Vibration-Based Strength Grading of Sawn Timber Using Piezoceramic Transducers and One- Dimensional Convolutional Neural Networks

1 (3)

Vibration-Based Strength Grading of Sawn Timber Using Piezoceramic Transducers and One- Dimensional Convolutional Neural Networks

Osama Abdeljaber, Anders Olsson and Welf Löwe

Core research areas: timber engineering (machine strength grading), computer science (deep learning)

1. Introduction

Nondestructive evaluation of structural timber properties such as strength and stiffness is a key issue for the sawmilling industry. Accurate strength grading of timber boards leads to a more effective utilization of material.

The timber grading process is often carried out using certified machines that classify boards into certain strength classes according to the three grade determining properties (i.e. bending strength, modulus of elasticity (MoE) and density). The grading machines measure certain indicating properties (IPs) that are believed to be strongly correlated with one or more of the grade determining properties [1].

In this seed project, we propose a new data-driven technique that utilizes piezoceramic transducers together with one-dimensional convolutional neural networks (1D CNNs) for accurate strength grading of structural timber.

The idea is to use a number of piezoceramic transducers as actuators that apply a wideband frequency excitation on the tested board. The vibration response of the board under this excitation is measured by another set of piezoceramic transducers, which act as sensors. The measured response is then analyzed by a 1D CNN to estimate the bending strength of the board.

2. Problem definition and value

Most of the current strength grading machines utilize axial dynamic excitation to calculate the IPs required for strength classification. Such machines apply an impact load on one end of the tested board and measure the response in the axial direction at the other end using a microphone. Fast Fourier Transform (FFT) is then applied to obtain the frequency spectrum of the response. The global dynamic MoE is calculated in terms of the frequency of the first peak of the spectrum, which corresponds to the first axial resonance frequency. The resulting quantity is then used as the basis for estimating the bending strength and MoE of the board [2].

This approach has been widely used in the sawmilling industry due to its simplicity and cost efficiency. However, experimental studies showed that the correlation between the axial dynamic MoE and the bending strength is rather weak. The reason behind this is the fact that the bending strength in timber is highly dependent on local variations in the density and fiber directions, which cannot be captured by a global IP such as axial dynamic MoE [1]. Therefore, axial dynamic MoE is often used in conjunction with high resolution X-ray or laser scanners that provide additional information regarding the local variations of material parameters [2].

As a potentially less expensive and more accurate alternative to X-ray and laser scanning, we propose a novel vibration-based strength grading system that makes use of piezoceramic transducers and 1D CNNs. The proposed technique involves using piezoceramic patch transducers to excite the tested board at several locations with a wideband Gaussian random excitation. Meanwhile, another set of patch transducers is used to measure the vibration response at certain points along the board for a certain amount of time. The low frequency content of the measured response mainly reflects the global properties of the board, while the high frequency content is more likely to be affected by the local variations in density and stiffness. Rather than relying on hand-crafted IPs, a 1D CNN is trained to identify the bending strength directly by processing the measured vibration signals in the time domain.

Techniques that involve piezoceramic sensors/actuators have shown great potential in several structural health monitoring applications in steel and concrete structures [3][4]. However, the literature is scarce when it comes to using piezoceramic transducers in timber engineering. The proposed project will probably be the first effort to utilize piezoelectric transducers in strength grading of structural timber.

1D CNNs are deep leaning tools that have become the de facto standard for classification of signals [5]. Unlike conventional classifiers that require extraction of certain hand-crafted features from the signals beforehand, 1D CNNs combine both feature extraction and classification into a single compact learning body [6]. This is particularly important when dealing with strength grading of timber because hand-crafted IPs extracted from the measured vibration signals might not optimally represent the most characteristic information in these signals.

On the other hand, optimal features extracted by a well-trained 1D CNN are more likely to reflect both global and local properties of the analyzed board, resulting in a better strength grading performance.

3. Objectives

Developing the proposed strength grading system requires taking several design parameters into consideration

including the number of sensors/actuators per unit length and their orientations, the frequency range, amplitude

2 (3)

and duration of the excitation signal, and the architecture of the 1D CNN. The best way for finding reasonable values for these design parameters is to conduct several experimental tests with different configurations.

Therefore, the objectives of the proposed seed project are:

1. To carry out nondestructive vibration tests on a relatively small sample of timber boards (about 30 boards) using different configurations of piezoceramic sensors/transducers as well as different excitations.

