and support vector machine (SVM) algorithm. Manwar et al. [21] utilized cutting force signals for toll condition assessment in micromilling using LSTM. In another study, Abdeltawab et al. [22] applied wavelet transforms to force signals and applied hybrid DL models for condition monitoring in the milling process. In their study, Dahe et al. [23] employed vibration signals and the random forest algorithm to forecast the state of a tool. Doukas et al. [24] extracted data from multiple sensors for the prediction of tool wear in the milling process. Cai et al. [25] developed a novel information system that uses a Long Short-Term Memory (LSTM) model to predict tool wear. The method entails employing a stacked LSTM to extract profound characteristics from the time series data collected by multiple sensors. Furthermore, Zhao et al. [26] employed an innovative deep neural network architecture designed to extract robust and informative local features from sequential input data. The experimental results indicate that this approach is highly effective in predicting the state of tool wear. Marinescu and Axnite [27] conducted a study to assess the efficacy of acoustic emission sensor data and highlighted its superior accuracy and higher level of detail compared to cutting-force measurements. Moreover, Kulandaivelu et al. [28] employed acoustic emission signals for monitoring tool wear for milling operations. The results of this research demonstrate the important influence that signals—especially those above 200 kHz—had a significant effect on TCM. Despite the fact that TCM has been the subject of several studies using various theories, further research in this field remains necessary for plenty of reasons.

Alternative methods such as variational autoencoders (VAEs), Wasserstein GANs (WGANs), and traditional data augmentation techniques were not considered in the present study due to the following reasons: VAEs were evaluated for their ability to generate synthetic data, but they were found to be less effective in capturing the fine-grained details necessary for accurate tool wear prediction. This limitation is particularly critical in TCM, where the precision of generated data directly impacts the model’s ability to predict subtle wear patterns. Wasserstein GANs (WGANs) were also considered due to their improved training stability over traditional GANs, which makes them attractive for generating high-quality synthetic data. However, WGANs did not offer significant advantages over TGAN in the context of tabular data generation, especially when considering the specific requirements of this study, such as the need to preserve the complex relationships inherent in TCM datasets. Traditional data augmentation techniques, such as rotations, flips, and noise addition, were also explored. While these methods are effective for basic image transformations, they were deemed insufficient for generating the diverse and high-fidelity data required for this study. These methods lack the ability to introduce the level of variability and detail needed to train models for complex TCM tasks, where the subtle nuances in data can be crucial for accurate predictions.

The manuscript addresses a critical challenge, which is the lack of experimental data, which is widely recognized as a major impediment to the development of trustworthy models for tool wear prediction [29]. This results in the presentation of a novel framework that considerably advances the field of tool condition monitoring (TCM). The significant contributions of proposed methodology as per available literature are as follows:

• By combining two generative models, conditional SinGAN (ConSinGAN) and tabular GAN (TGAN), the study presents a novel method. This combination of methods solves the problem of limited data, which is a major issue in ML model training, and marks significant progress in the area of TCM.
• To produce more spectrograms, ConSinGAN, one of the most sophisticated DL models, is used. This feature facilitates the development of DL models, which makes it particularly useful in situations when there is a lack of image data.
• In addition to introducing novel generative models, the framework incorporates them with well-known models like CNN, GRU, and ED-LSTM. The intricate and diverse model structure that emerges from this integration is well matched to the intricate complexity of tool wear prediction.
• The proposed approach has been thoroughly tested using publicly available milling datasets from NASA’s Prognostics Center of Excellence Data Repository. The experimental results demonstrate that the integrated approach significantly improves prediction accuracy and establishes a foundation for more effective TCM systems across several industries.

The paper’s structure is as follows: A thorough explanation of the working approach is provided in Section 2, along with a thorough explanation of the models used in the study. Section 3 then discusses the analysis of the results, and Section 4 contains the proposed methodology outcomes. Figure 1 represents the schematic flowchart of the proposed methodology.

2. Materials and Methods

2.1. Dataset

To assess the efficiency of the proposed technique, a series of tests were conducted using publicly accessible milling datasets collected from the prestigious NASA Prognostics Centre of Excellence-Data Repository [30]. Careful monitoring of tool wear is essential to guarantee that the produced components fulfill the required specifications and quality requirements. A milling machine was used to conduct face milling trials using a wide variety of machining settings.

Table 1. Process Parameters.

Parameter	Value
Depth of cut	1.5 mm & 0.75 mm
Feed Rate	0.5 mm/rev & 0.25 mm/rev
Material of Workpiece	Cast Iron & Stainless Steel J45

displays the testing settings, which included a cutting speed of 200 m/min, a feed rate range of 0.5 to 0.25 mm/rev, and a cutting depth variation of 1.5 to 0.75 mm.

