An Open Access Journal
From: Response-based methods to measure road surface irregularity: a state-of-the-art review
Response-based methods | Advantages | Disadvantages |
---|---|---|
1. Road profile reconstruction | ||
1.1. Model-based approach | can deal with unforeseen situations that are not included in the data-driven training datasets. | - an accurate model is required - not all required response information is measurable - often only time domains |
1.2 Kalman filter/estimator | convenient, fast and simple | - a priori information about model errors - the tuning of the covariance matrix is usually done heuristically |
1.1.2 Observer | can include tyre dynamics | generally required knowledge of many vehicle parameters |
i. Sliding mode observer | - convergence of the errors | rather complicated for practical application |
ii. Q-parameterisation | - less computing cost and complexity for real-time implementation - better performance than KF | - the problem of extensive modelling - the sensitivity to speed variation in almost methods |
iii. Algebraic estimator iv. H∞ observer v. State observer vi. Jump-diffusion | - can work effectively in the framework of the active suspension system - overcome the drawbacks of KF | |
1.2 Data-driven approach (MLs) | - can use fewer parameters (e.g. only sprung or unsprung mass) - various ML techniques to be applied - does not require excessive system characterisation - required fewer analytical skills than parametric model | - impractical for an online road estimation due to computationally costly training datasets (e.g. 4655 s are required to train the ANN-based moded) |
1.2.1 Only MLs (e.g. ANN) | - able to detect potholes | - spatial frequency only - many vehicle parameters - not high accuracy and sensitivity to speed variation |
1.2.2 Combined MLs and others | - higher accuracy and performance - feasible for speed independent classifiers | |
i. with feature selection (e.g. WPT, FFT, PCA) | - can combine both time and frequency domains - able to classify terrain conditions | further complex modelling and understanding vehicle dynamics control mechanism |
ii. with KF | determination of the process noise variance before estimation | |
iii. with TF | - speed independent classifier with less training effort - able to detect potholes | |
1.3 Transfer function and others | required fewer parameters than the model-based approach | |
1.3.1 The transfer function (TF) | - easy, convenient and fast - frequency domain only | - not directly yield the expression of the excitation - limited to a constant speed (can be eliminated when combined TF with small time span) |
1.3.2 Others | ||
i. Cross-entropy | using only sprung and unsprung mass accelerations | too much computing time |
ii. Control-constraints | non-linear and complex models | remains costly |
iii. Bayesian parameter | low cost regardless of vehicle models | a priori information of the road is required |
iv. Microphone | feasible for the combination of techniques | the susceptibility to signal contaminations |
v. Modulating function | fulfil the real-time and noise suppression requirements | particularly for off-road vehicles |
2. Road roughness estimation and pothole detection | ||
2.1 Threshold-based methods | ||
2.1.1 Thresholds only | simplest methods (for PD purpose) with fix thresholds | threshold value varies with different types of smartphones, roads, vehicles, the condition of vehicles. |
2.1.2 Combined thresholds and others | overcome drawbacks of the threshold-based methods | |
i. with signal processing approaches | - able to detect the severity of potholes, differentiate potholes and humps | |
ii. with MLs to train detectors | - clustering of different road anomalies with simple algorithms | training datasets required which are not able to collect in some cases |
2.2 Signal processing | - able include both PD and RE in the same system - deal with GPS errors, data aggregation, device installation and orientation, crowdsourcing - higher performance and accuracy - suitable for data aggregation regardless of different configuration (e.g. velocity, orientation, suspension) | complicated analysis |
2.2.1 PSD and RMS acceleration | calculate IRI value | not able to detect a pothole |
2.2.2 RIF transformation | - feasible for connected vehicles - both PD and RE considering a fleet of vehicles | advanced signal processing |
2.2.3 Adaptive threshold (e.g. DWT) | less training effort as compared to MLs | |
2.3 Data-driven approach (MLs) | - various techniques to be applied to select the best alternative - easier to implement in the smartphone for crowdsourcing | a huge amount of training datasets required which are not able to collect in some cases |
2.3.1 Only MLs (e.g. ANN) | simple using of raw acceleration data and filter | |
2.3.2 Combined MLs and feature extraction | - able to eliminate speed dependence, suspension variation - higher accuracy |