Neural networks (NN) have spurred significant interest in automating machine condition monitoring, with many recent studies focusing on networks that take raw diagnostic signals as input. However, ...the complexity of these NNs has raised concerns about their physical justification and trustworthiness. Although explainable artificial intelligence (XAI) has developed promising tools to address these issues, the nature and size of raw diagnostic inputs have limited the effectiveness of XAI techniques in condition monitoring.
This paper proposes a novel approach combining domain transformation and discretisation techniques with the XAI technique of Shapely value additive explanation (SHAP). This methodology allows the input domain for the network to remain in the original time-domain, while the explanation domain can be chosen as frequency or time–frequency and its discretisation adapted to the specific diagnostic case. This flexibility makes networks trained on raw signals more interpretable, while the controllable domain discretisation facilitates the implementation of XAI with manageable computational costs. The proposed methodology is tested on numerical cases and experimental signals, including the UNSW gear-wear and CWRU bearing-fault datasets. This enhancement of XAI is applicable to common black-box classifiers, providing researchers with an effective tool to give physical justification to their models and bolster trust in AI for machine condition monitoring applications.
•Addressing interpretability in time-domain processing NNs for condition monitoring.•Novel methodology to explain a NN in a transformed domain to their processed data.•Novel discretisation techniques to decrease computational costs of the calculation.•Application to two numerical examples and gear and bearing fault diagnostics.
•Proposal of a methodology for designing new condition indicators.•Statistical optimality of the proposed condition indicators for detection.•Delivery of statistical thresholds.•Consideration of ...fault symptoms produced by non-Gaussianity and nonstationarity.•Interactions between non-Gaussianity and nonstationarity.
Recent studies in the field of diagnostics and prognostics of machines have highlighted the key role played by non-stationarity – often in the form of cyclostationarity – or non-Gaussianity – often in the form of impulsiveness as characteristic symptoms of abnormality. Traditional diagnostic and prognostic indicators (e.g. the kurtosis) are however sensitive to both types of symptoms without being able to differentiate them. In an effort to investigate how the signal characteristics evolve with the different phases of machine components degradation, this paper proposes a new family of condition indicators able to track cyclostationary or non-Gaussian symptoms independently. A statistical methodology based on the maximum likelihood ratio is introduced as a general framework to design condition indicators. It arrives with the possibility of setting up statistical thresholds, as needed for a reliable diagnosis. The methodology is validated with numerically generated signals and applied to the dataset made available by NSF I/UCR Center for Intelligent Maintenance Systems (IMS). This particular application shows high potential for bearing prognostics by providing condition indicators able to describe different phases of the bearing degradation process.
Macropitting is one of the most common gear wear mechanisms, and monitoring its initiation and propagation is critical to ensure the safe operation of gearboxes. Wear debris analysis (WDA) is widely ...used to monitor and diagnose gear macropitting and its diagnostic capability relies on good understanding of the relationship between wear particle features and those of actual gear macropits. However, the study on this relationship has been impeded by on-going challenges in characterising gear macropits during the wear process. As a result, detailed experimental data of macropits and the quantitative relationships between wear particles and macropits evolving in the fatigue process are scarce.
In order to address the above issues, this study examines macropits and wear particles generated on gear surfaces and their evolution in a gear fatigue process through employing a moulding method to obtain 38 batches of gear teeth moulds in the wear process. 2D characterisations of macropits on the moulds and sampled wear particles were conducted using optical microscopy. Qualitative and quantitative relationships between different features of macropits and wear particles were investigated. Results showed the concentration of wear particles had a similar trend to pitted area growth, yet with a delay. This behaviour was also found in 4 non-dimensional numerical features characterising the geometry of particles and pits. The results of this study provide further insight in the evolution of the fatigue process of gears and can assist in its monitoring.
•Qualitative and quantitative relationships between macropits and particles were investigated.•The evolution of wear particles' concentration was similar to the pitted area derivative with a delay.•The evolution of 4 non-dimensional features of wear particles and macropits showed similar trends with the same delay.
Second order cyclostationary (CS2) components in vibration or acoustic emission signals are typical symptoms of a wide variety of faults in rotating and alternating mechanical systems. The square ...envelope spectrum (SES), obtained via Hilbert transform of the original signal, is at the basis of the most common indicators used for detection of CS2 components. It has been shown that the SES is equivalent to an autocorrelation of the signal׳s discrete Fourier transform, and that CS2 components are a cause of high correlations in the frequency domain of the signal, thus resulting in peaks in the SES. Statistical tests have been proposed to determine if peaks in the SES are likely to belong to a normal variability in the signal or if they are proper symptoms of CS2 components. Despite the need for automated fault recognition and the theoretical soundness of these tests, this approach to machine diagnostics has been mostly neglected in industrial applications. In fact, in a series of experimental applications, even with proper pre-whitening steps, it has been found that healthy machines might produce high spectral correlations and therefore result in a highly biased SES distribution which might cause a series of false positives. In this paper a new envelope spectrum is defined, with the theoretical intent of rendering the hypothesis test variance-free. This newly proposed indicator will prove unbiased in case of multiple CS2 sources of spectral correlation, thus reducing the risk of false alarms.