The milling experiments were performed utilizing a 70 mm face mill that was equipped with six KC710 inserts coated with TiC, TiC-N, and TiN to enhance toughness. Two designated locations—the table and the spindle—were used to collect data using sound emission sensors. KC710 refers to a specific grade of carbide cutting tool material and is optimized for metal cutting operations, providing a balance between toughness and wear resistance, making it suitable for various machining applications. The purpose of this study is to investigate spindle acoustic emission measurements acquired while milling a cast iron workpiece. Because spindle signals are directly connected to the cutting tool’s contact with the workpiece, they are an important component of tool wear monitoring systems. Due to their high sensitivity, these signals are perfect for use as markers to track the amount of tool wear. This characteristic aids researchers in gaining a deeper and more comprehensive understanding of the dynamic behavior shown by the cutting tool throughout various machining processes. The total number of 12 runs, as outlined in

Table 2. Experimental cases considered [30].

Case	Run	DOC (mm)	Feed (mm/rev)	Flank Wear (mm)
1	1	1.5	0.5	0
1	2	1.5	0.5	0.28
1	3	1.5	0.5	0.44
2	1	0.75	0.5	0.08
2	2	0.75	0.5	0.22
2	3	0.75	0.5	0.55
3	1	0.75	0.25	0
3	2	0.75	0.25	0.23
3	3	0.75	0.25	0.55
4	1	1.5	0.25	0.08
4	2	1.5	0.25	0.31
4	3	1.5	0.25	0.49

, was carefully determined to ensure a comprehensive exploration of the impact of varying machining parameters on tool wear in different scenarios. The number of runs was designed to cover a range of conditions that are representative of typical industrial milling operations, thereby providing sufficient data to validate the proposed models for tool condition monitoring (TCM). Face milling trials were conducted using a milling machine under various machining settings, as outlined in Table 1. The trials involved a cutting speed of 200 m/min, feed rates ranging from 0.5 to 0.25 mm/rev, and cutting depths between 1.5 and 0.75 mm, with workpieces made from cast iron and stainless steel J45. Four distinct scenarios were tested, each representing a unique combination of depth of cut (DOC) and feed rate: Scenario 1 with DOC of 1.5 mm and feed rate of 0.5 mm/rev, resulting in flank wear values of 0, 0.28, and 0.44 mm across three runs; Scenario 2 with DOC of 0.75 mm and feed rate of 0.5 mm/rev, producing flank wear values of 0.08, 0.22, and 0.55 mm; Scenario 3 with DOC of 0.75 mm and feed rate of 0.25 mm/rev, yielding flank wear values of 0, 0.23, and 0.55 mm; and Scenario 4 with DOC of 1.5 mm and feed rate of 0.25 mm/rev, resulting in flank wear values of 0.08, 0.31, and 0.49 mm. These scenarios were meticulously designed to analyze the impact of varying DOC and feed rates on tool wear during the milling process, providing valuable insights into the performance of the proposed model under different machining conditions. With sufficient variability in the input conditions (DOC and Feed Rate) and the corresponding outputs (Flank Wear), the models can generalize well across different machining conditions. This number of runs is adequate for capturing the complex, nonlinear relationships between machining parameters and tool wear, which is critical for the development of reliable predictive models. Figure 2 exhibits the acoustic emission signal plots captured at varying feed and DOC.

2.2. Signal Processing and Spectogram

The Walsh–Hadamard transform (WHT) is a mathematical technique used in signal processing and data analysis [31]. It is a signal processing technique that converts a signal from its frequency domain equivalent to its time domain counterpart. To aid in signal analysis, the working mechanisms of the WHT employ a set of orthogonal basis functions called the Walsh functions. The transformation is primarily an example of fast and efficient computation. Moreover, the Walsh functions used in the transform demonstrate orthogonality, indicating little mutual interference or overlap. The orthogonality criteria allowed for the efficient assessment of spectra and noise reduction. Preprocessing data as part of the WHT implementation results in spectrograms that properly predict tool wear based on the signals in the study. The transformation is advantageous in various application fields due to its many features. Figure 3 displays a variety of spectrograms generated using the WHT technique. A single set of operating parameters yielded 12 spectrograms.

2.3. Data Generation Using GAN

The two essential components of a generative adversarial network (GAN), a type of DL model, are a generator and a discriminator. The process involves creating fictitious data, including images, with the generator’s assistance by generating random noise. Training occurs in a competitive manner between the discriminator and generator, who want to improve their ability to distinguish genuine data from created data and the generator’s goal of creating data that is almost similar to actual data [32,33]. GAN could enhance generalization and model performance by producing distinct samples. GANs may be used to create synthetic data with variations, diversity, and challenging circumstances. Models are thus more able to pick up reliable representations and adapt to changing conditions in the real world. The next section provides a detailed analysis of the architectural designs used in TGAN and ConSinGAN.

2.3.1. ConsinGAN

ConSinGAN is a unique generative model designed exclusively for unsupervised learning, first presented by Hinz et al. [34]. This new approach to image generation differs from traditional GANs in that it achieves improved image synthesis of very high quality, which eliminates the need for large datasets [35]. ConSinGAN is very useful in situations like these since it can still perform well regardless of the absence of paired training data or with very little of it. ConSinGAN is a hierarchical generative model that can generate a broad variety of realistic variants while preserving the original structure and visual qualities by using contextual information and self-similarity from a single input picture. The model operates on many scales, with different generator and discriminator networks used at each size. These networks, operating at different resolutions, are able to comprehensively collect different properties in an efficient way. More detailed and precise pictures may be produced by controlling the higher-resolution counterpart with a lower-resolution generator. The architectural layout of ConSinGAN is shown in Figure 4. ‘Stage 0’ is where the model training starts using a low-power generator and low-resolution photos. The capacity of the generator and the picture resolution both rise with the number of steps. By creating 200 photos from each spectrogram, a total of 2400 spectrogram images were created for the present investigation. An example of a created spectrogram is shown in Figure 4.

