Development of a Wireless Predictive Maintenance Sensor Network for Lost Foam Casting

Transcription

1 Development of a Wireless Predictive Maintenance Sensor Network for Lost Foam Casting Copyright 2007 American Foundry Society M.J. Whelan and K.D. Janoyan Clarkson University, Potsdam New York ABSTRACT The mold production stage of the Lost Foam Casting (LFC) process utilizes flask vibration during sand raining to encourage sand flow and compaction. Due to the considerable weight of the flask and the sand in the mold, a significant amount of energy is transmitted throughout the compaction table during excitation. In addition to driving the flask, this energy excites many modes of vibration from other mechanical components within the compaction system. The contributions from the component vibration modes can be isolated and quantified using frequency-based analysis of sensor data. Using low-cost MEMS accelerometers, the vibration characteristics of critical locations on the compaction table can be monitored to detect changes in the system, such as loose bolts, mechanical unbalance, or gear and bearing damage. By actively monitoring the overall system health, the quality of the sand mold can be maintained and machine maintenance can be predicted prior to failure to reduce downtime. This study presents the development of a wireless accelerometer sensor network for industrial predictive maintenance including results from deployments on a Lost Foam Casting compaction line. The wireless nature of the accelerometer nodes permits permanent placement on the casting flask, which travels throughout the foundry during the casting cycle. This eliminates the time associated with routine placement and permits for a unique vibration signature to be developed for individual flasks. INTRODUCTION During the mold production stage of the LFC process, the compaction table executes a programmed sand raining and flask vibration schedule specific to the casting pattern (Fig. 1). The schedule, generally adopted through trial-and-error iterations, targets rapid fill times to maintain process efficiency, with consideration of the need for increased duration of fluidization around complex areas of the pattern or regions with difficult to fill cavities. In addition to filling cavities, flask acceleration must either be of significant magnitude or prolonged duration to achieve the densification necessary to provide sufficient confinement to the advancing molten metal during casting. Generally, increased flask acceleration is preferred to minimize the process length, though excessive accelerations can produce distortion of weak areas of the foam pattern, grain penetration into the foam, or anisotropic stiffness in the confining sand insufficient to confine the flow of metal (Kuraoka, 1994). Aside from sensors used to control compaction table operational tasks, the mold production process is generally equipped with limited, if any, sensors for quality control and process feedback. Operator feedback is typically limited to visual observations, which are heavily obstructed by the flask and raining equipment, in addition to drive frequency indication and perhaps single-location root-mean square acceleration on the compaction table. When employing an established pattern schedule on a compaction table in a production foundry, it is imperative that the compaction table produce repeatable sand molds within reasonable tolerance for defect-free castings. To ensure the longterm stability of the process quality, monitoring of the compaction table must be performed regularly to guarantee that the system dynamics remain relatively unchanged over the long-term cycle of pattern production. In other words, due to machine wear or damage, over time the process may produce unacceptable casting molds despite employing a programmed schedule that is optimized for a healthy compaction table. Machine condition changes can readily be identified through acceleration vibration measurements and comparison over time with developed vibration signatures (Mobley, 1999). Furthermore, localization of the machine elements exhibiting wear or damage can often be identified through distributed acceleration measurements coupled with spectral analysis of the vibration data. This study addresses these process issues through the development of a wireless accelerometer network. Wireless technology was pursued to enable permanent placement on features of the casting system that move throughout the foundry, such as the casting flask. Due to bandwidth and packet success rate considerations, the sensor network approaches the monitoring tasks through using two distinct architectures operating concurrently during sampling. A limited number of accelerometers placed directly on the flask sample and transmit the vibration data in real-time to provide in-process feedback. Flask accelerations can be compared to predetermined limits or profiles to provide the controller with measurements to adjust process control parameters. Additionally, wireless accelerometers are placed on the flask and/or on compaction table elements to record vibration time histories, which are recorded to onboard flash memory for post-cycle data transfer. The measurements from this distributed sensor network can be used to identify and locate changes in the health of compaction table components, thereby ensuring the quality of mold production and minimizing down-time for inspection and repair.

