This research is centered on the development of a monitoring system for a C-frame punch press, with a dual objective framework. The first goal is to detect abnormal vibrations during machine operation, a critical indicator of potential mechanical issues. The second objective is the identification of anomalies in tools, specifically dies and punch heads, during the stamping process. The monitoring locations are explained in Fig. 1.
A strategically positioned accelerometer is affixed to the crankshaft assembly, facilitating long-term monitoring. This setup is pivotal for observing any misalignments or wear patterns in both the crankshaft and bushing. Conversely, in the context of shorter-term monitoring, this device plays a crucial role in detecting abnormal vibrations within the mold during stamping operations. Such observations are instrumental in identifying wear-related issues in punch heads.
The sensors and their installation locations are depicted in Fig. 2.
An additional monitoring strategy involves the integration of sensors within the mold's backplate, as illustrated in Fig. 3.
The backplate, being a critical force-bearing element directly interacting with the punch head, provides a rich source of data for analyzing punch head wear. In our setup, strain gauges are utilized in lieu of more expensive load cells on the backplate. These gauges are meticulously installed as per the structural design of the mold and calibrated to translate strain measurements into tonnage values. Variations in these tonnage values during the forming process are then analyzed in relation to the baseline strain signals. Furthermore, this study has successfully established a correlation between the wear of the punch head and the forces exerted on the backplate, as well as the resultant machine vibrations.
The critical timing for sensor activation is controlled by a digital encoder mounted on the machine's crankshaft. This strategic placement considerably streamlines post-signal processing tasks, including signal stitching, and effectively eliminates extraneous signals. In the preliminary phases of our research, on-site technicians employ a specially designed application to mark anomalous time points, encompassing events like punch head fractures, mold irregularities, and machine malfunctions. This approach significantly narrows down the scope for pinpointing signal discrepancies and expedites the process of signal labeling. This process continues until our algorithm can proficiently isolate and identify aberrant signals, thereby enhancing the overall efficiency of our diagnostic system.
The forthcoming sections of this paper will delve into a comprehensive description of the monitoring system's design and implementation. This encompasses a thorough overview of the hardware components, which include sensors, signal processors, signal extractors, and industrial-grade computing units. Additionally, the paper will detail the software aspects, focusing on the architecture and functionalities of the monitoring system platform. This holistic examination aims to provide a clear understanding of both the technical and operational aspects of the system.
2.1 Hardware Measurement Architecture
The hardware framework of the monitoring system employs sensors capable of generating analog outputs. These outputs are interfaced with a signal conditioner (SC), which not only supplies Integrated Electronics Piezoelectric (IEPE) power to the accelerometers but also plays a pivotal role in initial signal preprocessing. This preprocessing includes the crucial step of filtering out power-related disturbances emanating from the industrial setting. Following this initial conditioning, the analog signals are then converted into digital format via a Data Acquisition (DAQ) system. This conversion is essential for subsequent data analysis and interpretation within the monitoring framework.
For the accelerometers, a 16-bit Analog to Digital Converter (ADC) is employed for data conversion. In the case of strain gauges, a high-specification 24-bit ADC is utilized, enabling precise measurement of metal deformation down to a strain level of\({\left(10\right)}^{-6}.\)
The two depicted snapshot cards represent physical signal lines that connect to a Data Acquisition (DAQ) unit, which is then connected to the monitoring system on the computer. The monitoring system is designed to accommodate both wireless Wi-Fi and physical Ethernet network connections, depending on the actual on-site configuration.
The software architecture for the monitoring system is tailored to align with Company A's unique operational needs. The user interface features a homepage that concisely presents crucial metrics, such as machine utilization rates and the count of stamping operations completed. This design ensures the data is accessible and comprehensible to on-site personnel. Subsequent interface sections are dedicated to the real-time display of raw signal inputs, enabling our team to continuously monitor signal line integrity and hardware performance. This setup is instrumental in promptly identifying and addressing any anomalies or irregularities in the system.
2.2 Methods
In concurrent studies, our research team has made significant strides in monitoring the service life of individual punch heads during the punching process. As detailed in our previous publication [5], we successfully demonstrated that both time and frequency analyses of extracted feature signals could establish a quantifiable relationship between the wear of punch heads and the resulting burr height. Furthermore, in study [6], we implemented logistic regression analysis and utilized the statistical overlap factor to define logical decision-making thresholds within the monitoring system. These experimental and analytical endeavors were conducted on a 50-ton hydraulic punching machine, situated within our university's smart factory. This real-world testing environment has been instrumental in validating our methodologies and findings.
The software component of our monitoring system is meticulously designed to cater to the distinct needs of Company A. The user interface is structured with a homepage that prominently displays essential statistics, including machine utilization rates and the total count of stamping operations executed. This data is presented in a user-friendly format, ensuring ease of understanding for on-site staff. Subsequent sections of the software interface are dedicated to real-time acquisition of raw signal data. This feature enables our technical team to continuously observe the status of signal lines and hardware components, facilitating prompt detection and response to any anomalies or irregularities that may arise.