2.3.2. Tabular Generative Adversarial Networks (TGAN)

TGAN concentrates on creating realistic tabular datasets, as opposed to traditional GANs, which are mostly used to generate pictures or sequential data [36,37]. The input and output formats of TGAN and GAN are the primary areas of distinction between them. Probabilistic input vectors are often used as the generator’s input in images-based GANs, where the resulting vectors are converted into image outputs. However, TGAN makes use of a random noise vector, also known as a feature vector, to develop a generated tabular dataset that simulates the statistical characteristics and organization of real data collected via experimentation. The main objective of TGAN is to contribute to making it possible for the generator network G(z) to be trained to create tabular data points to the point where D(z) is unable to discriminate between produced and genuine data.

Let T be a table comprising of Cn continuous random variables denoted as $C_n$ and $D_n$ represent discrete random variables denoted as $\left\{ D_1, D_2, D_3. . . , D_n\right\}$. The joint distribution of variables $C_1$ through $C_n$ and $D_1$ through $D_n$ is denoted as $P\left(C_1: C_n, D_1: D_n\right)$. The table comprises rows that correspond to independent samples drawn from the joint distribution. Each row is denoted by lowercase values $\left\{c_1, j, . . . , cn ,j, d_1,j, . . . , dn, j\right\}$ where j represents the index of the sample. The aim is to acquire knowledge from a generative model, labeled as $M \left(C_1: C_n, D_1: D_n\right)$, which can generate synthetic samples. The construction of the generator involved the utilization of LSTM, which incorporated a hidden vector $a_t=tanh\left(E_t{h}_i\right)$. Here, $h_i$ represented the LSTM output, while $E_t$ was a learned parameter of the network. The inputs provided to the LSTM included the random noise C, the hidden vector a, and the weight vector b. The computation of the output for discrete variables involved the determination of $w_i=x_i=y_i=SoftMax \left(E_t{a}_t\right)$. The cross-entropy loss function was employed in conjunction with the Kullback–Leibler divergence. The construction of $D\left(z\right)$ involved the utilization of MLP, which comprises n layers. The inputs, namely $w_{1:n}, x_{1:n}$ and $y_{1:n}$ were concatenated. The initial and ith layers were calculated in the following manner:

$$a_1=LeakyReLU\left(BN\left(E_1^{\left(k\right)}\left(w_{1:n} ⨁x_{1:n}⨁y_{1:n} \right)\right)\right)$$

(1)

$$a_i=LeakyReLU\left(BN\left(E_i^{\left(k\right)}\left(a_{i-1}^{\left(k\right)} ⨁diversity\left(a_{i-1}^{\left(k\right)} \right)\right)\right)\right),i=2:n$$

(2)

where ⊕ is the concatenation operator, E1 is the learned parameter, batch normalization is BN (.), and the activation function is Leaky ReLU. The generator is optimized using KL divergence as:

$$G_l={-E}_{l~N\left(\mathrm{0,1}\right)}logD\left(G\left(l\right)\right)+{\sum }_{i=1}^{n}KL\left(x_i^{\prime },x_i\right)+{\sum }_{i=1}^{n}KL\left(y_i^{\prime },y_i\right)$$

(3)

Similarly, using conventional cross-entropy loss, discriminator is optimized as:

$$D_l=-E_{w_{1:n},x_{1:n},y_{1:n} ~P\left(T\right)}logD\left(w_{1:n},x_{1:n},y_{1:n}\right)+E_{l~N\left(\mathrm{0,1}\right)}logD\left(G\left(l\right)\right)$$

(4)

2.4. Feature Extraction

Feature extraction is the process of identifying and extracting important information or features from raw data [38]. In the context of tool condition monitoring, where enormous quantities of sensor readings are frequently involved, feature extraction is critical in translating complex raw data into a more comprehensible and understandable representation. These retrieved features are used as variables in ML models or other types of analysis. The features extracted from spectrograms in the current study are listed in

Table 3. Statistical features extracted from generated spectrograms.

Sr. No.	Feature	Sr. No.	Feature
1	Root Mean Square Error (RMSE)	7	Variance
2	Peak Signal-to-Noise Ratio (PSNR)	8	Mean
3	Mean Absolute Error (MAE)	9	Standard Deviation (STD)
4	Entropy	10	Mean Squared Error
5	Structural Similarity Index Measure (SSIM)	11	Erreur Relative Globale Adimensionnelle de Synthèse (ERGAS)
6	Kurtosis

. A feature vector of size 2400 × 11 was created. This feature vector is then used to train the TGAN model, which, in later stages, feeds to the DL models for prediction.