2 Figure 1. Lost Foam Casting Process Schematic DEVELOPMENT OF WIRELESS ACCELEROMETER NETWORK To facilitate process monitoring and routine machine condition assessment, a wireless sensor network was developed using commercial wireless transceivers, custom signal conditioning hardware, and low-cost MEMS accelerometers (Fig. 2). The wireless nature of the sensor nodes permits permanent placement on the large number of flasks that circulate through the foundry, thereby eliminating routine setup time and costs. The sensor nodes actively sample the accelerometer channels during the mold production cycle and utilize ultra-low power shutdown of hardware components during remainder of the process to preserve limited battery resources. Two software architectures are available for sampling to maintain 100% data recovery despite the limited bandwidth of the transceivers: 1) Real-time scheduled transmission of sensor data from a limited number of accelerometers for in-process feedback, and 2) Data logging to onboard flash memory for subsequent data transfer to facilitate the deployment of a large number of sensors in the network. Both software applications communicate with a central base transceiver that is used for in-network programming of sampling parameters, task signaling, and data collection. ACCELEROMETER 5-th ORDER ANTI- ALIASING FILTERS WIRELESS TRANSCEIVER Figure 2. Prototype Wireless Accelerometer Node MEMS accelerometers were selected to minimize the system costs associated with a large sensor array necessitated to instrument the flasks and machinery in a production foundry. The accelerometers selected are dual-axis capacitance-based sensors with a bandwidth of 5kHz, which is more than sufficient to capture the table frequency (40-60Hz) and significant harmonics. Full-scale ranges of +/- 10g and +/- 2g were used to monitor vibrations in the primary direction of vibration and off-axis directions, respectively. Analog low-pass filtering hardware was developed using 5-th order switched capacitor filters to prevent signal aliasing and increase signal resolution by reducing the noise bandwidth. The low pass filter design permits in-network reprogramming of the filter cutoff frequency to tailor the sensor bandwidth to the needs of the system. The Tmote Sky wireless sensor platform, developed and marketed by the Moteiv Corporation, was utilized to enable ultralow power, remote data acquisition and wireless transmission of the sensor signal. The onboard microcontroller has an internal, eight-channel 12-bit analog-to-digital converter with hardware timed sampling as well as access to programmable timer modules and digital I/O signals for interfacing with external sensors and signal conditioning circuits. The chip transceiver incorporated is an IEEE (ZigBee) compliant device capable of up to 250 kbps data throughput with

3 digitally programmable output power and channel frequency. Additionally, the transceiver provides Direct Sequence Spread Spectrum (DSSS) digital modulation of the radio signal, thereby enabling enhanced performance in the increasingly noisy 2.4GHz frequency band. The wireless accelerometer nodes were programmed with custom software applications and interfaced to the host computer for network communication and data collection through a developed LabVIEW application. This custom software interface was developed to control the sensor network with a base transceiver interfaced through a USB connection. Through two-way wireless messaging, the base mote can signal either individual motes or the entire network to configure sampling parameters, initiate sampling, terminate sampling, recover data wirelessly, and erase the flash memory. During post-sampling data recovery, radio acknowledgement messages are sent in response to received packets passing error frame checking in order to establish a two-way communication control to ensure complete recovery of data. Alternatively, the motes can be connected to a USB interface for faster reading of the external flash storage. INLINE SENSOR DEPLOYMENT The sensor nodes were programmed with applications developed to sample data from each axis of the accelerometers using the internal 12-bit ADC with hardware timing at a sampling rate of 1kHz. The anti-aliasing filters were set to 250Hz to permit accurate representation of the primary table frequency as well as the first four harmonics while sufficiently suppressing frequency content above the Nyquist rate. Generally, only two-axes of acceleration were transmitted in real-time for indication of acceleration characteristics during the cycle. By maintaining such a small load on the limited transceiver bandwidth, full packet success was assured. With proper arbitration and acknowledgements of packet reception, it is however feasible to wirelessly transmit data at the same sampling rate from several transceivers in a reasonably sized network. For the data logging nodes, the samples were written to a 1MB external flash memory chip during the mold production cycle and then recovered from the wireless transceiver afterwards. The 1MB capacity of the external flash permits storage of up to 512,000 samples, which correlates to approximately two minutes and 50 seconds when sampling across three channels at a rate of 1kHz. Fortunately, this length of time coincides with the approximate length of the typical mold production cycle, however, within this research, flash cards, such as those used in digital cameras, were also interfaced to permit longer length sampling as well as faster acquisition rates. The sensor array was deployed on a horizontal compaction table equipped with a rectangular flask having an approximately 0.75m square cross-section and 1m depth (Whelan, 2005). The flask was clamped at the rear at three locations while the leading bottom edge of the flask was supported by a pneumatic support with rotational freedom of motion (Fig. 3). Dynamic, sinusoidal flask excitation was provided by a belt-driven off-balance flywheel rotated at a primary table frequency of approximately 47Hz. Acceleration measurements across the flask indicate the presence of a significant degree of pitching motion in addition to the primary translation in the direction of forced sinusoidal motion. RMS acceleration in the primary direction varies with height, though is generally around 1.5g, while in the vertical direction it is typically around 0.5 g. As expected, accelerations on the flask decrease as a function of the depth of sand in the flask, though the acceleration measured at the compaction table on one of the clamps was found to increase with the addition of sand. Motor Clamps Flask Rocker Support Primary Vibration Direction Figure 3. Horizontal Compaction Table for Mold Production in LFC FREQUENCY SPECTRUM One of the most basic approaches to predictive maintenance using vibration measurements is the comparison of the frequency content of the signal to a developed signature frequency spectrum or baseline. In reciprocating and rotating machines, harmonics and ½ harmonics often arise from specific phenomenon, such as imbalance, mechanical looseness, misalignment, eccentricity, gear and component wear, and other deteriorations in the overall system health (Mobley, 1999). A waterfall display, or stacked plot, of the frequency spectrum as measured across the cycle over time can serve as a very effective means of identifying changes in machine condition and signal maintenance. The Fourier Transform was used with a Hanning windowing function to compute the frequency spectrum across the duration of the cycle in both the primary and vertical direction of flask motion (Fig. 4). The spectrums for each direction are similarly characterized by a strong

