Development of Deep Learning-Based Algorithm for Extracting Abnormal Deceleration Patterns ^†

Abstract

A smart regenerative braking system for EVs can reduce unnecessary brake operations by assisting in the braking of a vehicle according to the driving situation, road slope, and driver’s preference. Since the strength of regenerative braking is generally determined based on calibration data determined during the vehicle development process, some drivers could encounter inconveniences when the regenerative braking is activated differently from their driving habits. In order to solve this problem, various deep learning-based algorithms have been developed to provide driving stability by learning the driving data. Among those artificial intelligence algorithms, anomaly detection algorithms can successfully separate the deceleration data in abnormal driving situations, and the resulting refined deceleration data can be used to train the regression model to achieve better driving stability. This study evaluates the performance of a personalized driving assistance system by applying driver characteristic data, obtained through an anomaly detection algorithm, to vehicle control.

1. Introduction

The smart regenerative braking system for electric vehicles can provide driving convenience for drivers and avoid unnecessary brake operations by automatically controlling regenerative braking in response to driving environments. Additionally, the driver can select the preferred level of braking force and timing in user options, which are categorized as soft, medium, and strong. However, some drivers could encounter inconveniences when the regenerative braking is activated differently from their driving habits because the strength and timing of regenerative braking are controlled based on calibration data predetermined during vehicle development.

Various efforts have been made to alleviate driving discomfort resulting from automated brake control, and recent initiatives have involved diverse research endeavours leveraging artificial intelligence [1,2,3,4]. Applying anomaly detection algorithms to various artificial intelligence algorithms enables the identifying and refining of abnormal driving scenarios, such as the unexpected merging of the leading vehicle, drowsy driving, and abrupt braking. Subsequently, the vehicle can be controlled to enhance driving stability with an intelligent driver assistance system trained on refined data acquired through an anomaly detection algorithm.

Among various studies aimed at refining abnormal driving data, one study utilizes DBSCAN-DTW (density-based spatial clustering of applications with noise-dynamic time warping) to cluster driving data, excluding data that fail to form clusters as outliers [5]. However, the DBSCAN algorithm forms clusters based on the density of the data, leading to a significant impact on cluster results based on the quantity of data. Therefore, hyperparameters such as eps and min_samples substantially influence the clustering outcomes, making it challenging to generalize the clustering algorithm. Thus, in this study, deep learning-based algorithms are developed to perform anomaly detection based on the features of the data without being influenced by hyperparameters.

Subsequently, the anomaly detection algorithm is utilized to refine data from abnormal driving scenarios and driver characteristic data are applied to vehicle control systems. Furthermore, the regenerative braking system incorporating personalized data is assessed to determine whether it controls braking intensity in a manner similar to that of actual driver behaviour, thereby improving driving consistency and reducing discomfort.

2. Background

2.1. Smart Regenerative Braking System

The smart regenerative braking system is a driver convenience feature that assists in braking the vehicle based on driving conditions, road incline, and the driver’s deceleration tendency, even without the driver pressing the brake pedal. The fundamental control of the smart regenerative braking system blends the Constant Time Gap Policy (1) and the Uniformly Accelerated Motion Equation (2) according to the distance and relative speed of the vehicle ahead. The Constant Time-Gap Policy [6,7,8,9,10], aimed at improving string stability, offers the advantage of preventing congestion in trailing vehicles compared to the use of constant spacing. Additionally, the Uniformly Accelerated Motion Equation is a formula that calculates target deceleration based on the equation of uniformly accelerated motion.

$$a_{CTG}=-{\displaystyle \frac{1}{h}}\left(\dot{\epsilon }+\lambda \delta \right)$$

(1)

$$a_{UAM}={\displaystyle \frac{{{\nu }_{front}}^2-{{\nu }_{ego}}^2}{2\epsilon }}$$

(2)

Here, ε represents the distance to the vehicle ahead, λ is the control parameter, h is the target time gap, δ is the distance error, ν_front is the speed of the vehicle ahead, and ν_ego is the speed of the ego vehicle.

In addition, by adjusting the intervention timing and braking intensity of smart regenerative braking, it is possible to improve the driving discomfort experienced by the driver due to automatic control. However, since the intensity of regenerative braking is determined based on predefined data during the vehicle development process, some users may feel uneasy while driving due to the regenerative braking activating too late or too strongly. This study proposes a smart regenerative braking system that learns the driver’s driving data and controls the vehicle in a manner identical to the driver’s usual habits, providing a sense of driving stability.

2.2. Dataset Description

The real driving data used to measure the model’s utility are summarized in

Table 1. Dataset statistics.

