Influential factors for the emission inspection results of urban in-use vehicles: From an ensemble learning perspective

Figures & data

Table 1. Data composition.

Feature	Subdivision item (corresponding integer code)	Feature type
Vehicle weight (A1)	<3995 kg	Numerical
Displacement (A2)	[0,1], (1,1.6], (1.6,2.5], (2.5,4.0], and (4.0,] (1–5)	Sequencing
Engine type (A3)	Others (0) and ignition (1)	Category
Transmission type (A4)	Manual (0), automated-manual (1), and automatic (2)	Category
Driving mode (A5)	Front-wheel drive (0), rear-wheel drive (1), multi-axis drive (2), timely four-wheel drive (3), time-sharing four-wheel drive (4), full-time four-wheel drive (5), and others (6)	Category
Fuel supply system (A6)	Closed-loop electronic injection, open-loop electronic injection, mechanical pump, in-line pump, high-pressure common rail injection, and carburetor	Category
Three-way catalytic converter (A7)	No (0) and yes (1)	Category
VIN (A8)	VIN8 (first 8 digits of the VIN): 5232 category, VIN3 (first 3 digits of the VIN): 302 category. In this paper, VIN3 was mainly used.	Category
Emission standard (A9)	National I–VI emission standards (1–6)	Sequencing
Fuel specification (A10)	92# (0), 95# (1), 98# (2), and others (3)	Sequencing
Dual fuel (A11)	CNG and electricity (0), LPG and electricity (1), gasoline and CNG (2), and no (3)	Category
Vehicle type (A12)	9 categories including small-sized buses, light trucks and minibusses	Category
Accumulated mileage (B1)	<600,000 km	Numerical
Vehicle usage days (B2)	<9,995 d	Numerical
Vehicle use character (B3)	16 categories including non-operating, net lease, learner-driven vehicle, taxi passenger transport, and freight transport	Category
Inspection environment temperature (C1)	<45°C	Numerical
Inspection environment relative humidity (C2)	<85%	Numerical
Testing equipment manufacturer (C3)	8 manufacturers	Category
Emission inspection agency/testing system (C4)	65 emission inspection agencies and 385 testing systems. An emission inspection agency contains multiple testing systems. In this paper, emission inspection agencies were focused on.	Category
Predicting object: inspection results	qualified (0), unqualified (1)	Category

Notes: The corresponding relationships between integer codes of some features are explained in the table.

Table 2. Model evaluation indexes (Bruce and Bruce 2018).

Parameter for evaluation	Definition
Accuracy	Percentage of correct classification, a measure of overall error
Precision	True positive samples /(true positive samples + false positive samples), the ability of the model to distinguish negative samples
Recall	True positive samples /(true positive samples + false negative samples), the ability of the model to recognize positive samples
F1 score	Harmonic mean of precision and recall
AUC	Area under the receiver operating characteristic (ROC) curve, to evaluate the quality of sample rank predicted
Kappa coefficient	Kappa is an omnibus index of agreement, and can also be used to measure the effect of classification

Notes: True positive samples refer to the samples that are positive in fact and prediction, false-positive samples are the samples that are falsely predicted as positive samples (true samples are negative samples), and false-negative samples mean the samples that are falsely predicted as negative samples (true samples are positive samples).

Table 3. Results of model evaluation.

Model No.	1		2		3
VIN	VIN8		VIN8		VIN3
Agency/testing system	Agency		Testing system		Agency
Evaluation method	Testing set	10-fold cross validation	Testing set	10-fold cross validation	Testing set	10-fold cross validation
Accuracy	71.71%	78.11%	70.77%	77.55%	74.00%	79.52%
Precision	24.70%	76.12%	24.28%	75.47%	25.95%	77.75%
Recall	77.35%	82.77%	78.84%	82.58%	74.20%	83.12%
F1_score	0.3744	0.7922	0.3713	0.7878	0.3845	0.8030
AUC	0.7418	0.8560	0.7431	0.8489	0.7409	0.8710
Kappa	0.2500	– –	0.2449	– –	0.2654	– –

Table 4. Comparisons of evaluation indexes before and after parameter adjustment (testing set validation, Model 3).

Model evaluation	Accuracy	Precision	Recall	F1 score	AUC	Kappa
Original model	74.00%	25.95%	74.20%	0.384	0.740	0.265
After parameter adjustment	78.71%	29.92%	70.41%	0.420	0.750	0.314
Change	+4.71%	+3.97%	−3.79%	+0.035	+0.010	+0.049

Figure 1. Model calculation importance.

Figure 2. Changes in the evaluation results after removing single feature.

Figure 3. Feature SHAP value for a single sample.

Table 5. Proportion of the feature SHAP values.

Feature	Emission inspection agency	VIN3	Vehicle usage days	Transmission type	Accumulated mileage	Emission standard	Temperature	Driving mode
Proportion	18.38%	15.01%	9.60%	9.41%	9.21%	5.82%	5.54%	5.50%
Feature	Vehicle use character	Relative humidity	Vehicle weight	Fuel specification	Vehicle type	Equipment manufacturer	Displacement	Dual fuel
Proportion	4.24%	3.86%	3.52%	3.33%	2.63%	2.61%	0.67%	0.66%

Figure 4. SHAP summary plot [(1)] and feature ranking plot [(2)].

Figure 5. SHAP dependency graph (category features).

Figure 6. SHAP dependency graph (numerical features).

Figure 7. Differences in test pass rates among emission inspection agencies.

Bruce, P., and Bruce, A. 2018. Practical statistics for data scientists. Translated by G. Lei, Beijing: Posts & Telecom Press.

 Google Scholar