4 fundamental frequency with significant harmonics as well as smaller magnitude ½ harmonics. However, the relative magnitude of the harmonics to the primary, or fundamental, frequency is larger in the vertical direction than in the forced direction and ½ harmonics are less evident. Consequently, the relative contribution of the harmonics is larger in the vertical direction and, likewise, changes in the spectrum may be slightly more apparent in this direction. Figure 4. Frequency spectrum produced for the duration of the casting cycle in a) the primary direction of flask excitation, and b) the vertical direction HARMONIC DISTORTION CORRELATION TO PROCESS PARAMETERS Since the flask is subject to sinusoidal excitation, the acceleration waveforms are expected to exhibit a primarily sine wave profile. During post analysis, the accelerometer data was streamed in smaller time windows to examine the characteristics of the vibration time series. In the primary direction of excitation, the sinusoid is clearly apparent and to a large degree maintains its integrity over the course of the filling cycle. Although there is a small degree of distortion that develops during the cycle, the magnitude of harmonic distortion is very small and the acceleration remains dominated by the table drive frequency. In contrast, the vertical acceleration displays a response that alters in consequence to the fill and compaction (Fig. 5). During the cycle, the time series is initially a relatively pure sine wave, then gradually develops into a nearly sawtooth function, and by the end of the cycle resembles a stepped sawtooth. Joint time-frequency analysis reveals that the distortion is the product of significant secondary frequencies that arise in the vertical direction with the addition of sand to the casting flask.

5 Figure 5. Development of Harmonic Distortion with Addition of Sand Wavelet analysis is a multiresolution analysis method that provides a means of deconstructing a signal s frequency content while maintaining the signal representation in the time domain. Consequently, this deconstruction permits the visualization of both the frequency content and time series simultaneously, a shortcoming of the traditional Fourier transform. Essentially, the method is an efficient means of digital filtering the time sequence that successively divides the bandwidth of the signal into a collection of filter banks. Wavelet analysis was performed on the acceleration time history in order to separate the contributions of the harmonic components of the signal. Wavelet decomposition of the signal using the discrete Meyer wavelet is provided in Figure 6. In each subplot, the largest magnitude waveform is constructed from the lower 62.5Hz of the bandwidth that contains the table primary frequency, the second strongest signal is the filtered signal containing the bandwidth from 62.5Hz to 125Hz that contains the first harmonic of the table frequency, and the weakest signal is constructed from the bandwidth from 125Hz to 250Hz, which encompasses the next three harmonics. Further decomposition can be performed using wavelet packet analysis to separate the contributions from the higher order harmonics, though the degree of filtering utilized sufficiently isolated the peak frequency and the first harmonic, which are the focus of the following discussion. The wavelet decomposition clearly indicates that the first harmonics is largely responsible for the distortion of the sinusoidal motion in the vertical direction that is noted with the addition of sand to the casting flask. Bast et al. (2000) presented a mathematical model of sand flow in casting flasks on horizontal compaction tables in addition to experimental measurements of the vibrations as measured in the sand. Both the model and the accelerometer data indicated significant vertical acceleration in the sand at a fundamental frequency of twice the horizontal table frequency, i.e. the first harmonic of the table frequency. Furthermore, the vertical sand acceleration was found to be approximately 90 out-of-phase with table acceleration. In regard to the frequency and phase, the characteristics of the waveform measured in the vertical direction in this study are consistent with those predicted and measured within the sand by Bast, et al. (2000). This tends to indicate that the sand accelerations can be measured non-intrusively by monitoring accelerations on the flask, with particular interest in the first harmonic frequency, rather than within the sand. However, there are many other mechanical factors that contribute to the first harmonic of the drive frequency of machine systems, so this correlation can only be applied loosely.