Vehicle Types	Drivers	Dataset Size	Anomaly Ratio
IONIQ5	Driver 1	55	0.67
IONIQ5	Driver 2	35	0.60
G80EV	Driver 2	29	0.63
G80EV	Driver 3	23	0.56

. These data were collected during actual drives from Yongsan in Seoul to Gunpo in Gyeonggi Province in South Korea. Among these, data instances where the driver’s foot was off the accelerator pedal were considered to indicate a lack of intention to accelerate and a desire to reduce speed; thus, only these instances were acquired through preprocessing. Each dataset consists of 11 features, and the labelling was manually conducted by an expert to classify solely the stable deceleration data. During this process, states showing a similar pattern of the driver pressing the brake pedal where the longitudinal deceleration values increased in the negative direction and then converged to zero were defined as the driver’s braking habits and labelled accordingly. Additionally, unlabeled driving datasets were also used to qualitatively assess the anomaly detection model’s potential for generalization.

3. Unsupervised Anomaly Detection

3.1. Motivations

In time series anomaly detection, benchmark datasets typically feature fewer anomalies than normal data, challenging the effectiveness of traditional supervised learning algorithms and leading to a focus on unsupervised learning for anomaly detection. These algorithms, using normal data, train models for tasks such as prediction or reconstruction, enabling them to recognize normal data patterns. However, when these models encounter unseen anomalous data, they exhibit significant performance drops, indicating anomalies.

For developing smart regenerative braking systems, accurately extracting drivers’ deceleration patterns is key. Our study assumes that patterns where a leading vehicle’s deceleration is followed by the trailing vehicle are the most representative. While general clustering algorithms can identify these patterns without extensive labelling, their hyperparameter sensitivity complicates real-world application.

To address this issue, we advocate for an anomaly detection approach. By defining data that reflect a driver’s deceleration style as ’normal’, we train anomaly detection models to identify and exclude ’anomalous’ profiles during deployment. This method streamlines data filtering and reduces hyperparameter tuning needs, though it requires some labelled data and advanced generalization techniques for diverse drivers and vehicles.

3.2. Anomaly Detection Algorithms

We explore the use of EncDec-AD [11] and Transformer [12] models in identifying unique deceleration patterns. Our evaluation extends to the models’ generalizability across various driving conditions, vehicle types, and driver behaviours, highlighting their potential for real-world applications.

3.2.1. EncDec-AD

In Figure 1,The EncDec-AD model is a Long Short Term Memory (LSTM) network-based Encoder–Decoder scheme for anomaly detection that is designed to learn to reconstruct ’normal’ time series behaviour and uses reconstruction errors to detect anomalies. This approach is robust and can detect anomalies from predictable, unpredictable, periodic, aperiodic, and quasi-periodic time series. It is particularly useful in scenarios where anomalous data are not available or are sparse, making it difficult to learn a classification model over the normal and anomalous sequences. This is especially true for machines that undergo periodic maintenance and therefore are serviced before anomalies show up in the sensor readings.

The model works by first training the LSTM Encoder–Decoder to reconstruct the normal time series. The LSTM encoder learns a fixed-length vector representation of the input time series, called the feature vector, and the LSTM decoder uses this representation to reconstruct the time series in reverse order. Here, a special technique called teacher forcing is applied which uses the original data as the input to the decoder with a certain probability (p) during training. The reconstruction errors are then used to obtain the likelihood of a point in a test time series being anomalous. A higher anomaly score indicates a higher likelihood of the point being anomalous.

The EncDec-AD model is shown to be effective in detecting anomalies in various scenarios, including different datasets with characteristics such as periodicity, predictability, and dimensionality. It has been tested on datasets from diverse domains, demonstrating its ability to detect anomalies from short time series (length as small as 30) as well as long time series (length as large as 500).

3.2.2. Transformer

In Figure 2, The Transformer model offers an alternative approach to sequence modelling by replacing traditional recurrent and convolutional neural networks with an architecture based solely on attention mechanisms. This design enables the model to directly compute relationships between all timestamps in the input and output sequences, irrespective of their positional distances. The architecture comprises an encoder and a decoder, each consisting of six identical layers that incorporate multi-head attention mechanisms and position-wise fully connected feed-forward networks. This configuration allows the model to dynamically allocate attention, effectively capturing long-range dependencies within sequences.

