Review of Current Trends and Uses of Machine Learning for Discrete Acoustic Emission Interpretation

Numerous reviews exist, describing AE and its various applications in details [23] and in regard to other NDT techniques [24]. Many reviews present the usage of AE on a specific material: concrete [25] and corrosion of reinforced concrete [28], composite damage [17, 29], and steel in the context of Stress Corrosion Cracking (SCC) [26]. While some reviews discuss in details the application of ML methods to specific materials [17, 26] there is, to the authors’ best knowledge, no review discussing the usage of ML methods on multiple materials trying to pinpoint common trends and intrinsic differences exist. This review will focus on discrete AE studies as the methods applied to discrete AE are distinct from those used for continuous AE. The discussed materials are concrete, composite and steel with a focus on corrosion evaluation, and some additional references to other domains such as wood or bearings. Despite important differences, methodologies developed for specific materials have been successfully generalized to different structures [30]. This review will try to explore how ML can help to interpret recorded AE signatures and datasets. The multiple contributions, requirements and limitations of ML will be discussed, such as the capacity to discriminate specific damage types or phases of deterioration, to correlate and evaluate the most pertinent characteristics as well as offer data cleaning and visualisation possibilities.

AE is applied to a variety of applications and structures with a similar measurement chain, schematized in Fig. 1:Although discrete AE aims at “listening” to occurring phenomena, especially degradation phenomena, independently of the considered material, they may have very different natures in practice. For composite structures, they include fibre cracks, matrix breaking, fibre-matrix debonding and delamination [17, 29, 36]. For concrete samples, shear cracks, tensile cracks as well as corrosion for reinforced concrete structure are expected phenomena [24, 28]. Regarding corrosion, various forms exist such as SCC, pitting corrosion, uniform corrosion, crevice corrosion, galvanic corrosion etc. [26]. For pure steel structures such as bridge cables or vessel, phenomena include plasticity [3], crack initiation or crack propagation [12, 37, 38]... In addition to the differences in the multiple damaging mechanisms, these materials have specific characteristics influencing the wave propagation: metallic structures tend to be homogeneous and isotropic, composite ones are anisotropic and concrete structures are inhomogeneous, inducing scattering effects and reducing the quality of the measured signal. Damping is also very different, as it tends to be high in composite and concrete and lower in steel structures, affecting the amplitude and the frequency components of the signal [39‐41]. To adjust to these differences and to record accurately the degradation of structures, several acquisition parameters need to be set. The value of these parameters depends on the physical characteristics of the structure (size, damping, elasticity) and environmental characteristics (temperature, background noise, etc.). To estimate the propagation of AE waves in the structure, Hsu-Nielsen tests are commonly realised at increasing distances of the sensors to determine temporal acquisition parameters such as peak detection time (PDT), hit definition time (HDT) and hit lockout time (HLT). The sampling frequency and the sensor can be chosen depending on the application using literature knowledge. If no knowledge is available, broadband sensors must be used as a first step. To determine the acquisition threshold, background noise must be measured before the test. The threshold must be fixed a few dB over the noise level to accommodate for noise instability. Noise measurement must not be neglected, and external noise sources must be investigated thoroughly to determine analogue and numerical filter’s frequency or more specific data cleaning strategies adapted to the studied case. Different data cleaning strategies are discussed in Section 2.3.

For long term monitoring, multiple experimental precautions can be taken to ensure the quality of measurements. A first necessity is to evaluate the noise sources over a representative period of time prior to testing. Background noise can also be recorded at fixed intervals when no event is measured during the entire test to control its stability and help adjust acquisition parameters if needed. Finally, an additional sensor can be added to the experiment al setup to record external noise and to help detect external events, as shown by Sibil et al. [32]. A two sensors setup was used to localize defects and to filter signals associated with background noise. This is important as noise and outliers can influence classification results. In particular, differences in the background noise may alter measured signals and induce a bias during training. The classification will hence represent experimental differences rather than the presence of multiple damage mechanisms. This has been pointed out by Pomponi and Vinogradov [45], who developed a real-time clustering method and showed that not taking background noise into account could significantly alter the results. This has also been investigated by Doan et al. [46], who proposed a ML framework that estimates and projects background noise in the feature space. Calabrese et al. [47] performed corrosion monitoring for approximately 7 months, demonstrating the influence of day and night cycles on AE data, and proposed a framework to filter experimental noise. Barat et al. [48] investigated different noise filtering strategies adapted to electromagnetic noise, mechanical impact and friction noises. Finally, sensor-structure coupling may also be monitored during the test, as the coupling quality has been noted to influence amplitude and frequency parameters [17, 49, 50]. As it is common to cluster datasets acquired from different experiments [42‐44], specific experimental precautions must be taken to perform repeatable sensor coupling and to have similar background condition. The impact of potential differences is discussed in Section 3.3.4.