•Logarithmic variance stabilisation in CS2 indicators.•Unbiased test for multiple CS2 sources.•Automated damage diagnostics in complex machines.
A novel methodology for automated wear mechanism and severity assessment combining surface replication, imaging and deep learning is proposed. A large dataset of images of gear teeth moulds was built ...and covers abrasive wear, macropitting and scuffing, and three severity levels for each mechanism, i.e., mild, moderate and severe. A two-level inference methodology was implemented, based on a first convolutional neural network (CNN), which contains multiple convolutional layers and is commonly used for image classification, for wear mechanism identification, followed by three CNNs for wear severity estimation. The first level obtained a test classification accuracy of 98.22% and the second of 95.16% on average. The two-level system was also applied to full tooth flank mould images to generate wear mechanism and severity maps showing the geographical distribution of wear.
•Automated gear wear mechanism and severity assessment achieved by deep CNNs.•A large dataset of gear teeth mould images was built from gear tests.•Data labelling against established guidelines to minimise subjectivity.•A two-level inference methodology was implemented and obtained high test accuracies.•Wear mechanism and severity maps for full tooth flanks generated.
Neural networks (NNs) have attracted a lot of attention in recent years in automated condition monitoring assessment. Scientists have successfully demonstrated that NNs can analyze mould and oil ...images simply on the basis of labeled data, without the need for expert knowledge, and provide an accurate indication of the wear mechanism and wear severity present. However, these results also raise questions about the generalizability and way of finding solutions of NNs in the field of severity and wear mechanism analysis. Since NNs are designed as a black-box, it is not obvious without further analysis which features of the input data are particularly relevant for the classification. Methods from the field of explainable Artificial Intelligence (XAI) can provide this analysis. If it is known which input areas of an image are particularly relevant for the NN, a targeted analysis can be used to make a statement about the meaningfulness of the solution finding and thus strengthen the trust in the NN. In this work, Layer-wise Relevance Propagation (LRP) is applied to an established convolution neural network (CNN) for tribological image analysis. Subsequently, an automated global meta-analysis of the LRP results on the relevant clusters of the mould and oil gearbox images is used to make a statement about the way the network is classified and compared with the approach of experts. This analysis proves that a CNN is able to base its classification on similar guidelines as a domain expert. Furthermore, it is shown that a misclassification of the network, which lowers the accuracy, is not due to a wrong assignment, but rather to the unclear manually performed labeling. The output of the method is an overview, which lists the most defining features for each severity class. Based on this, a domain expert can reliably and quickly determine whether the NN uses meaningful features to determine the solution.
•CNN interpretability & trust in gearbox wear images with LRP.•First XAI adaptation for global wear debris analysis.•Extract meta parameters, compared to expert guidelines.•Auto-extract crucial classification features.
•Traditional band-selection methods focus on impulsiveness or cyclostationarity.•The two properties are often entangled in fault signals, giving misleading results.•The properties can be separated, ...and cyclostationarity appears most important.•Log-cycligram proposed to capture cyclostationarity separately from impulsiveness.•Performance of new method validated against existing tools on multiple datasets.
The demodulation of machine signals is a key step for the diagnostics and prognostics of components such as rolling element bearings. Whereas diagnostic approaches could perform a cyclostationary analysis over the full spectral band (i.e. using cyclic-spectral maps), in order to extract time domain and statistical features for prognostics, a pre-processing filtering step is required to extract the often-weak fault-symptomatic signal components. A series of techniques derived from the original idea of the kurtogram have been proposed in previous studies for the selection of this optimal demodulation band. All of these methodologies have been designed to identify signal components with high impulsiveness (non-Gaussianity) or strong second-order cyclostationarity, both assumed to be typical characteristics of bearing fault signals. However, a recent series of theoretical works has shown how non-Gaussianity and non-stationarity (in its cyclic form), despite being clearly distinct properties, are practically entangled in bearing signals, and can be easily confused by indices such as kurtosis (implicitly assuming stationarity) or second-order cyclostationary (CS2) indicators (implicitly assuming Gaussianity). In addition, it has been shown that generalised Gaussian cyclostationary models are effective tools to describe and separate these two properties in bearing signals. Partial evidence seems to show that the cyclostationary property is dominant and more clearly indicative of a bearing fault, whereas impulsiveness is potentially misleading and not uniquely attributable to the bearing component of the signal. In this paper, we therefore propose a new statistically robust band selection tool which can capture cyclostationarity separately from non-Gaussianity. The tool, coined the log-cycligram (LC), is based on the strength of target cyclic frequency components in the spectrum of the log envelope (LES), and so potential fault frequencies must be known in advance. The effectiveness of the method is validated against the traditional kurtogram and a range of other existing techniques on both numerical and experimental datasets.
