MONITORING BOLT TIGHTNESS USING PERCUSSION AND MACHINE LEARNING

Abstract

The systems and methods described herein are for monitoring the tightness of bolts. The systems and methods may be used with a mechanism to apply a percussive tap and with a recording or monitoring device for detecting and recording the acoustic signals that are generated by the percussive tap. The acoustic signals generated by percussive taps applied to bolts in various looseness states are analyzed and a machine learning model is developed that allows for determining bolt looseness based on the acoustic signals.

Claims

1. A method for monitoring the tightness of a bolted connection, comprising: using an impact device to apply at least one percussive tap to a portion of a plurality of known bolted connections, wherein each known bolted connection has a known tightness, to produce a model acoustic signal for each known bolted connection; recording the model acoustic signals using a recording device to produce model acoustic data; extracting features from the model acoustic data wherein the features are correlated with the known tightness of each known bolted connection; using the features and the known tightness of each known bolted connection to generate a decision tree model for predicting tightness of a bolted connection using an acoustic signal; using the impact device to apply at least one percussive tap to a portion of an unknown bolted connection having an unknown tightness to produce an unknown acoustic signal for the unknown bolted connection; recording the unknown acoustic signal for the unknown bolted connection using a recording device to produce unknown acoustic data; extracting the features from the unknown acoustic data; and using the decision tree model to predict the tightness of the unknown bolted connection using the features extracted from the unknown acoustic data.

2. The method of claim 1, further comprising: using the impact device to apply at least one percussive tap to a portion of a plurality of test bolted connections, wherein each test bolted connection has an known tightness, to produce a test acoustic signal for each test bolted connection; recording the test acoustic signals using a recording device to produce test acoustic data; extracting features from the test acoustic data; using the decision tree model to predict the tightness of the test bolted connections using the features extracted from the test acoustic data; and comparing the tightness of the test bolted connections predicted by the decision tree model to the known tightness of the test bolted connections in order to validate the decision tree model.

3. The method of claim 1, wherein the impact device is a hammer.

4. The method of claim 1, wherein the features are power spectrum density (PSD) at selected frequency ranges.

5. The method of claim 4, wherein the selected frequency ranges are 200-300 Hz, 300-400 Hz, 400-500 Hz, and 500-600 Hz.

6. A system for monitoring the tightness of a bolted connection, comprising: an impact device, wherein the impact device is positionable to apply at least one percussive tap to a portion of an unknown bolted connection, to produce an acoustic signal for the unknown bolted connection; a recording device, wherein the recording device s capable of recording the acoustic signal for the unknown bolted connection to produce recorded acoustic data; and a processor in communication with the recording device, wherein the processor is programmed with a decision tree model for predicting tightness of a bolted connection based on features extracted from acoustic data, and wherein the processor is programmed to extract the features from the recorded acoustic data and predict tightness of the unknown bolted connection.

7. The system of claim 6, wherein the impact device is a hammer.

8. The system of claim 6, wherein the features are power spectrum density (PSD) at selected frequency ranges.

9. The system of claim 8, wherein the selected frequency ranges are 200-300 Hz, 300-400 Hz, 400-500 Hz, and 500-600 Hz.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0014] FIG. 1 shows a schematic and a flow chart of a preferred embodiment of the present method.

[0015] FIG. 2A shows an image of a subsea flange used in an experimental analysis.

[0016] FIG. 2B shows an image of a selection of one bolt from a flange used in an experimental analysis.

[0017] FIG. 2C shows an image of sound recordings being taken from a selected bolt in a flange in an experimental analysis.

[0018] FIG. 3 shows impact-induced sound signals recorded by a smart phone in an experimental analysis at a bolt looseness of (a) 0 lb.ft, (b) 20 lb.ft, (c) 40 lb.ft, (d), 60 lb.ft, and (e) 80 lb.ft.

[0019] FIG. 4A shows 3D scatterplots for different bolt looseness conditions as determined by extracted Features 0, 2, 3.

[0020] FIG. 4B shows 3D scatterplots for different bolt looseness conditions as determined by extracted Features 0, 1, 3.

[0021] FIG. 5 shows an exemplary decision tree model for identifying bolt looseness from acquired impact data, in accordance with preferred embodiments described herein.

[0022] FIG. 6 shows accuracy percentage of test data obtained using an exemplary decision tree model to identify bolt looseness from acquired impact data repeated 100 times.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

[0023] The present disclosure relates to systems and methods for monitoring the tightness of bolts. In preferred embodiments, the systems and methods utilize percussion and machine learning.

