Sliding Detection for Friction Welding Machine Tools Using AI

Raw Data - Training

The data is obtained from the friction welding machine’s control PLC through the MQTT protocol. The PLC Siemens 1500 includes a program that sends JSON messages containing the data values and timestamp to a Mosquitto broker installed on a Linux server. On the same server, a service implemented in Python reads these messages and generates CSV files that store all the collected data. To implement the real‑time detection system, another similar service is used, which periodically calls the detection function with a window of the most recent data, at a configurable time interval.

Two distinct Designs of Experiments (DOEs) were carried out, resulting in the production of 81 welds with a diameter of 20 mm and 20 welds with a diameter of 40 mm. For each diameter, half of the welds were manufactured under standard production conditions, while the remaining half were intentionally produced with an induced defect in order to generate a sliding phenomenon. In the welds with induced defects, two different behaviors were observed: in some cases, the welding process had to be interrupted due to excessive sliding, whereas in others the process was able to reach completion despite the sliding presence.

Using the developed platform, the following variables are monitored: Timestamp, Spindle Torque (Nm), Spindle Speed (rpm), Slide Force (kN), Slide Position (mm), Step, Spindle Clamp Pressure (bar), Spindle Clamp Position (mm), Slide Clamp Pressure (bar) and Slide Clamp Position (mm). As the output variable, a binary label is recorded to indicate whether the sliding phenomenon occurred during each welding cycle. This label constitutes the ground truth for the subsequent analysis and model development. It is important to note that, although the label confirms the presence or absence of sliding, the exact timestamp at which the sliding phenomenon occurs within the time series is not available.

Raw Data - Testing/Production

Preprocessing - Training Data

Once we have examined the ranges, structure, and patterns of each variable as part of the descriptive analysis, we proceed to analyze the linear correlations between them. In both correlation plots, we can confirm what we observed previously: our label is highly correlated with two specific variables whose values differ clearly between slicing and non‑slicing cases.

However, we do not want the model to learn from these lower absolute values; instead, our goal is for the model to learn from the underlying temporal patterns. Therefore, we will apply a preprocessing stage to these time series with the goal of removing the influence of absolute values and focusing exclusively on the dynamic patterns of the signals. This preprocessing includes two key steps: signal smoothing and signal normalization.

First, smoothing allows us to reduce high‑frequency noise present in the raw measurements, making it easier to identify the true underlying trends of the welding process. Techniques such as moving averages, Savitzky–Golay filters can help highlight the general shape of the signal without distorting it.

Second, we will apply normalization to make the signals comparable across different welds and to prevent the machine learning model from relying on anomalously low or high magnitudes. Min–max normalization or standardization (z‑score) have been used.

With this preprocessing, our goal is for the model to learn the relevant temporal characteristics—how the signal evolves over time, its changes, peaks, and transitions—rather than the differences in absolute magnitudes that we already know distinguish slicing from non-slicing cases. This will lead to a more robust and generalizable model.

Preprocessing - Testing/Production Data

Dataset - Training

The training dataset is composed of multivariate time‑series signals recorded during friction welding operations, including torque, pressure, and force measurements. To enable the use of traditional machine learning models on each individual weld, these raw signals are transformed into compact feature vectors by extracting descriptive statistical features—specifically the mean, standard deviation, minimum, and maximum of each time series. This process summarizes the global behavior of the weld while reducing dimensionality. After feature extraction, the dataset is normalized using a MinMax Scaler, preparing it for robust model training. Given the limited number of samples, particularly for the 40 mm diameter group, the dataset is evaluated using Leave‑One‑Out Cross‑Validation (LOOCV), which allows us to assess model performance by iteratively training on all welds except one and testing on the remaining instance.

Dataset - Testing/Production

Model Training

The model training followed two complementary approaches. First, statistical features were extracted from each weld and used to train a Random Forest classifier. To ensure robust evaluation with the limited dataset, Leave‑One‑Out Cross‑Validation (LOOCV) was applied. All features were scaled using a MinMax Scaler before training.

In the second approach, an LSTM Autoencoder was used to detect where slicing occurs within the weld. All weld signals were concatenated, and sliding windows were created to preserve temporal structure. The Autoencoder was trained exclusively on non‑slicing welds to learn normal behavior. During testing, reconstruction errors were used to identify abnormal windows, allowing precise localization of the slicing phenomenon within the welding process.

Trained Model

The trained model consists of two complementary components. First, a Random Forest classifier was developed using statistical features extracted from each weld. The features were normalized using a MinMax Scaler, and the model was trained through Leave‑One‑Out Cross‑Validation to ensure robustness with the limited dataset. A second model, an LSTM Autoencoder, was trained exclusively on non‑slicing welds to learn the normal temporal patterns of the welding process.

Model Testing

The Random Forest model was tested iteratively through LOOCV, evaluating its performance separately for 20 mm and 40 mm weld diameters. Confusion matrices, accuracy, and weighted F1 scores were used to quantify classification performance. For the LSTM Autoencoder, testing involved computing the reconstruction error on sliding windows generated from both good and defective welds. Windows with abnormally high errors were identified as indicators of slicing.

Tested Model

The final model combines global weld‑level classification with localized anomaly detection. The Random Forest provides a reliable binary decision for 20 mm welds, while the LSTM Autoencoder enables temporal localization of slicing events within the process. Together, they form a robust solution capable of both detecting the presence of sliding and identifying the specific regions where it occurs.