2.5. Deep Learning Models

Deep learning, a branch of artificial intelligence, uses multi-layered neural networks to learn and predict data, eliminating the need for manual feature engineering typical of traditional ML algorithms. Its ability to automatically learn features makes it especially valuable for complex tasks like condition monitoring, fault diagnosis, and tool wear prediction. This study examines three models—GRU, CNN, and ED-LSTM—for accurate tool wear prediction.

2.5.1. Gated Recurrent Unit (GRU)

The GRU is a recently developed architecture for sequence prediction, designed to optimize the flow of information in sequential data through its gating mechanisms. Unlike traditional LSTM models, which use three gates, the GRU combines the input and forget gates into a single update gate, streamlining the structure. This design also includes a reset gate, which captures transient dependencies within sequences, helping the model retain relevant contextual information. Due to its simpler structure with just two gates, the GRU is more computationally efficient than LSTM, leading to shorter training times and lower memory usage, making it ideal for scenarios with limited computational resources [39]. This efficiency, combined with its strong performance, makes GRUs well-suited for a wide range of sequence-related tasks in machine learning. Figure 5 shows the architecture of the GRU model.

2.5.2. Convolutional Neural Network (CNN)

A significant advancement in artificial neural networks is the development of convolutional neural networks (CNNs), particularly designed for grid-like inputs such as images. CNNs have transformed fields like image recognition and computer vision by automatically generating hierarchical representations from raw data. The architecture consists of convolutional layers that detect local patterns, pooling layers that reduce dimensionality while retaining essential features, and fully connected layers that enable comprehensive learning and decision-making. This structure allows CNNs to effectively analyze complex relationships and make accurate predictions based on the extracted features [40]. CNN’s main characteristic is its ability to automatically generate hierarchical representations from raw input data. The architecture is composed of a series of carefully designed linked layers that are meant to extract and assess hierarchical representations from the incoming data. By achieving hierarchical representation learning over the previous layers, the network is able to comprehend complex connections and provide predictions based on the characteristics that were extracted.

2.5.3. Encoder Decoder-LSTM (ED-LSTM)

The ED-LSTM is an advanced variant of the classic LSTM model, particularly suited for sequential data. Its architecture consists of two main components: the encoder and decoder networks. The encoder reads the input sequence and converts it into a fixed-length vector representation, which the decoder then uses to generate the output sequence, one token at a time. LSTM cells are typically used in both the encoder and decoder, with the encoder’s final hidden state containing the compressed data of the entire input sequence. The decoder gradually generates the output sequence by using this context vector and the previous hidden state at each step. The ED-LSTM architecture is widely applicable in fields like machine translation, photo captioning, and speech recognition [41,42]. Figure 6 illustrates the ED-LSTM model’s structure.

Table 4. Model parameters of deep learning algorithms.

Parameter	GRU	CNN	EDLSTM
Number of Layers	2 layers (1 GRU layer, 1 Dense layer)	7 layers (2 Conv1D layers, 2MaxPoolind, 1 Flatten layer, 2 Dense layers)	4 layers (2 LSTM layers, 1 Repeat Vector layer, 1 Time Distributed Dense layer)
Units	GRU Layer: 512 units Dense Layer: 1 unit	Conv1D Layers: 64, and 128 filters respectively Dense Layers: 128 and 1 units respectively	LSTM Layers: 256 units in the first LSTM layer, 128 units in the second LSTM layer Time Distributed Dense Layer: 1 unit per time step
Layer Types	GRU, Dense	Conv1D, MaxPooling1D, Flatten, Dense	LSTM, Repeat Vector, Time Distributed
Activation Functions	GRU Layer: Tanh for the activation of the cell state and Sigmoid for the update and reset gates. Dense Layer: Linear	Conv1D Layers: ReLU (Rectified Linear Unit) Dense Layer: ReLU and Linear	LSTM Layers: Tanh for the LSTM cell state and Sigmoid for the LSTM gates Time Distributed Dense Layer: Linear
Optimizers	RMSprop	Adam	Adam
Loss Function	Mean Absolute Error	Mean Squared Error	Mean Squared Error
Learning Rate	0.001	0.001	0.001
Batch Size	32	32	32
Epochs	100	100	100

represents the details about the model parameters considered in this study. The parameters were selected based on the need to balance model complexity with computational efficiency, ensuring that each model could effectively capture the relevant patterns in the data. The choices, such as the number of layers, units, and activation functions, were guided by best practices for each model type to optimize prediction accuracy while preventing overfitting. The consistency in optimizers, learning rates, and batch sizes was maintained to facilitate stable and efficient training across all models. The models were implemented using Python, utilizing libraries such as TensorFlow and Keras for DL and Scikit-learn for data preprocessing and model evaluation. TensorFlow and Keras were chosen for their flexibility and robust support for developing and deploying models using the Google Colab interface, which has an NVIDIA Tesla T4 GPU with 16 GB of VRAM.

