2.1. Underground Coal Mine Drilling Process

In underground coal mining, the drilling process is essential for roadway excavation, gas drainage, and geological exploration. It involves the rotation and axial advancement of the drill bit to fracture rock formation, and high-pressure flushing fluid carries cuttings out of the borehole. The process requires equipment capable of providing stable axial thrust and rotational torque under complex geological condition. An illustration of the underground coal-mine drilling process is shown in Fig. 1.

The tunnel drilling rig is the central equipment used to support this process. It integrates mechanical structures, hydraulic power, and control system to ensure safe and efficient drilling. The hydraulic system is mainly composed of two functional circuits: the feed circuit and the rotary circuit. The feed circuit controls the vertical movement of the drill bit through hydraulic cylinders, delivering the required thrust for penetration. The rotary circuit provides the torque necessary for drill bit rotation using hydraulic motors. These circuits are supplied by multiple pumps and coordinated through an integrated control console that manages the distribution of hydraulic power and monitors system performance in real time.

During operation, hydraulic pumps extract oil from the tank and pass it through suction and pressure filters. The pressurized oil then enters the control console, where it is distributed to different subsystems through load-sensing proportional valves. Pump I supplies the power head, feed cylinder, and travel motor. Pump II controls the slow feed function, the pull-up mechanism, and the manipulator. Pump III regulates the front prop cylinder, body angle adjustment, and stabilizing devices. Through the coordinated actions of these subsystems, the drilling rig achieves precise control over bit rotation, feed motion, positioning, and gripping functions, thereby supporting stable and continuous drilling under underground conditions.

2.2. Fault Analysis

The working environment of tunnel drilling rig in underground coal mines is characterized by confined spaces, high humidity, strong vibration, and complex geological conditions. Under such circumstances, the electro-hydraulic system of the rig is prone to various faults, many of which occur within internal oil circuits and are difficult to detect. These faults are often interrelated, and their symptoms can overlap, which adds further complexity to detection and maintenance.

• Sticking of valve cores in electro-hydraulic pressure relief valves: this issue is typically caused by wear, contamination, or inadequate lubrication. When the valve core becomes stuck, it impairs the system’s ability to regulate pressure effectively, leading to instability during operation.

• Internal leakage in the feed cylinder: it is often associated with degraded seals or worn piston surfaces, resulting in a loss of axial thrust and poor feed control. In severe cases, this may lead to complete failure of the drill bit’s forward movement.

• Hydraulic filter clogging: This, especially in suction or high-pressure filters, can restrict oil flow and reduce overall system pressure. This condition is usually caused by oil contamination or a lack of routine maintenance, and it can contribute to overheating and accelerated wear of hydraulic components.

Due to the strong interdependence of the feed and rotary hydraulic systems, a fault in one circuit can rapidly propagate and influence the performance of the other, leading to complex coupled failure modes. Traditional threshold-based or rule-driven detection methods struggle to capture such nonlinear interactions and often require extensive labeled fault data, which are scarce in underground coal mine operations. This motivates the need for unsupervised representation learning methods. In this paper, a sparse autoencoder is developed to model the normal operating patterns of drilling rig. The sparsity constraint encourages compact and discriminative latent representations, which in turn make subtle deviations from normal behavior easier to detect. Thus, sparse autoencoder-based fault detection offers a practical and effective way to enhance fault tolerance, enable condition-based maintenance, and ensure safe and efficient drilling operations in underground coal mines.

2.3. Scheme Design

To address the challenges of fault detection under the harsh operating conditions of tunnel drilling rig, this paper proposes a data-driven detection scheme based on sparse autoencoder. The overall framework is illustrated in Fig. 2. It consists of three main components: data denoising, normal behavior modeling for drilling rig, and online fault detection.

First, the drilling data, including feed pressure, rotary speed, and advance rate, are collected from the drilling rig. To improve the signal quality and eliminate the noise interference, some signal preprocessing methods such as outlier removal, normalization, and smoothing are applied. This step enhances the robustness of the model and reduces the influence of environmental noise on fault indicators.

Next, a normal behavior model is trained and constructed using a sparse autoencoder with multivariate time-series data under normal operating condition. This model is designed to learn the underlying patterns of healthy system behavior by minimizing reconstruction error. Sparsity constraints are imposed on the hidden layer to encourage the extraction of compact and informative features, which improves generalization and reduces overfitting.