Bild vergrößern

To ensure the quality of data, data filtering process may be applied during or after the acquisition. Most commercial acquisition systems offer many tools to clean up data as to avoid recording signals of bad quality. The triggering of acquisition based on an amplitude threshold is a first form of filter by itself, but other criteria can be added freely based on extracted parameters: an example is the amplitude-duration-based filter known as Swansong II, which is used in many concrete studies [78, 79] originating from Fowler et al. [80, 81]. The hypothesis behind this filter is that signals with high amplitude will naturally have a higher duration. Signals not following this hypothesis may originate from external unwanted noises such as mechanical friction or electrical noise. This filter has been refined by Zheng [82] to include peak frequency. Comparable data filtering strategies exist: Van Steen et al. [60] proposed a filtering method based on the ratio of amplitude between two time window of a signal. If the ratio is smaller than 2, the signal is considered noisy. Appalla et al. [83] considered the Root Mean Square (RMS) and the average frequency to filter out noisy signals following observations during their experiments. Calabrese et al. [47] proposed a noise filtering method based on the frequency of events, removing events during high activity periods. Data filtering strategies using thresholds should be implemented with precautions: first, thresholds in the literature are only relevant for the studied structure and environmental conditions and may not be reusable for different structures and conditions. In particular, these thresholds rely on the distance between the source and the sensors. Furthermore, removing signals outside a specific range may result in a loss of important information to assess the structure degradation. Nonetheless, because AE sensors have a high sensitivity, they are very prone to record experimental noise which can alter the results of classification or clustering and data cleaning techniques must be considered.

Modern ML frameworks for AE are composed of multiple important steps: data acquisition, cleaning and preparation, processing and interpretation. Many unsupervised and supervised models require the data to be normalised in order to ensure that all features are dimensionless and on comparable scales. There are several standardisation strategies such as standard scaling, robust scaling or min-max scaling. This step is important as this will alter the distances between samples, impacting clustering and classification algorithms. The importance of data cleaning is linked to the necessity of normalisation as outliers will impact the normalisation process because of the statistical bias they create. As a result, the data structure may not be representative without data cleaning. The impact of different normalisation strategies on classification performances has been explored by Singh and Singh [84]. The most common scalers are the standard scaler, especially for composite [44, 85], corrosion [86] and ceramics [32] damage monitoring and the min-max scaler, with applications to composite [35] and metal damage [87]. The min-max scaler is used whenever authors wanted a common limited range of values across features, while the standard scaler is used traditionally to obtain dimensionless, comparable, and centered features. The standardisation of data is also a mandatory step when using Principal Component Analysis (PCA).

Among current weaknesses of clustering for AE, Muir et al. [17] remarked in particular the lack of quality of some extracted features. The importance of selected subset of feature was also demonstrated by Sibil et al. [32] by comparing classification results over several subsets of features. A first solution to select an optimal subset is to remove the worst features or to keep only the best ones. Many feature selection strategies exist, with no clear agreement on the best one for AE purposes. It is important to separate the different strategies: dimensionality reduction algorithms will aim at resuming information in newer dimensions, feature suppression algorithms will remove features of low quality (for example of low variance) and, finally, feature selection algorithms will aim at selecting the best features. All these procedures are referred to as feature selection, which can bring confusion. In a supervised learning context, Ai et al. [18, 68] proposed a simple feature selection process involving the use of backward elimination, whereby classification performance is evaluated on all possible subsets of features iteratively, demonstrating an improved accuracy after suppression. In an unsupervised learning context, a basic procedure for selecting the best feature partition has been proposed by Sause et al. [88] based on the test of all possible feature combinations. For each number of clusters to investigate, the clustering performances are evaluated for each combination of features, resulting in a costly but thorough exploration of all feature partitions. However, it is preferable to use data-driven feature selection algorithms which do not rely on clustering performances, as a method relying on clustering metrics such as the silhouette score or the Davies-Bouldin index may not achieve the desired partition [46]. This is further discussed later on in Section 3.4.4. For example, Almeida et al. [35] used datasets containing only specific types of damage and Kendall’s Tau correlation to identify features highly correlated with identified damage mechanisms before using only this optimized subset of feature to perform unsupervised clustering. However, having access to datasets with distinct damages is difficult and cannot be generalized. An example of a simple data-driven feature selection algorithm was proposed by Traore [89], which is designed to remove redundant features. This algorithm suppresses iteratively features if they can be described as a linear combination of the others by checking whether the \(R^2\) value of the linear regression is greater than a given threshold. The complete algorithm is described in Appendix.