[0024] The present systems and methods can be used in preferred embodiments in any conditions, including in difficult to reach bolted connections. In preferred embodiments, the systems may be used with a mechanism to apply a percussive tap and with a recording or monitoring device for detecting and recording the acoustic signals that are generated by the percussive tap. In further preferred embodiments, the acoustic signals generated by percussive taps applied to bolts in various looseness states are analyzed and a machine learning model is developed that allows for determining bolt looseness based on the acoustic signals. In additional preferred embodiments, the systems and methods are used in conjunction with a monitoring application to provide initial data and long-term data regarding the tightness of each bolt.

[0025] FIG. 1 shows a schematic and a flow chart of a preferred embodiment of the present method. Percussion is performed on the bolted structure to assess the health condition of the bolted connection. In this embodiment, an impact device, such as a hammer with a metal tip, is utilized to apply one or more taps to a portion of the bolted stricture, such as a corner of the nut. The sound generated by the bolted structure is recorded on a suitable recording device, such as the “voice record” function of a smart phone. Signal features are preferably extracted from the recorded signals. In a preferred embodiment, power spectrum density (PSD), with selected frequency ranges, is used as a signal feature to recognize and classify the recorded acoustical tapping data. The frequency details of the features in this preferred embodiment are listed in Table 1 below. For each feature, the total energy of each PSD segment is computed by the summation of the PSD sampling values at the corresponding frequency range.

TABLE-US-00001 TABLE 1 Feature name Frequency Feature (0) 200-300 Hz Feature (1) 300-400 Hz Feature (2) 400-500 Hz Feature (3) 500-600 Hz

[0026] In preferred embodiments, to identify the level of bolt looseness, a simple machine learning model using a decision tree method is adopted. A decision tree is a type of machine learning model that arrives at a final output based on a series of decisions made from multiple conditions. The decision tree may be visualized as an algorithm containing conditional if-then statements. In preferred embodiments, a portion of test results collected from different bolt looseness values may be used as training data to build the decision tree model. The remaining portion of the data may be used to verify the model accuracy.

[0027] Preferred embodiments described herein relate to a method for monitoring the tightness of a bolted connection. In initial steps of the method, a decision tree model is developed. The method may include using an impact device, such as a hammer, to apply at least one percussive tap to a portion of multiple bolted connections, wherein each bolted connection has a known tightness, to produce an acoustic signal for these known bolted connections. The acoustic signals for these known bolted connections are then recorded on a recording device to produce acoustic data corresponding to a variety of known bolt tightnesses. Features from the acoustic data are extracted, where the features are correlated with the known tightness of each bolted connection. The features and known tightness of each bolted connection are then used to generate a decision tree model for predicting tightness of a bolted connection using an acoustic signal. Once the decision tree model has been generated, further steps in the method may include using the decision tree model to predict tightness for a bolted connection having an unknown tightness. In these steps, the impact device is used again to apply at least one percussive tap to a portion of the bolted connection having an unknown tightness to produce an acoustic signal. A recording device records the acoustic signal for the unknown bolted connection to produce unknown acoustic data. The same features are extracted from the unknown acoustic data and the decision tree model is used to predict the tightness of the unknown bolted connection.

[0028] Additional preferred embodiments include a method for validating the decision tree model, optionally before it is used on an unknown bolted connection. In these steps, the impact device is used to apply at least one percussive tap to a series of bolted connections having known tightnesses. Acoustic signals are recorded by a recording device to produce acoustic data, and the features used in the decision tree model are extracted from the acoustic data. The decision tree model is then used to predict the tightness of the bolted connections. The tightness of the bolted connections as predicted by the decision tree model is then compared to the actual known g less of these bolted connections in order to validate the decision tree model.

[0029] Additional preferred embodiments relate to a system for monitoring the tightness of a bolted connection, which includes an impact device. The impact device may be positioned to apply at least one percussive tap to a portion of a bolted connection, to produce an acoustic signal for the unknown bolted connection. The system also includes a recording device capable of recording the acoustic signal for the bolted connection to produce acoustic data. The system also includes a process in communication with the recording device. Any suitable processor can be used. The processor is programmed with a decision tree model for predicting tightness of a bolted connection based on features extracted from acoustic data, and is also programmed to extract these features from the acoustic data that is recorded and to predict tightness of the bolted connection.

[0030] In further preferred embodiments, the impact device is a hammer. In preferred embodiments, the extracted features are power spectrum density (PSD) at selected frequency ranges, preferably the frequency ranges of 200-300 Hz, 300-400 Hz, 400-500 Hz, and 500-600 Hz.