6 Figure 6. Harmonic distortion revealed through wavelet decomposition Since the first harmonic provides such as significant contribution to the content of the acceleration spectrum and can be correlated to sand flow models and measurements, the use of its magnitude for in-process indication may be beneficial. The magnitudes of the harmonic frequencies across the duration of the cycle are presented in Figure 7. Generally, only the primary table frequency and the first harmonic display profiles with intuitive trends. As expected, the primary table frequency projects a decrease in acceleration as sand filling occurs, which is a result of the increased mass-loading of the system. The compaction cycle analyzed in this figure featured filling in four stages, which are clearly divided at the primary frequency by the presence of constant acceleration when the filling was paused. Alternatively, the first harmonic, while significant in magnitude, does not project a linear dependence on the mass of sand in the flask. It is the author s opinion that, under the assumption that the first harmonic is strongly correlated to the sand flow, the profile may be consequent of the depth of the fluidized sand layer. The fluidized layer is the region in which dynamic sand motion occurs, which is known to be at a frequency of twice the table speed. As the layer increases with the addition of sand, the vertical acceleration transferred to the flask increases in magnitude, which is noted in the profile. However, as sufficient overburden stress develops with the filling of the flask and as the sand compacts assisted by the vibration, the deeper layers of sand cease to flow dynamically and travel rather as a block in phase with the flask. This may account for the decrease in magnitude of the first harmonic as the cycle progresses in the second half of the compaction stage. However, additional correlations with process parameters and machine condition indicators are necessary before the profile of the first harmonic amplitude can be attributed to definitive in-process phenomenon.

7 Figure 7. Contribution of harmonics to flask vertical acceleration CASE STUDY: RESONANCE DURING FLASK SHAKEDOWN Monitoring the harmonic frequencies, or orders, during ramp-up or shake-down of a cyclical machine is particularly important because machine faults or changes are often most evident during changes in fundamental speed (Han, et al., 1999). For the mold production stage of LFC, the table frequency inherently sweeps across a range of low frequencies as flask vibration is initiated and later as the flask vibration is ramped back to stationary. While the ramp-up of flask vibration may provide a means of investigating the condition of machine components through excitation of low frequency modes, the shakedown of the flask was the primary focus of this case study. At this stage in the cycle, the sand mold has been completely formed. Ideally, the flask should smoothly decelerate until dead-stop in order to avoid harsh dynamics that might cause re-fluidization of the sand mold and possible disruption of local confining stresses on the foam pattern. However, it was visually and audibly evident on the compaction table used for the test sequences that the deceleration was not smooth and possesses a potential quality control issue in the mold production stage (Fig. 8). This section of the paper presents a case study of the vibration characteristics during the shakedown of a filled flask for possible isolation of the root cause of dynamic resonance. Figure 8. Evidence of excited resonance during sweep of table frequency ORDER ANALYSIS Order analysis is a form of frequency analysis that transforms the signal from the time domain to a scale based on multiples of the driven system frequency rather than absolute frequency. It is particularly advantageous for machine condition monitoring applications since it enables the orders to be tracked even while the primary frequency changes, which is generally a shortcoming of joint time-frequency analysis. During frequency changes, such as during the ramp-up or shake-