2.3 Process
The segmentation of a complete stamping vibration signal into three distinct phases is depicted in the diagram, as identified in our research on stamping processes and corroborated by the findings in reference [4]. These phases are delineated based on the relative timing within the stroke cycle of the stamping machine:
-
Upper Die Pressing Stage: This phase encompasses the interval during which the upper die applies pressure to the material.
-
Punching Stage: This phase represents the moment when the punch head actively pierces through the material.
-
Upper Die Lifting Stage: This final phase occurs when the upper die is retracted or lifted away from the material.
Each of these stages is critical for understanding the dynamics of the stamping process and is instrumental in the analysis of vibration signal patterns associated with different operational phases.
Through rigorous statistical analysis conducted in the referenced study, it was determined that the vibration signal captured during the punching stage yields more accurate representations of the stamping process's inherent characteristics. Additionally, a positive correlation was established between the degree of punch wear and the vibrational patterns observed during this punching stage. This finding underscores the significance of the punching stage vibration as a reliable indicator for assessing punch wear in the stamping process. The diagram illustrating the research process for time-domain signal extraction is presented as Fig. 4.
This analysis involves measuring z-axis vibrations using an accelerometer, enabling the categorization of the signals into four distinct signal amplitude phases during a complete punching process. These phases are characterized as follows:
-
Bed Movement: The upper bed moves downward from the top.
-
Die Pressing Stage: The vibration signal generated when the pressing plate first contacts the sheet metal and the lower die.
-
Punching Stage: The moment when the punch extends downward to pierce the sheet metal.
-
Lifting Stage: After the punching is complete, the upper bed moves upward, and the pressing plate disengages from the sheet metal.
The explanation diagram for time-domain signal extraction is depicted as Fig. 5.
In accordance with the signal extraction methodology outlined, the pivotal third stage, namely the punching stage, is isolated for detailed analysis, as depicted in Fig. 6. This figure methodically illustrates the punching stage from various analytical perspectives. From top to bottom, it showcases the time-domain vibration signal, the frequency-domain representation of this vibration, followed by the punch strain, and finally the external strain measurements. This comprehensive representation effectively encapsulates the entire time-domain extraction process. Subsequent sections of this paper will elaborate on the total number of experiments conducted and the methodology employed to transform each punching event into quantifiable energy values. This transformation is crucial for facilitating a comparative analysis of the variances across different punching events.
Based on the signal extraction process, the third stage, the punching stage, is isolated, as shown in Fig. 6. From top to bottom, the figure displays the punching stage in the time domain vibration, frequency domain vibration, punch strain, and external strain, completing a time-domain extraction process. The following will explain the overall number of experiments and how each punching event is converted into energy values to compare their differences.
In this paper, the energy statistics method used is the Root Mean Square (RMS), as shown in Eq. 1. It converts the extracted vibration and strain values into representative numerical quantities, effectively illustrating how the values change as the number of stamping cycles increases over time. In this experiment, as presented in Table 1, there were a total of 15,410 data points, with 3,044 data points in the stamping range of 1-5k cycles and 12,366 data points in the range of 25k-46k cycles.
Root Mean Square (RMS): \({x}_{rms}\)=\(\sqrt{\frac{\sum _{\text{n}=1}^{\text{N}}{\text{x}\left(\text{n}\right)}^{2}}{\text{n}}}\)
(1)
Table 1
Stamping Experiment Data.
| Item Interval | Stamping Interval (Cycles) | Number of Data Points (Entries) |
| 1 | 1-5k | 3,044 |
| 2 | 25k-46k | 12,366 |
The main components include the time-domain and frequency-domain representations of vibration signals. Additionally, energy trends calculated via the Root Mean Square (RMS) method are presented in both panels. Data from two distinct stamping intervals are juxtaposed for comparative analysis. Examination of these panels reveals that in the 25k-46k cycles interval, approximately 50% of the readings approach the peak values of 0.4 in the time-domain and 0.03 in the frequency-domain. In contrast, the 1-5k cycles interval exhibits an absence of readings near these maximum values, indicating minimal variance between the original time-domain and frequency-domain values. To adequately capture the significance of the observed increase in stamping interval values, further extraction of feature signals utilizing the Ensemble Empirical Mode Decomposition (EEMD) method is essential. The time-domain and frequency-domain plots are presented in Fig. 7.
In Fig. 8, the upper and lower plots represent the extracted strain signals for punching forming and appearance forming, respectively. The strain sensors located behind the punch provide a direct indication of the force conditions and wear level during the punching process. In the upper plot, the strain signal for punching forming exhibits a linear growth from interval 1 to interval 2, increasing from\({2.3*10}^{-3}\)to \({2.125*10}^{-3}\) in strain magnitude. On the other hand, in the lower plot, the strain magnitude for appearance forming increases from \({1.1*10}^{-3}\) to\({1.35*10}^{-3}\). It is observed that the strain linearity in appearance forming is less pronounced compared to punching forming, primarily due to the differences in the shape and dimensions of the punch heads. As shown in Fig. 9, the left image represents a circular punch head, while the right image represents a concave-shaped appearance punch head, highlighting the noticeable differences.