WHT was employed primarily for its efficiency in converting time-domain signals into a form that emphasizes key features critical for tool wear analysis. This transformation enhances the feature space, making it easier to distinguish subtle wear patterns that might otherwise go unnoticed. The generation of spectrograms, despite adding an additional layer of processing, was crucial for capturing the time-frequency characteristics of the tool wear data. This step provides a richer signal representation, allowing our models, particularly CNNs, to leverage both temporal and spectral features. This dual representation significantly enhances the predictive accuracy of the models, as tool wear typically manifests in both time and frequency domains. While it is acknowledged that these preprocessing steps introduce additional computational complexity, the benefits in terms of improved prediction accuracy justify the costs. To address concerns about real-time applicability, the implementation has been optimized by parallelizing the WHT and spectrogram calculations on GPU-accelerated hardware, which substantially reduces processing time. Additionally, the use of reduced-resolution spectrograms and selective application of the WHT based on signal characteristics are being explored to further lower computational demands. These optimizations, combined with the inherent parallelizability of the operations, suggest that the proposed preprocessing pipeline could be feasible for real-time industrial applications, particularly in environments equipped with modern computational resources.

3. Results and Discussion

The current study centers on forecasting tool wear by analyzing acoustic signals obtained from a face milling machine [30]. In the signal processing phase, wavelet hash transform (WHT) was implemented on the acquired data, leading to the creation of a spectrogram using the transform coefficients. The ConSinGAN model was then employed for the generation of a multitude of spectrograms derived from the original representations. From the generated spectrograms, a feature vector is created by extracting 11 features (Table 3), which in later stages are utilized to build synthetic feature vectors using TGAN. The feature vectors were utilized to train various models: GRU, CNN, and ED-LSTM models, for the purpose of predicting tool wear.

Table 5. Statistical comparison of features: RMSE, PSNR, MAE, entropy.

	RMSE		PSNR		MAE		Entropy
	Original Feature	Generated Feature	Original Feature	Generated Feature	Original Feature	Generated Feature	Original Feature	Generated Feature
Mean	18.33	20.31	26.06	25.44	133.90	133.46	4.89	4.57
Std	25.47	27.66	5.66	6.12	9.80	10.51	2.13	2.37
Min	9.18	9.39	7.92	7.92	109.61	111.45	0.00	0.00
25	9.55	9.60	27.58	27.36	130.19	129.16	5.76	5.73
50	10.09	10.32	28.05	27.80	133.84	133.54	5.83	5.81
75	10.65	10.84	28.53	28.46	137.89	137.79	5.89	5.88
Max	102.47	102.48	28.87	28.66	153.09	153.09	6.00	5.95

,

Table 6. Statistical comparison of features: SSIM, Kurtosis, Variance, Mean.

	SSIM		Kurtosis		Variance		Mean
	Original Feature	Generated Feature	Original Feature	Generated Feature	Original Feature	Generated Feature	Original Feature	Generated Feature
Mean	0.64	0.64	28.18	31.01	224.43	207.27	175.21	173.73
Std	0.03	0.032	64.01	71.53	99.38	109.64	21.25	23.24
Min	0.58	0.58	0.00	0.00	0.00	−1.52	105.00	105.00
25	0.62	0.62	9.07	9.24	249.34	238.34	179.28	179.22
50	0.63	0.63	9.83	10.00	265.90	260.80	181.21	181.04
75	0.65	0.65	10.66	10.85	277.44	275.10	182.86	182.66
Max	0.73	0.73	246.10	247.71	305.55	288.68	186.00	185.83

and

Table 7. Statistical comparison of features: STD, MSE, ERGAS.

	STD		MSE		ERGAS
	Original Feature	Generated Feature	Original Feature	Generated Feature	Original Feature	Generated Feature
Mean	13.88	13.03	984.67	1081.54	2159.75	2392.29
Std	5.63	6.21	2869.89	3133.09	3168.95	3445.82
Min	0.00	0.00	84.28	86.90	1016.80	1039.01
25	15.79	15.47	91.24	90.81	1072.50	1071.69
50	16.31	16.21	101.84	103.52	1137.42	1174.85
75	16.66	16.60	113.45	116.26	1254.73	1267.75
Max	17.48	17.04	10,500.86	10,500.86	12,643.56	12,643.56

present statistical comparisons of the features extracted from ConSinGAN images (referred to as original features) and the features generated by TGAN (referred to as generated features).