In this work, we utilize the Transformer model as a time series predictor within a prediction-based unsupervised anomaly detection framework. Given the past w timestamps (where w represents the window size), the Transformer predicts values for the subsequent h timestamps (where h denotes the horizon length). Unlike LSTM models the Transformer processes the entire sequence in parallel, which can cause the positional information of individual values within the sequence to be lost. To address this, sinusoidal positional encoding is added to the input data, enabling the model to account for the order of values in the sequence. The transformed input is then passed into the Transformer encoder, starting with the multi-head attention layer. This layer employs a “Scaled Dot-Product Attention” mechanism, which calculates the output as weighted sums of the inputs, with weights being determined based on the compatibility between queries and keys. This mechanism allows the model to flexibly adjust its focus across the sequence, facilitating parallel data processing compared to LSTM, thereby reducing training time and enhancing the efficiency of sequence modelling tasks. The multi-head attention layer performs scaled dot-product attention computations in parallel for a number of heads specified as a hyperparameter, collecting information from diverse perspectives. Subsequently, the outputs pass through a feed-forward network, and the encoder generates the context vector, which serves as the representation of the input sequence over the w timestamps.

During training, the Transformer decoder employs the teacher forcing technique, similar to the EncDec-AD architecture, to facilitate learning. Specifically, the target output sequence is used as part of the decoder input, starting with a special start token. The masked multi-head attention layer within the Transformer decoder restricts the scaled dot-product attention operation to only consider positions preceding the current one. This masking ensures that the model generates predictions in an autoregressive manner, preventing any information leakage from future timestamps. Following this, the multi-head attention layer in the decoder does not use self-attention but instead uses the encoder’s context vector as keys and values. This operation enables the decoder to leverage the encoded representation of the input sequence while predicting the output sequence. Finally, the decoder’s output passes through a projection layer to generate predictions for the subsequent h timestamps. Through this process, the Transformer decoder predicts the probability distribution of the next token in the sequence based on the entire input sequence.

In the inference phase, since teacher forcing is not applied, the decoder receives only the start token, which is initialized as a zero vector. Using an iterative decoding process, the decoder is repeatedly invoked, with each predicted output token being fed back as input to generate the output sequence incrementally. Ultimately, the predicted output sequence is compared with the ground truth to compute the prediction error, which is subsequently refined into an anomaly score. This score quantifies the level of anomaly for the given timestamps.

3.3. Experimental Results

We conducted various experiments and visualizations to compare and analyze whether the EncDec-AD or Transformer model more effectively detects abnormal deceleration driving profiles and generalizes to unseen vehicles and drivers. For quantitative anomaly detection evaluation, we utilized precision, recall, and the F1-score metric, which is one of the most commonly used evaluation metrics for anomaly detection tasks in machine learning. Unlike accuracy, which can be misleading due to the inherent class imbalance nature of anomaly detection problems, where normal samples significantly outnumber anomalies, the F1-score provides a more reliable evaluation by harmonically averaging precision and recall. We placed an additional emphasis on recall due to the task’s characteristics by also employing an fβ-score with β = 2. The formula for the fβ-score is as shown in Equation (3). The threshold for the anomaly score was set to the value that yields the highest f1 score in the test data.

$$F_{\beta }=\left(1+{\beta }^2\right)·{\displaystyle \frac{precision ·recall}{\left({\beta }^2·precision\right)+recall}}$$

(3)

The results of all experiments are based on the values obtained using the feature set combination that shows the highest detection performance out of a total of 11 features. For EncDec-AD, the dimensions of the LSTM hidden layers and the number of layers were chosen from {5,10} and {1,2,3}, respectively. For the Transformer, the dimensions of the hidden layers, the number of heads, and the number of layers were chosen from {4,6}, {2,3}, and {1,2}, respectively. The final hyperparameter configurations were determined through a grid search.

The results of the quantitative comparison between the two models for each vehicle type, IONIQ5 and G80EV, are documented in

Table 2. Quantitative results for EncDec-AD and Transformer Dataset statistics.

Dataset	IONIQ5		G80EV
Model	EncDec-AD	Transformer	EncDec-AD	Transformer
Precision	0.9686	0.9336	1.0000	0.9767
Recall	0.7708	0.8792	1.0000	1.0000
F1-score	0.8585	0.9056	1.0000	0.9882
Fβ-score	0.8037	0.8896	1.0000	0.9952

. Across all experiments, and regardless of the model or vehicle type, it was observed that anomaly detection techniques for extracting abnormal deceleration driving patterns performed significantly well. This demonstrates that learning through deep learning reduces dependency on hyperparameters and can yield stable, high performance, serving as evidence that the approach shifting the problem’s paradigm to anomaly detection is justified. As seen in Table 2, for IONIQ5 data, the Transformer showed quantitatively better performance. However, for G80EV data, the limited amount of labelled data presents challenges in comparing the performances of the two models. Therefore, a qualitative analysis through a substantial amount of unlabelled data is necessary to select the model most suitable for actual deployment.