During the online fault detection phase, sensor data are processed in sliding time windows and fed into the trained normal behavior model. The reconstruction error between the input and output signals is computed and compared against a statistical threshold derived from training data. If the reconstruction error exceeds this threshold, the system flags the current window as anomalous, indicating a potential fault.

3.1. Data Denoising

The raw data collected from drilling sensors typically include feed pressure, rotary speed, and feed rate. These time-series signals are often corrupted by environmental noise, mechanical interference, and sensor instability, especially under harsh underground mining condition. To enhance data quality and ensure the reliability of fault detection, signal preprocessing techniques are applied.

Common preprocessing steps include outlier removal, time-domain segmentation, and normalization. Additionally, to improve the signal-to-noise ratio, wavelet transform is adopted for denoising purposes. Compared to Fourier-based methods, wavelet transform offers better localization in both time and frequency domains, and it is suitable for processing non-stationary signals.

Given a discrete signal \(x(t)\), its discrete wavelet transform is defined as

3.2. Normal Behavior Model Based on Sparse Autoencoder

Due to the limitations of sensor precision and system redundancy, multivariate time series from the drilling rig under normal condition often exhibit smooth and low-rank characteristics. These properties can lead to overfitting in conventional autoencoder models. To address this, a sparse autoencoder is adopted, with sparsity constraints imposed to encourage selective activation in the hidden layer.

Conventional autoencoder is a fully connected feedforward neural network composed of an encoder and a decoder. The sparse autoencoder introduces sparsity constraints to the conventional autoencoder, inhibiting the output of some hidden layer neurons to achieve a sparse structure. Let \(X = \{X^{1}, X^{2}, \dots, X^{m}\}\) represent the collection of \(m\) multivariate time series in the drilling process. The multivariate time series \(X^{i}\) can be represented as a matrix \(X^{i} = \{x^{i}_{jt}\}\) (\(j = 1,\dots, r\); \(t = 1,\dots, n\)), Let \(X_{i} \in \mathbb{R}^{r \times n}\) denote the \(i\)-th multivariate time-series segment extracted from the drilling process using a sliding window, where \(r\) is the number of monitored drilling parameters and \(n\) is the window length. The element \(x^{i}_{jt}\) represents the value of the \(j\)-th parameter at time step \(t\) within the \(i\)-th window.

To optimize the parameters of the sparse autoencoder, the following loss function is used:

The \(L_{2}\) regularization term is defined as

The sparsity regularization term is defined as

As the discrepancy between \(\rho\) and \(\widehat{\rho_{d}}\) increases, the KL divergence also increases accordingly. After training the sparse autoencoder using multivariate time series data of the drilling process obtained via sliding windows from normal drilling history, the reconstruction error between the original and the reconstructed multivariate time series can be computed as a monitoring metric:

Here, \(h_{X}^{i}\) denotes the reconstruction error of the \(i\)-th sliding window, which is used as the monitoring indicator throughout this paper.

In fields such as image anomaly detection, reconstruction error from reconstructed images is often used as a monitoring metric. Since multivariate time series data are also matrix data, using reconstruction error within a sliding window for monitoring is a reasonable approach. The monitoring metric calculated on the training set is denoted as

For training the sparse autoencoder, we use the following practical pipeline and hyperparameters. Each multivariate time-series segment (sliding window) is denoted by \(X^{i} \in \mathbb{R}^{r \times n}\) where \(r\) is the number of sensors and \(n\) is the window length (in experiments, \(n = 60\)). In this work, we flatten \(X^{i}\) into a vector \(\mathrm{vec}(X^{i}) \in \mathbb{R}^{rn}\) before feeding it to the fully-connected encoder.

The encoder and decoder are fully-connected networks with one hidden layer of size \(N\) (hidden dimension). Encoder: \(\Theta = \sigma(W^{(1)}\mathrm{vec}(X^{i}) + b^{(1)})\), decoder: \(\widehat{\mathrm{vec}}(X^{i}) = \psi(W^{(2)}\Theta + b^{(2)})\). In our experiments, we set \(N=64\), activation functions \(\sigma = \textrm{ReLU}\) and \(\psi = \textrm{sigmoid}\).

The objective in Eq. 6 is minimized, where \(\Omega_{l}\) denotes the \(L_{2}\) weight regularization and \(\Omega_{s} = \sum_{d=1}^{N} \mathit{KL}(\rho{\,}\|{\,}\hat{\rho}_{d})\) represents the sparsity penalty. The average activation \(\hat{\rho}_{d}\) of hidden unit \(d\) is computed over the training set as

3.3. Fault Detection Based on Reconstruction Error

After the model is trained using normal drilling data, it can be used for online monitoring. For each new data segment, the reconstruction error is computed and used as a fault indicator. Since the reconstruction error distribution under normal conditions can be estimated, a threshold \(d\) corresponding to a confidence level \(\alpha\) is established using kernel density estimation. If the reconstruction error of incoming data exceeds this threshold, the system flags it as anomalous.