Table 5 presents a statistical analysis of the original and newly developed attributes. There are considerable changes in various parameters when comparing the created and original characteristics. The RMSE slightly increased from 18.33 to 20.31, indicating a marginal reduction in prediction accuracy due to the variability introduced during data augmentation. The PSNR decreased from 26.06 to 25.44, suggesting a minor decrease in the fidelity of the generated data, though it remains within an acceptable range. The MAE showed a slight decrease from 133.90 to 133.46, reinforcing the close similarity between the generated and original data. The entropy values decreased from 4.89 to 4.57, indicating less complexity in the generated features, which could simplify model training. The standard deviation analysis highlights increased variability in the generated data across all metrics, which is essential for improving model generalization and robustness in real-world conditions. The generated features have a small downward inclination, as seen by the consistent trend of percentile numbers at the 25th, 50th, and 75th percentiles. Comparisons among maximum values may be detected, with RMSE and PSNR exhibiting relatively small changes. Despite these variations, the changes are within acceptable limits, underscoring the effectiveness of the TGAN and ConSinGAN frameworks in enhancing tool wear prediction by introducing beneficial variability and maintaining overall data consistency. The metrics of the generated and original features—SSIM, kurtosis, variance, and mean—are compared in Table 6. There is agreement between the produced and original features, according to the SSIM mean values (0.64). Kurtosis, a gauge of the distribution’s tail heaviness, increases in the produced version from 28.18 in the original to 31.01, suggesting a shift in the direction of a heavier tail. The variance of the produced features has significantly decreased, from 224.43 to 207.27, recommending that the range of values has shrunk. The derived features’ mean values demonstrate a modest reduction from the original 175.21 to 173.73. Standard deviation values show rising values for SSIM, kurtosis, variance, and mean among the created features. As long as SSIM and Kurtosis are constant between the original and generated features, minimum values exhibit comparable trends. The variance, with the produced number at −1.52 and the initial value at 0.00, shows an unexpected disparity. All metrics demonstrate that the highest values among the original and produced characteristics are still rather near, apart from minor variations at the 25th, 50th, and 75th percentiles.

Similar statistical analyses are shown for three more important features (STD, MSE, and ERGAS) in Table 7. The average values of all three attributes are 13.88, 984.67, and 2159.75, respectively. Furthermore, the features that were produced show average STD values of 13.03, MSE values of 1081.54, and ERGAS values of 2392.29. The data suggest a minor decrease in the STD and increase in both MSE and ERGAS for the produced features as compared to the original features. Standard deviations for the generated features are 6.21 for STD, 3133.09 for MSE, and 3445.82 for ERGAS. Standard deviation values for STD, MSE, and ERGAS in the original features, on the other hand, are 5.63, 2869.89, and 3168.95, respectively. The produced features’ larger standard deviation values indicate a greater level of unpredictability when compared to the original features. The little fluctuations in the 25th, 50th, and 75th percentiles demonstrate how the data’s distribution and central tendency have altered after switching to the quartiles. The TGAN characteristics typically show minor variations in comparison to the original features, as can be seen by comparing the three tables and looking at the statistical values. Notably, certain original qualities have been successfully captured and recreated by the generating process. A graphic depiction of the distribution of TGAN’s original and derived features is shown in Figure 7a–l. This figure displays a comparison of the histogram and kernel density estimation (represented by red and blue lines respectively) between the original features and the features generated by TGAN.

DL models have received a lot of attention in the field of tool wear monitoring because of their capacity to automatically learn and extract intricate patterns from difficult data. To comprehend and represent hierarchical aspects in the input data, these models use neural networks with deep architectures. In the proposed work, feature vectors created by the ConSinGAN-generated images and the TGAN-generated features, are evaluated from three models: GRU, CNN, and ED-LSTM for the purpose of tool wear prediction. These models are trained on the dataset of original and generated features to effectively capture the variations and patterns present in both sets of features. The trained models are then utilized to make accurate tool wear predictions based on testing and 10-fold cross validations. Initially, the feature vectors are partitioned in a conventional 70:30 ratio, whereby 70% of the data is allocated for model training and 30% reserved for model testing. Further to reduced biasedness due to random split of data, 10-fold cross validation results are evaluated. The present study assessed the efficacy of three models in predicting tool wear through the utilization of two performance metrics, specifically RMSE and MAE. Figure 8a,b compares the training and testing outcomes from models for tool wear prediction. The algorithms’ performance was evaluated using both real and artificially generated feature vectors. When trained on real feature vectors, the CNN model achieves an RMSE of 0.029 and an MAE of 0.020. Nonetheless, after training with a TGAN-generated feature vector, the RMSE rises to 0.048, while the mean absolute error MAE rises to 0.031. The results of the testing show that the CNN model performs slightly better on real feature vectors to predict tool condition monitoring, as evidenced by its lower RMSE of 0.028 and MAE of 0.020 when compared to the generated feature vector, which yielded an RMSE of 0.053 and an MAE of 0.035. Similarly, in the case of the GRU model, there is a significant variation in tool wear prediction between the actual feature vector and the generated feature vector. The GRU model has an RMSE of 0.062 and an MAE of 0.048 after being trained with the original feature vector. Nonetheless, when trained with the resultant feature vector, the RMSE increases to 0.094, while the MAE increases to 0.067. During the testing phase, the GRU model produced equivalent results for real and created feature vectors, as shown by an RMSE of 0.062 and MAE of 0.047 and 0.066, indicating a superior tool wear prediction accuracy. The ED-LSTM model predicts with rather constant performance when applied to both actual and generated feature vectors, with an RMSE of 0.192 and an MAE of 0.164 when trained with real feature vectors. Similarly, after being trained with the given feature vector, it achieves an RMSE of 0.191 and an MAE of 0.165. During the testing phase, the ED-LSTM model performed similarly on both real and created feature vectors, as demonstrated by RMSE values of 0.197 and 0.193, and MAE values of 0.169 and 0.167, respectively. To summarize, the CNN model has demonstrated superior performance in tool wear prediction compared to other models. It has consistently achieved the lowest RMSE and MAE values during both training and testing phases, irrespective of data type. The GRU model’s effectiveness is below optimal, whereas the ED-LSTM model consistently demonstrates the highest RMSE and MAE values, indicating lower accuracy in predicting tool wear.