EXAMPLE

[0031] The specimen used in this example was a 12-bolt subsea flange recovered. from the field, as shown in FIG. 2A. One bolt from the flange was selected as the test sample, as shown in FIG. 2B. Five different looseness conditions of the bolt from fully loosened (0 lb.ft) to 80 lb.ft with 20 lb.ft increment were demarcated. For each looseness condition, the bolt was impacted 50 times. The smart phone was fixed at a pre-determined location and recorded the impact-induced sound from the bolt, as shown in FIG. 2C. The sampling rate of the signal was 48000 Hz.

[0032] The impact-induced sound signals recorded by the smart phone are shown in FIG. 3(a)-3(e), for each 20 lb.ft increment of bolt looseness from FIG. 3(a) 0 lb.ft to FIG. 3(c) 80 lb.ft. The 50 peaks on each plot represent the 50 hammer impacts. Due to the innate randomness resulting from manually controlled impacts, the magnitudes of the peaks are not uniform.

[0033] Four features were selected based on the summation of energy from four frequency bands of the data's PSD plot. FIG. 4A presents the 3D scatterplots for different bolt looseness conditions as determined by Features 0, 2, 3 and FIG. 4B shows the 3D scatterplots for different bolt looseness conditions as determined by Features 0, 1, 3. The figures suggest that the torque level of the bolt can be distinguished based on the combination of the selected features. Different bolt looseness levels can be clearly distinguished from the 3D scatterplots of the signal features extracted from the 200-600 Hz frequency range. However, the results also suggest that the correlation between features and the bolt looseness is nonlinear and this complicates the finding of a trend between feature values and bolt looseness.

[0034] In order to overcome the nonlinearity in the correlation, the decision tree method was utilized to further identify the bolt looseness from acquired impact data. For each torque level, 40 randomly selected data points from 50 impact signals were used to build a decision tree model, as shown in FIG. 5. The remaining 10 impact signals were used to test the feasibility and accuracy of the model. Therefore, a total of 200 impacts signals for all the torque level conditions were used for as training data and the remaining 50 signals were used as test data. This process was repeated 100 times, and the accuracy percentage of the test data for each time was calculated, as shown in FIG. 6. The average accuracy percentage of the test data was 96.94% for the process repeated 100 times. Thus, the decision tree model developed by the training data could accurately identify the bolt looseness condition for the test data.

REFERENCES

[0035] The following documents and publications are hereby incorporated by reference.

[0036] U.S. Pat. No. 6,962,082

[0037] U.S. Patent Pub. No. 2010/0089161

[0038] May, D. B. and Munt, P. W., 1979, Physiologic effects of chest percussion and postural drainage in patients with stable chronic bronchitis. Chest, 75(1), pp. 29-32.

[0039] Gallon, A., 1991. Evaluation of chest percussion in the treatment of patients with copious sputum production. Respiratory medicine, 85(1), pp. 45-51.

[0040] McGee S. R., 1995. Percussion and physical diagnosis: separating myth from science. Disease-a-Month, 41(10), pp. 645-692.

[0041] Kaewunruen, S. and Remennikov, A. M., 2007. Field trials for dynamic characteristics of railway track and its components using impact excitation technique. Ndt & E International, 40(7), pp. 510-519.

[0042] Sohn, H. and Farrar, C. R., 2001. Damage diagnosis using time series analysis of vibration signals. Smart materials and structures, 10(3), p. 446.

[0043] Deraemaeker, A., Reynders, E., De Roeck, G. and Kullaa, J., 2008. Vibration-based structural health monitoring using output-only measurements under changing environment. Mechanical systems and signal processing, 22(1), pp. 34-56.

[0044] Giurgiutiu, V., Zagrai, A. and Jing Bao, J., 2002. Piezoelectric wafer embedded active sensors for aging aircraft structural health monitoring. Structural Health Monitoring, 1(1), pp. 41-61.

[0045] Staszewski, W. J., Lee, B. C., Mallet, L. and Scarpa, F., 2004. Structural health monitoring using scanning laser vibrometry: I. Lamb wave sensing. Smart Materials and Structures, 13(2), p. 251.

[0046] Li, R. and He, D., 2012. Rotational machine health monitoring and fault detection using EMD-based acoustic emission feature quantification. IEEE Transactions on Instrumentation and Measurement, 61(4), pp. 990-1001.

[0047] Safavian, S. R. and Landgrebe, D., 1991. A survey of decision tree classifier methodology. IEEE transactions on systems, man, and cybernetics, 21(3), pp. 660-674.