8 down of rotating or reciprocating machinery, short-term Fourier Transform tends to spread the fundamental frequency and its harmonics, thereby providing poor resolution of the signal composition. Traditionally, order tracking is performed either: 1) using acquisition hardware that employs a tachometer or encoder signal to trigger sampling of acceleration signals at constant angular intervals, or 2) through software resampling of data sampled at constant time increments using a simultaneous encoder measurement of angular position to interpolate the signal at constant angular intervals. The test sequences performed did not utilize any transducer for measuring the rotational position of the compaction table motor or flywheel or the direct lateral displacement of the flask at the clamps, so these traditional methods could not be implemented. As an alternative, a software algorithm was implemented that uses a high-resolution method of joint-time frequency analysis, the Wigner-Ville Distribution, of small time windows of data to estimate the instantaneous frequency of the signal, which was then numerically integrated to obtain angular position (Han, 1999). INVESTIGATION OF SHAKEDOWN STAGE OF CYCLE Vibrations were measured at several locations across the flask as well as on the lower clamp of the compaction table. Since the table frequency tends to dominate the signal in the primary direction of flask motion, vertical accelerations were selected as the primary focus of the following order analysis. Using the algorithm described above, acceleration data from the postcycle system ramp-down was resampled at constant angular intervals of 20. Since order analysis is not as widely recognized and utilized as traditional Fourier transforms, the transforms of the original data and the resampled uniform angular signal for a shakedown cycle are presented in Figure 9. The comparison clearly indicates the ineffectiveness of equal time spaced sampling across non-stationary signals and illustrates the usefulness of order analysis in tracking harmonic contributions during a sweep in fundamental system speed. Figure 9. Comparison of spectrums obtained during shakedown for a) constant-time sampling; and, b) constant-angle sampling The magnitude of the orders, or harmonics, throughout the duration of the ramp-down of the flask for both a location on the flask as well as on the lower table clamp is presented in Figure 10. In particular, the first and third orders feature a low frequency resonance condition that results in a reversal in the deceleration of the flask. The phase response of the cycle during the shakedown reveals a phase reversal that verifies the presence of the resonance. Identification and isolation of the source of this resonance is desirable, as the amplitude and phase reversal of flask acceleration could produce potentially counter-productive sand fluidization within the final mold resulting in loosening of compacted areas sand or counter-flow from cavities. The order analysis of the shakedown is advantageous as different modes of vibration specific to machine components or wear conditions reflect differently in the frequency, or order, content of the signal (Mobley, 1999). The content of the order spectrum indicates several possible sources of the resonance, which necessitates further sensor

9 deployment to identify the source of the resonance with certainty. Natural frequencies in the support structure of the flask, mechanical out-of-balance, or resonance of the rocker flask support are all plausible sources. An additional potential cause of resonance could be the belt that drives the compaction table from the motor. If the belt resonance is below the running speed of the compaction table, one or more of its mode shapes will be excited during the shakedown, though it is unlikely that the increase in magnitude of flask acceleration would be as significant as noticed. Due to the relatively high magnitude and low frequency of the main resonance during the ramp-up and shakedown of the flask, the logical source is resonance of the pneumatic rocker support at the toe of the flask. However, a comprehensive vibration measurement and analysis study of the compaction table is necessary to identify the resonance source with certainty, which could potentially lead to a higher rate of casting success and, possibly, reduced mold production time. Figure 10. Order analysis Tracking harmonics during shakedown: a) on flask, b) on clamp CONCLUSION Presented is the development of a wireless sensor network for industrial process monitoring using accelerometers to measure vibration characteristics. The deployment of these sensors on a compaction table during the mold production stage of the Lost Foam Casting process has provided insight into the process dynamics. Advanced signal processing methods have been applied to the accelerometer measurements to extract relationships among signal frequency content and plausible physical sources. Consequently, the technology offers promise for future development of in process parametric monitoring for controller indication and control feedback. Additionally, long-term monitoring of the vibration spectrum offers a means of implementing a predictive maintenance program for early machine damage identification and stricter quality control regulation. ACKNOWLEDGEMENTS This project is partially funded by the New York State Energy Research and Development Authority (NYSERDA) and General Motors Powertrain (GMPT). Additional funding is provided by the New York State Office of Science, Technology, and Academic Research (NYSTAR) Center for Advanced Material Processing (CAMP). The authors would like to individually recognize Tom Gustafson of General Motors Powertrain and Scott Biederman of Metal Casting Technology, Inc for their continued support. Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not reflect the views of the funding agencies. REFERENCES Bast, J., Nikolov, K. and Clegg, A.J. Analysis of vibrations during sand casting in the evaporative pattern casting process. International Journal of Cast Metals Research. Vol 13. p (2000) Han, Y-S. and Lee, C-W. Directional Wigner Distribution for Order Analysis in Rotating/Reciprocating Machines. Mechanical Systems and Signal Processing. 13(5) p (1999) Kuraoka, Senro Anisotropic stiffness and circulation flow of sand; application for the expendable pattern casting. PhD Dissertation, University of Wisconsin-Madison, Department of Civil & Environmental Engineering. 230 p. (1994). Mobley, R. Keith Vibration Fundamentals Butterworth-Heinemann. Woburn, MA. 295 p. (1999) Whelan, M.J. Development of Advanced Measurement and Monitoring Tools for the Sand Fill and Compaction Stage of the Lost Foam Casting Process. Masters of Science Thesis, Clarkson University, Potsdam, NY. 90 p. (2005).