Simply evaluating a DL model’s dependability based on how well it performs during training and testing may not be accurate. This is due to the fact that accurate data utilized in both the training and testing phases of a model’s life is critical to its performance. It’s possible that these data don’t fully capture the dataset or that they can’t be very broadly applied to brand-new, unidentifiable data. To overcome this limitation, predictive studies often use the well-known 10-fold cross-validation technique. Using 10-fold cross-validation, ten equivalent subsets, or “folds”, are produced from the dataset. The remaining folds serve as the training set, and a fresh fold serves as the testing set for each of the ten iterations of the model that are trained and assessed. Assessment robustness is improved, and the model’s predictive capability is measured more accurately by averaging scores across these ten rounds. Through a thorough study that overcomes the potential biases resulting from a single data split, this method increases confidence in the model’s performance. Figure 9a,b displays the tool wear estimation outcomes of a 10-fold cross-validation using all three models and both feature vectors. Among the three models, the CNN model shows the lowest prediction error when real and created feature vectors are evaluated. The CNN model produces an MAE of 0.943 for the created feature vector and 0.640 for the real feature vector, as seen in Figure 9a. Additionally, the CNN model achieves an RMSE of 0.063 for the original feature vector and 0.028 for the produced feature vector using the 10-fold cross-validation method. By contrast, the GRU model performs better than the ED-LSTM model but shows somewhat higher prediction errors than the CNN model. When tested on an actual feature vector, the GRU model shows an RMSE of 0.062 and an MAE of 0.776 over 10-fold cross-validation. As shown in Figure 9a,b, the created feature vector yields an MAE of 0.889 and an RMSE of 0.040. With an RMSE of 0.205 and an MAE of 0.176 for the actual feature vector and an RMSE of 0.191 and an MAE of 0.164 for the produced feature vector, as shown in Figure 9a,b the ED-LSTM model exhibits the highest prediction errors among the three models.

In summary, the GRU model trails closely behind the CNN model, which shows the lowest prediction error for both generated and actual feature vectors. The CNN model is believed to have an advantage over the GRU and ED-LSTM models in terms of tool wear prediction, which may be attributed to many important reasons. CNNs excel in capturing spatial hierarchies in data due to their convolutional layers, which can efficiently identify local patterns and features in the input signals. This makes CNNs particularly effective in tasks where the detection of intricate patterns or anomalies in the sensor data is crucial. The local receptive fields and shared weights in CNNs allow them to generalize well across varying signal conditions, leading to consistently better performance in the experiments conducted. Convolutional layers employ their innate capacity to represent intricate spatial connections to their advantage in identifying minute alterations that point to degradation. Furthermore, CNNs use convolutional kernels to exploit parameter sharing, which allows them to recognize patterns in a picture independent of where they are precisely located. This feature is particularly useful when there is tool wear, since it prevents the precise location of worn parts from changing. Additionally, in scenarios where tool wear is not largely dictated by the temporal sequence, such as image-based wear analysis, CNNs’ reduced sensitivity to sequence length relative to GRU and ED-LSTM models for sequential data becomes favorable. ED-LSTM’s ability to capture long-term dependencies in the data could explain its superior performance in certain scenarios, especially when dealing with sequences that require the model to remember and relate information over extended time periods. On the other hand, while GRU is designed to be a more computationally efficient alternative to LSTM by simplifying the gating mechanisms, this efficiency might come at the cost of reduced capacity to capture very complex temporal patterns. This trade-off could contribute to GRU’s comparatively lower performance in cases where the complexity of the temporal dependencies exceeds what the GRU architecture can efficiently model.

One of the primary reasons for utilizing GAN-based models, particularly TGAN and ConSinGAN, is their exceptional capability to handle scenarios where the available dataset is limited. In the context of tool condition monitoring, obtaining large, labeled datasets is often challenging due to the time-consuming and costly nature of data collection in industrial environments. Traditional DL models, such as ED-LSTM, GRU, and CNN, typically require substantial amounts of data to achieve high performance. In contrast, GAN models can effectively generate high-quality synthetic data to augment the limited real-world data, thus enhancing model training without the need for massive datasets. Authors agree that in real-world industrial settings, computational resources may be constrained, and the scalability of models is a crucial factor. However, the GAN models employed, specifically TGAN, have been chosen and optimized for their efficiency in scenarios where data is scarce. By generating high-quality synthetic data, TGAN allows for augmenting the limited dataset effectively, which in turn reduces the need for extensive real-world data collection—a process that is often both time-consuming and resource-intensive in industrial environments. Scalability is a multifaceted issue that touches on every aspect of model deployment, from computational resources to data management and real-time application constraints. While the models discussed in our study offer significant advantages in terms of predictive accuracy and the ability to handle limited datasets, their scalability in industrial settings requires careful consideration. By addressing these challenges through optimization, distributed computing, and strategic model deployment, it is possible to enhance the scalability of these models, making them more practical for widespread industrial use.