-
Ensemble Empirical Mode Decomposition (EEMD): [1][2][3] The Hilbert-Huang Transform (HHT), proposed by Dr. Norden E. Huang and his colleagues in 1998, comprises two main steps: Empirical Mode Decomposition (EMD) and the Hilbert Transform (HT). In this study, we employ Ensemble Empirical Mode Decomposition (EEMD), also introduced by Dr. Norden E. Huang in 2008, with the aim of improving the Intrinsic Mode Functions (IMFs) within the EMD. In the standard EMD, when selecting local maxima and local minima, it is required that these extrema cross zero points. If the next extremum does not cross a zero point, it is ignored, and the following zero-crossing extremum is considered. This can lead to the extraction of amplitude components larger than those present in the original signal, causing signal distortion. To mitigate such distortion, EEMD introduces minimal amplitude white noise conforming to a Gaussian process into the time-domain signal before the decomposition process begins. This establishes an ensemble of multiple samples, and the true IMF is defined as the average of the ensemble samples obtained through EMD filtering. The decoupling process of EEMD is outlined as follows.
In EMD, it is essential that both local maxima and local minima cross zero points. If an Fig. 10 shows the vibration signal from the 24th stamping cycle in the Company A on-site stamping process, subjected to the EEMD decomposition process. The X-axis represents data points, and the Y-axis represents vibration values. The signal is decomposed into eight IMF components, with the frequency decreasing from IMF1 (high-frequency) to IMF8 (low-frequency). These eight IMFs are considered as significant features, and after statistical analysis, they are used as input parameters to construct the punch head failure monitoring model.
In Fig. 11, we applied Eq. 1 to conduct statistical calculations for the first eight modes of Ensemble Empirical Mode Decomposition (EEMD) within two distinct stamping intervals, delineated by black dashed lines. These intervals correspond to the stamping ranges specified in Table 1, specifically, 1-5k strokes and 25k-46k strokes. The X-axis in the figure represents a total of 15,410 data points from stamping records, while the Y-axis represents vibration energy values obtained through Root Mean Square (RMS) analysis for each stamping event. From top to bottom, the plot illustrates the energy values for EEMD modes 1 through 8.
Mode 1, representing the original time-domain signal, shows minimal differences in values between the two intervals, making it unsuitable for use as an input parameter. Modes 2 to 4, on the other hand, exhibit significant variations in values, making them highly suitable as the primary parameters for training. Modes 5 to 8 show moderate differences in values and can also be used as supplementary training data.
-
Neural Network Model (Bayesian regularization, BR): The Bayesian regularization (BR) algorithm serves as the foundation for the model used in this research, aiming to expedite the training speed of neural networks. The Bayesian regularization method is specifically designed to utilize a mean square error loss function. As mentioned in reference [7], trainbr can train neural networks as long as network weights, inputs, and transfer functions are differentiable.
Bayesian regularization minimizes the linear combination between the square error and the weights, modifying this combination to ensure the network's generalizability upon training completion, as discussed by Foresee and Hagan in reference [8]. The algorithm incorporates Bayesian regularization within the Levenberg-Marquardt optimization method. Backpropagation is employed to compute the performance gradient with respect to weights and bias variables, denoted as jX. Each variable is adjusted according to Levenberg-Marquardt as follows.
$$jj = jX * jX je = jX * E dX = -(jj+I*mu) \backslash je$$
(2)
Where E represents all error terms, I presents the identity matrix, and mu is an adaptive value increased by mu_inc until the above change leads to a decrease in performance. Subsequently, the network is updated, and mu is decreased by mu_dec. Training stops under the following conditions:
-
Reaching the maximum number of steps (epochs).
-
Exceeding the maximum time limit.
-
Performance dropping to a minimum threshold.
-
Performance gradient falling below 'min_grad.'
-
'mu' exceeding
-
'mu_max.'
Figure 12 illustrates the architecture of the punch head failure monitoring model constructed by our team. The input training data consists of punching strain values, appearance strain values, and IMF (Intrinsic Mode Function) 1 ~ 7 vibration magnitudes, totaling 9 sets. The neural network has 3 hidden layers, with 28, 21, and 14 neurons in the 1st, 2nd, and 3rd layers, respectively. The output results are represented by numbers 0, 1, 2, and 3, indicating different stamping intervals.
2.4 Model Evaluation
To assess the fault diagnosis results, we employ a confusion matrix to illustrate four possible scenarios [9]. True Positive (TP), False Negative (FN), False Positive (FP), and True Negative (TN). The details are presented in Table 2. Additionally, accuracy is used as the experimental metric in the validation section. The equations for calculating is shown in Equations (3).
Table 2
Confusion matrix for fault diagnosis
| Actual Condition |
| Prediction Condition | | Failure | Normal |
| Failure | TP (True Positive) | FP (False Positive) |
| Normal | FN (False Negative) | TN (True Negative) |
$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$
(3)