To set the anomaly detection threshold \(d\), a confidence level \(\alpha\) can be specified, and the probability density function of \(h_{X}\) can be used to determine the threshold.

In summary, the kernel density estimation is used to model the statistical distribution of the reconstruction error under normal operating conditions, and the anomaly detection threshold is determined as the \(\alpha\)-quantile of this distribution. This data-driven thresholding strategy avoids the use of fixed empirical thresholds and enables adaptive fault detection under varying drilling conditions.

The flowchart of fault detection for multivariate time series based on sparse autoencoder is illustrated in Fig. 3. In the offline training phase, a sparse autoencoder is trained using multivariate time series data of drilling parameters collected under normal drilling conditions. The reconstruction error between the original and reconstructed multivariate time series is calculated as a monitoring metric to assess the drilling process state. The alarm threshold is designed based on a given confidence level so that anomalies in the drilling process can be identified during the online detection phase.

For clarity, the overall implementation procedure of the proposed sparse autoencoder-based fault detection method is summarized as follows. First, multivariate drilling signals are segmented into overlapping samples using a sliding window of fixed length. These windowed data are used to construct the training dataset under normal operating conditions. Second, the sparse autoencoder is trained offline by minimizing the reconstruction error with sparsity and weight regularization, enabling the model to learn the normal behavior patterns of the drilling rig. Third, after training, the reconstruction errors of the normal training samples are calculated, and kernel density estimation is employed to model their statistical distribution. Based on a predefined confidence level, an adaptive anomaly detection threshold is then determined. Finally, during online monitoring, incoming drilling data are processed using the same sliding-window strategy and fed into the trained model. The reconstruction error of each window is computed and compared with the threshold to determine whether the current drilling state is normal or anomalous.

3.4. Fault Diagnosis Based on Detection Results

Building upon the anomaly detection results, a fault diagnosis mechanism is further developed to determine the potential fault type and the subsystem where it occurs. Following the principle of “detect first and then diagnose,” only data segments identified as abnormal by the reconstruction error are further analyzed, which reduces unnecessary computation and avoids misinterpretation under fluctuating drilling conditions. After an anomaly is confirmed, the system extracts key sensor parameters within the corresponding time window, including feed pressure, main pump pressure, rotary speed, return-oil pressure, and forward-rotation pressure. These parameters reflect the core hydraulic behaviors of the feed and rotary circuits, and their coupling relationships provide useful indications of possible fault propagation.

Based on the deviation patterns of these parameters from normal operation trends, and by comparing them with characteristic signatures of typical failure modes such as stagnant pressure increase, attenuated rotary speed, or abnormal return-oil accumulation, the system infers the most likely fault category and locates the affected subsystem. Meanwhile, the continuity of anomalies is considered to eliminate transient disturbances and avoid false alarms. Through this two-stage procedure combining anomaly detection and focused diagnostic analysis, the system achieves reliable localization of hydraulic abnormalities and provides interpretable diagnostic outputs that support on-site maintenance decisions.

4.1. Semi-Physical Simulation Experiment

Direct on-site validation under underground drilling conditions is the most straightforward way to carry out fault detection experiments and verify the proposed method. However, tests in tunnel environments involve high safety risks and many operational limitations. The harsh drilling conditions require minimizing interference during construction to avoid affecting borehole stability and equipment reliability. In addition, drilling still relies on manual remote operation, and the complex site environment often causes unexpected interruptions, making systematic experiments inefficient and difficult to reproduce.

To overcome these limitations, this section adopts a semi-physical experimental approach. By constructing a controllable laboratory system that replicates real drilling conditions, the experimental environment becomes safer, more stable, and easier to instrument. The semi-physical system used in this work, as shown in Fig. 4, is modified from the ZDY4500LFK, which is the primary tunnel drilling rig used in underground drilling. The test rig maintains the same mechanical structure and power output characteristics as the field equipment, allowing full simulation of rotary drilling, feed motion, and drill-pipe handling.