•Anomaly detection in condition monitoring using healthy vibrations for training.•Design of data-driven model architectures that integrate a physical/statistical understanding of different ...faults.•LSTM/SVM capable of detecting new deterministic components introduced by gear faults.•A two-step LSTM regression + one-class SVM for detecting new random components caused by bearing faults.•SVM capability in fault detection when vibrations collected using different sensors.
Fault detection is a critical step for machine condition monitoring and maintenance. With advances in machine learning technologies, automated faulty condition identification can be achieved by training an artificial intelligent (AI) model if sufficient data is available. In most practical applications, it is unlikely that enough data from faulty cases are available for supervised training of AI models for anomaly detection. This work is based on training data that are exclusively constituted of healthy signals (i.e., semi-supervised) and aims to develop an automated algorithm capable of identifying any abnormal mechanical behaviour captured by vibration measurements. A deep learning method with long short-term memory (LSTM) architectures combined with a one-class support vector machine (SVM) is used to separate abnormal data from normal vibration signals collected during an endurance test of a reduction gearbox and helicopter test flight data measured by multiple sensors situated at different locations of the aircraft. For the gearbox dataset, a detailed understanding of the physical mechanisms by which different types of faults (gear wear and bearing faults) affect the vibration signal led to the design of two anomaly detection architectures: i) an LSTM regression + one-class SVM for detecting new deterministic components introduced by gear faults; and ii) a two-step LSTM regression + one-class SVM for detecting new random components caused by bearing failures. For the helicopter dataset, which does not contain consecutive time-series, we show that the LSTM regression is not advantageous, and a better performance can be achieved by a simpler one-class SVM outlier detection based on statistical features. This work contributes to the field of machine condition monitoring by introducing a novel two-step LSTM configuration for removing deterministic components associated with dominant gear signals in the first step and removing the ‘residual deterministic’ components associated with varying gear signals in the second step.
•A vibration-based wear mechanism identification procedure is proposed.•Wear evolution is tracked using an indicator of vibration cyclostationarity (CS).•The correlation between surface features and ...vibration characteristics is investigated.•Methods validated using lubricated and dry gear wear tests.
Fatigue pitting and abrasive wear are the most common wear mechanisms in lubricated gears, and they have different effects on the gear transmission system. To develop effective methods for online gear wear monitoring, in this paper, a vibration-based wear mechanism identification procedure is proposed, and then the wear evolution is tracked using an indicator of vibration cyclostationarity (CS). More specifically, with consideration of the underlying physics of the gear meshing process, and the unique surface features induced by fatigue pitting and abrasive wear, the correlation between tribological features of the two wear phenomena and gearmesh-modulated second-order cyclostationary (CS2) properties of the vibration signal is investigated. Differently from previous works, the carrier frequencies (spectral content) of the gearmesh-cyclic CS2 components are analysed and used to distinguish and track the two wear phenomena. The effectiveness of the developed methods in wear mechanism identification and degradation tracking is validated using vibration data collected in two tests: a lubricated test dominated by fatigue pitting and a dry test dominated by abrasive wear. This development enables vibration-based techniques to be used for identifying and tracking fatigue pitting and abrasive wear.
Gear wear introduces geometric deviations in gear teeth and alters the load distribution across the tooth surface. Wear also increases the gear transmission error, generally resulting in increased ...vibration, noise and dynamic loads. This altered loading in turn promotes wear, forming a feedback loop between gear surface wear and vibration. Having the capability to monitor and predict the gear wear process would bring enormous benefits in cost and safety to a wide range of industries, but there are currently no reliable, effective and efficient tools to do so. This paper develops such tools using vibration-based methods.
For this purpose, a vibration-based scheme for updating a wear prediction model is proposed. In the proposed scheme, a dynamic model of a spur gear system is firstly developed to generate realistic vibrations, which allows a quantitative study of the effects of gear tooth surface wear on gearbox vibration responses. The sliding velocity and contact forces from the model are used in combination with the well-known Archard wear model to calculate the wear depth at each contact point in mesh. The worn gear tooth profile is then fed back into the dynamic model as a new geometric transmission error, which represents the deviation of the profile from an ideal involute curve and is thus zero for perfect gears. New vibration responses and tooth contact forces are then obtained from the model, and the process repeated to generate realistic gear wear profiles of varying severities. Since the wear coefficient in the model is not constant during the wear process (and in any case is difficult to estimate initially), measured vibrations are compared with those generated by the model, so as to update the coefficient when a deviation from predictions is detected. With the continually updated dynamic wear model, the wear process can be well monitored, and at any particular time the best possible prediction of remaining useful life can be achieved. The paper illustrates the ability and effectiveness of the proposed scheme using measurements from a laboratory gear rig.
•We propose a vibration-based scheme for updating gear wear prediction.•Prediction based on models of gearbox dynamics and abrasive wear.•Updating of wear constant based on comparing simulated and measured vibration.•Scheme allows reliable wear prediction using simple modelling tools.•Scheme experimentally validated on run-to-failure dry test with spur gears.