This can be seen as a generalisation of feature analysis through Pearson correlation [66, 90], where correlation matrices are used to remove highly correlated features and to evaluate feature importance. A parallel can be made with the PCA, a common dimension reduction technique which produces a reduced number of uncorrelated dimensions. These new dimensions are a linear combination of input features that aims to maximise the variance. By grouping correlated features and reducing the number of dimensions, PCA can improve classification results and interpretability. It was used in many AE studies, such as corrosion classification [42], composite damage clustering [10, 32, 91, 92] or concrete crack monitoring [54]. PCA is also notably used before applying the k-means algorithm as the two tools are linked, as demonstrated by Ding and He [93]. However, PCA relies on variance as a criterion for meaningfulness, which may not always be the most suitable approach. As highlighted by Placet et al. [94], PCA did not yield satisfactory results in their case study.

The Laplacian score (LS) algorithm is another popular feature selection algorithm that assigns a score to each feature based on their locality preserving power. The Laplacian score can be defined as [96]:where \(f_{ri}\) is the r-th feature of data point i and S is a weighted nearest neighbour graph.

As explained by the authors [96], \(\sum {_{ij}(f_{ri}-f_{rj})^2S_{ij}}\) is minimum when \(S_{ij}\) is maximum and \((f_{ri}-f_{rj})\) is minimum, meaning there is an edge between data points, and they are close to each other for this feature. Furthermore, this criterion prefers features with high variance. A good feature is then defined by a feature that respects the graph structure and has a high variance, resulting in a low Laplacian score. Note that another formulation has been used by authors and is in particular implemented in MATLAB, leading to having the most relevant features having the highest scores instead of the lowest ones. As this algorithm is dependent on the neighbour graph and distances, it is susceptible to the curse of dimensionality [95]. Therefore, different distances such as the Manhattan, Euclidean or Mahalanobis distances may be used. One difficulty of the Laplacian score is knowing which features to select once they are ranked. Most studies will select a subset of an arbitrary number of features corresponding to the best clustering scores. The Laplacian score has been successfully applied in recent works for concrete crack monitoring [54] and composite damage classification [87, 97, 98] to reduce the number of dimensions and for steel crack damage monitoring [99] in combination with a random forest framework. However, the proper choice of the number of dimensions, which should be the best trade-off between loss of information and reduction of the number of dimensions, is still an open question.

Xu et al. [31] compared the Laplacian score to other feature selection methods such as the Multi Cluster Feature Selection (MCFS) score and infinite feature selection for composite wind turbine monitoring. The MCFS score is a modern score first introduced in [100] by Deng Cai, Xiaofei He and Chiyuan Zhang. For reference, Deng Cai and Xiaofei He also authored the original Laplacian score paper [96]. It is presented as an improvement compared to Laplacian score or maximum variance based indices that may fail in multi-cluster scenarios. Based on spectral analysis techniques, this algorithm assumes that the characteristics have different powers to distinguish the separate clusters. Infinite feature selection is a novel graph-based feature selection method presented in [101] based around Markov chains. The original paper proposes not only a feature ranking method but also a subset selection method, which is lacking for the Laplacian score and MCFS. Xu et al. [31] demonstrated that 5 out of 15 features were described as relevant for their study case by all three algorithms, including RMS and average signal level. These two feature are exponentially correlated as noted by the authors, which is not accounted for by the feature selection algorithm. To avoid input ting the same information twice and altering clustering results, Xu et al. recommended filtering the correlated features to reduce redundancy. The impact of redundancy is detailed in Section 3.4.

As previously stated, one of the limitations of employing time-frequency representations is the large number of features that must be analysed and interpreted. A significant proportion of these features may contain little or no information, making time-frequency representations almost sparse. To reduce the number of dimensions and select the best or most relevant wavelet coefficients, many studies explored the use of Shannon’s entropy [74, 102]. Satour et al. [69] proposed a method based on CWT to characterise composite damage mechanisms, representing signals using the k most representative coefficients according to Shannon’s entropy and setting the rest of coefficients to 0. Zhang et al. [103] compared multiple wavelets by computing the ratio of energy to entropy, determining the most informative wavelets for rail defect detection. Ekici et al. [104] used Shannon’s entropy to determine the level of decomposition required to represent precisely the signal, increasing the frequency resolution until the entropy reached a fixed threshold and using wavelet packet energy and entropy as inputs for defect location. These techniques do not improve feature’s quality on their own, but allow creating a more representative set, eliminating redundancy and low added value features such as invariant features.