To process the data, a fixed-length sliding window approach was applied for segmentation. The sampling interval of the drilling monitoring system is \(\Delta t = 0.2\) s. Therefore, a sliding window of 25 samples corresponds to a temporal duration of 5 s. This window length is sufficient to capture the short-term dynamic coupling characteristics of the feed and rotary hydraulic systems during stable drilling, while maintaining adequate temporal resolution for online fault detection. Data representing non-productive periods, such as idle time or rod addition phases, are excluded from the dataset.

To ensure data quality, wavelet transform is applied to the raw sensor signals, effectively suppressing environmental noise and filtering out abnormal segments that did not align with continuous and stable drilling operations. For training and validation, 80% of the preprocessed data is randomly selected for training the sparse autoencoder model, while the remaining 20% is used for testing.

As shown in Figs. 5 and 6, the reconstruction errors of the training samples are mainly concentrated within a narrow low-value range, indicating that the sparse autoencoder accurately captures the characteristic patterns of normal drilling behavior. When applied to the test set, the model maintains good reconstruction performance in most intervals. However, as illustrated in Fig. 7, a small portion of the test samples exhibit reconstruction errors that exceed the threshold. These short intervals deviate from the learned normal patterns, suggesting the presence of mild operational disturbances or transient inconsistencies. The limited and localized nature of these exceedances confirms that the model is sensitive to subtle deviations while still maintaining stability for the majority of normal segments.

Considering that field data inevitably contain noise and transient fluctuations, an alarm optimization strategy is introduced: isolated threshold exceedances are not immediately classified as faults, whereas continuous exceedance over several consecutive windows is required to trigger an alarm. This strategy effectively suppresses false alarms and enhances the reliability of the detection results.

4.2. Actual Drilling Data-Based Experiment

Section 4.1 validates the proposed method under controlled semi-physical conditions, while Section 4.2 further examines its robustness and applicability using real drilling data. Together, these two experiments provide a coherent and complementary validation of the proposed fault detection framework.

To further evaluate the practical effectiveness of the proposed fault detection scheme, actual drilling data are collected from the Zhashui Experimental Base and used for the experiment. The underground drilling operation is carried out using a full-scale ZDY4500LFK tunnel drilling rig, as shown in Fig. 8. The equipment operates in a harsh outdoor environment with complex geological conditions, strong humidity, and significant mechanical vibration, which provides a realistic validation scenario for the proposed method.

Six key drilling parameters closely related to the hydraulic and mechanical behavior of the rig are selected for analysis. These parameters include penetration rate, rotational speed, main pump pressure, feed pressure, forward-rotation pressure, and return-oil pressure. Together, they characterize both the working status of the power head and the internal dynamics of the feed and rotary hydraulic circuits. The real on-site data enables rigorous verification of the model’s robustness under practical disturbances such as fluctuating load, sensor noise, and transient operational adjustments.

A comparison of the raw and denoised signals is presented in Fig. 9, confirming the effectiveness of the wavelet-based preprocessing in suppressing high-frequency noise and improving signal quality. As shown in Figs. 10 and 11, the sparse autoencoder successfully learns the underlying patterns of normal drilling behavior, with the reconstruction errors of the training samples concentrated within a narrow range and clearly separated from the estimated threshold. This demonstrates that the model has strong learning capability and good generalization performance across different datasets. Furthermore, Fig. 12 illustrates that all reconstruction errors in the field test set remain below the detection threshold, indicating that the drilling operation during data collection was stable and free of abnormal events. The absence of threshold exceedance also verifies that the proposed method maintains reliable performance and does not produce false alarms under normal condition.

From an engineering perspective, the proposed fault detection method has been validated using both semi-physical simulation experiments and actual drilling data collected from a full-scale tunnel drilling rig at the Zhashui Experimental Base. The results demonstrate that the method can operate stably under realistic drilling conditions characterized by strong vibration, load fluctuation, and environmental disturbances. In practical underground applications, the proposed approach can be integrated into existing monitoring systems by processing real-time sensor data through sliding-window analysis and providing anomaly alarms based on reconstruction error, without requiring additional fault labels or intrusive modifications to the drilling equipment.

Despite its effectiveness, the proposed method has several limitations. First, as a data-driven unsupervised approach, its performance depends on the representativeness of the normal-condition training data, and significant changes in geological conditions or operating modes may affect detection accuracy. Second, the model parameters and detection threshold may require recalibration when applied to different drilling rigs or configurations. In addition, the proposed method focuses on fault detection rather than precise fault type identification, which will be addressed in future work by incorporating additional diagnostic information and adaptive learning strategies.