The authors compared their suggested technique with already published literature. Interestingly, the dataset utilized to measure tool wear was the same for all the studies referenced in

Table 8. Comparison of results with other similar works.

Reference	Sensor	Algorithm	RMSE
Hanachi et al. [43]	Current sensors	Sipos	0.42
		Adaptive neuro-fuzzy inference system (ANFIS)	0.56
		Regularized particle filter (RPF)	0.22
Yu et al. [44]	All sensors	Bi Directional LSTM	7.14
		BiLSTM-ED2	11.27
Kumar et al. [45]	Vibration sensors	Vanilla LSTM	0.1129
		Bidirectional LSTM	0.0982
		EDLSTM	0.0586
		Hybrid LSTM	0.0364
Proposed work	Acoustic sensors	Original Data
		CNN	0.0625
		GRU	0.0623
		EDLSTM	0.2049
		TGAN Data
		CNN	0.027
		GRU	0.039
		ED-LSTM	0.190

. When compared to other studies on the topic, this comparative study demonstrates the distinctive contributions and breakthroughs achieved by the proposed technique. Hanachi et al. [43] employed current sensors with models such as Sipos and ANFIS, achieving RMSE values of 0.42 and 0.56, respectively. In contrast, the proposed approach using TGAN-augmented data with CNN achieved an RMSE of 0.027, significantly lower than the values reported by Hanachi et al. This highlights the effectiveness of TGAN in generating high-quality synthetic data that enhances model performance. Yu et al. [44] utilized all available sensors and applied BiLSTM and BiLSTM-ED2, with RMSE values of 7.14 and 11.27, respectively. The proposed approach, even with the original data, outperformed these methods, with GRU achieving an RMSE of 0.0623. When TGAN data was used, the RMSE further reduced to 0.039 with GRU, demonstrating superior performance in comparison to Yu et al.’s approach. Kumar et al. [45] explored various LSTM-based models using vibration sensors, with the Hybrid LSTM achieving an RMSE of 0.0364. The proposed approach using TGAN data with CNN achieved a slightly better RMSE of 0.027. This comparison indicates that the integration of TGAN and ConSinGAN with CNN not only matches but also slightly exceeds the performance of advanced LSTM-based models, particularly when utilizing acoustic sensor data. The comparative analysis underscores the advancements made by the proposed framework in tool wear prediction. The integration of TGAN and ConSinGAN, combined with CNN and GRU, results in significantly lower RMSE values compared to other state-of-the-art methods reported in the literature. These results confirm that the proposed approach offers a substantial improvement in predictive accuracy, particularly in scenarios where traditional methods struggle to achieve similar levels of performance.

4. Conclusions

In this study, the predictive capabilities of three distinct DL models—CNN, GRU, and ED-LSTM—were comprehensively evaluated for tool wear prediction. The assessment involved a detailed analysis of the training and testing datasets, with 10-fold cross-validation employed to validate the accuracy of the predictions. Selecting appropriate evaluation metrics was crucial to demonstrating the models’ effectiveness, with RMSE and MAE identified as key indicators. The models were rigorously evaluated, and their performance was discussed in depth, leading to several important conclusions:

• The CNN model consistently exhibited superior predictive performance for tool wear compared to the GRU and ED-LSTM models during both training and testing phases.
• The 10-fold cross-validation results further underscored the CNN model’s robustness, showing significantly lower RMSE and MAE scores, highlighting its adaptability even as the GRU model presented higher prediction errors than ED-LSTM.
• Depending on the evaluation criteria and the relative importance of predicted versus actual feature vectors, the CNN and GRU models emerge as the most suitable choices for tool wear prediction.

This study successfully demonstrated the superior predictive capabilities of CNN, GRU, and ED-LSTM models for tool wear prediction, with CNN consistently outperforming the other models. The comprehensive evaluation, including 10-fold cross-validation, validated the accuracy and robustness of these models, particularly in their ability to minimize RMSE and MAE scores. The findings suggest that CNN, due to its high adaptability and precision, is particularly well-suited for applications in real-time tool condition monitoring in industrial settings. These models hold significant potential for enhancing predictive maintenance strategies, leading to reduced downtime and optimized machining processes across various manufacturing sectors. Future research should prioritize validating the proposed framework using real-world industrial datasets from diverse machining environments, moving beyond the NASA milling dataset used as a benchmark. This validation would provide a more thorough assessment of the model’s robustness and its ability to generalize across various machinery and operating conditions. Additionally, investigating the integration of additional sensors, such as thermal cameras, optical sensors, or force sensors, alongside the existing acoustic and vibration sensors, could enhance the accuracy and comprehensiveness of tool wear predictions through advanced data fusion techniques. While the current models demonstrate high accuracy, their computational complexity may limit their feasibility for real-time industrial applications. Therefore, future work could focus on developing lightweight versions of the TGAN and ConSinGAN models or optimizing them for faster inference, ensuring their suitability for real-time monitoring systems. These future research directions provide a roadmap for enhancing the current framework, ensuring its continued relevance and effectiveness in the rapidly evolving field of tool condition monitoring.