2. To conduct destructive tests in order to obtain the actual bending strength of each board.

3. The resulting dataset, consisting of vibration signals as inputs and the corresponding bending strengths as outputs, will be used to train and validate several 1D CNN models (a 1D CNN for each configuration).

4. To test the trained 1D CNNs over a limited number of boards (about 10 boards) in order to determine the best configuration.

5. To investigate interest and opportunities for collaboration with industry partners within established networks.

The outcome of this seed project (i.e. the preliminary design of the proposed system) will be the basis for an externally funded project. The objectives of the externally funded project are:

1. To optimize the preliminary design in order to improve the accuracy and grading speed of the proposed system.

2. To conduct large-scale nondestructive and destructive tests over a large sample of timber boards (500 boards). The resulting dataset will be used to train and validate the 1D CNN.

3. To build a prototype strength grading machine based on the proposed technique.

4. To evaluate possibilities and limitation of the technique for use under sawmill production conditions and speed.

4. Expected Outcomes

1. The expected outcomes of the seed project are:

2. A preliminary design of the proposed strength grading system.

3. A seminar for research colleagues, with relevant research interests, at Linnaeus University.

4. A conference paper to be presented in a scientific conference.

5. An external funding proposal to be submitted to either FORMAS or KK Foundation.

5. Consortium

Osama Abdeljaber has hands-on expertise in utilizing deep learning tools in structural health monitoring and vibration control. Osama will lead the seed project and carry out the majority of the work, including nondestructive and destructive tests as well as training and validation of 1D CNNs. He will also contribute in the writing of the conference paper and external funding proposals.

Anders Olsson has long experience from research on strength grading of sawn timber and a large network with industry. Anders will act as advisor, take part in planning and interpretation of results, initiate discussions with industry representatives and contribute in the writing of research proposals.

Welf Löwe is a Professor at the Department of Computer Science with insight regarding interests and competences of colleagues at LNU who could participate in the research and help developed the ideas presented in this proposal. His role in the seed project will be to help develop the network of researchers.

6. Activities & Time plan

The project will be carried out according to the objectives stated in Section 3. The activities and the time plan are detailed below:

Activity May

2020

June 2020

July 2020

Aug 2020

Sept 2020

Oct 2020 1 Nondestructive tests

2 Destructive tests

3 Training and validation of 1D CNNs 4 Final testing of the trained 1D CNNs 5 Preparation of an external funding proposal 7. Budget

The total budget of the seed project is 100 KSEK distributed as follows:

 Material and equipment for testing: 5 KSEK

 Working time for Osama Abdeljaber: 80 KSEK

 Working time for Anders Olsson 15 KSEK

 Working time for Welf Löwe Other DISA funding

3 (3)

References

[1] Olsson A, Pot G, Viguier J, Faydi Y, Oscarsson J. Performance of strength grading methods based on fibre orientation and axial resonance frequency applied to Norway spruce (Picea abies L.), Douglas fir (Pseudotsuga menziesii (Mirb.) Franco) and European oak (Quercus petraea (Matt.) Liebl./Quercus robur L. Ann For Sci 2018;75.

https://doi.org/10.1007/s13595-018-0781-z.

[2] Olsson A, Oscarsson J. Strength grading on the basis of high resolution laser scanning and dynamic excitation: a full scale investigation of performance. Eur J Wood Wood Prod 2017;75:17–31. https://doi.org/10.1007/s00107- 016-1102-6.

[3] Park S, Kim JW, Lee C, Park SK. Impedance-based wireless debonding condition monitoring of CFRP laminated concrete structures. NDT E Int 2011;44:232–8. https://doi.org/10.1016/j.ndteint.2010.10.006.

[4] Talakokula V, Bhalla S, Gupta A. Monitoring early hydration of reinforced concrete structures using structural parameters identified by piezo sensors via electromechanical impedance technique. Mech Syst Signal Process 2018;99:129–41. https://doi.org/10.1016/j.ymssp.2017.05.042.

[5] Abdeljaber O, Avci O, Kiranyaz S, Gabbouj M, Inman DJ. Real-time vibration-based structural damage detection using one-dimensional convolutional neural networks. J Sound Vib 2017;388:154–70.

https://doi.org/http://dx.doi.org/10.1016/j.jsv.2016.10.043.

[6] Kiranyaz S, Ince T, Gabbouj M. Personalized Monitoring and Advance Warning System for Cardiac Arrhythmias.

Sci Rep 2017;7. https://doi.org/10.1038/s41598-017-09544-z.