3.4. Generalizability Analysis

From an anomaly detection viewpoint, the method for extracting abnormal deceleration patterns, unsuitable for training smart regenerative braking systems, demonstrates a lower dependence on hyperparameters and superior performance over traditional clustering-based methods, enhancing its practicality. However, a notable limitation is the requirement of a specific amount of labelled data to facilitate effective training. Despite being an unsupervised learning technique which typically does not necessitate labelled data for training it still requires a dataset of high-purity normal estimates. Moreover, labelled data are essential for periodic evaluation of the model’s performance. Given these challenges and the real-world difficulty in obtaining high-quality labelled data, the success of deploying anomaly detection-based filtering largely hinges on the model’s ability to generalize. Should a model trained on a limited selection of vehicle types and drivers demonstrate robust generalization across a broad spectrum of vehicle types and drivers, it would affirm the practical viability of the anomaly detection-based method.

3.4.1. Experiment Settings

The experimental setup for evaluating the generalization performance of the anomaly detection model across different vehicle types and drivers is constructed by combining the datasets in Table 1, as shown in

Table 3. Qualitative evaluation in generalizability experimentation settings.

Scenario	Train Dataset	Test Dataset
Scenario 1	IONIQ5 + G80EV normal dataset	IONIQ5—Driver 1
		IONIQ5—Driver 2
Scenario 2	IONIQ5 + G80EV normal dataset	G80EV—Driver 2
		G80EV—Driver 3
Scenario 3	IONIQ5 + G80EV normal dataset	IONIQ5—Driver 2
		G80EV—Driver 2

. Scenario 1 and 2 involves using one anomaly detection model per vehicle type to assess whether a pre-trained model can generalize across all drivers. In contrast, Scenario 3 examines whether a single pre-trained universal anomaly detection model can uniformly generalize across vehicle types.

Figure 3 illustrates the deceleration profile during a vehicle braking scenario. The braking process is divided into four stages: Coasting, Deceleration, Adjustment, and Termination. We analyzed the braking initiation point (ⓐ deceleration start point) and braking intensity (ⓑ deceleration rate) during the deceleration phase, where the driver’s braking tendencies are most pronounced with scatterplots. Additionally, a quantitative evaluation was conducted to assess whether the data refined by the anomaly detection algorithm could effectively classify representative data reflecting the driver’s braking tendencies. The results and analysis for each scenario are discussed in the following Section 3.4.2.

3.4.2. Analysis for Scenario 1 and Scenario 2

For qualitative evaluation, we analyzed scatterplots representing gradient deceleration and vehicle speed according to distance for both normal and abnormal data. It is common for normal data to exhibit an upward trend in both scatterplots, so we examined whether the yellow dots used in training and the green dots identified as normal during inference displayed a distinct upward trend. Furthermore, a well-generalized anomaly detection model is expected to have a clearer decision boundary and identify points deviating from the pattern as outliers, marked with a red ’x’. All qualitative evaluations were conducted with the consultation of domain experts.

Figure 4 and Figure 5 illustrate the results of Scenario 1 in Table 3, presenting the outcomes of outlier removal applied to driving data collected from Driver 1 and Driver 2 using the IONIQ5. Both results indicate, consistent with pre-trained data, that braking intensity increases as the distance to the preceding vehicle decreases (with an increase in the gradient of deceleration in the negative direction). Furthermore, when employing the Transformer model, the boundaries between normal and abnormal data are more distinctly separated, suggesting that the Transformer model might have detected anomalies more conservatively to cover a wider range of normality. The IONIQ5 data showed significant differences in data distribution across drivers. At the same vehicle speed, Driver 1 demonstrated a characteristic tendency to commence braking at shorter distances from the preceding vehicle. Thus, it can be observed that the processed data inherently reflect the distinct driving characteristics of individual drivers.

Figure 6 and Figure 7 present the results of Scenario 2 in Table 3, applying the same anomaly detection algorithm to the data collected from Driver 2 and Driver 3 using the G80EV. An examination of the driving data distribution for Driver 3 reveals a clearer distinction between anomalous and normal data. The application of the anomaly detection algorithm to this dataset exhibits a pronounced effect, effectively categorizing most data points deviating from the pre-trained dataset as anomalous.

3.4.3. Analysis for Scenario 3

Figure 5 and Figure 6 show the results of the same driver operating two vehicles, the G80EV and IONIQ5, providing a basis for a comparative analysis to evaluate Scenario 3, as outlined in Table 3. The results of Scenario 3 indicate that the models demonstrated a certain degree of generalization success, consistent with the findings from Scenarios 1 and 2. It was confirmed that anomalous driving data falling outside the domain of the pre-trained dataset could be consistently classified, regardless of changes in the vehicle. However, due to the change in the vehicle being driven, the tendencies in the driver’s characteristic data appear differently. Ultimately, the evaluation of all scenarios makes it clear that generalized results can be achieved using the same anomaly detection model, regardless of the vehicle or driver.

4. Vehicle Testing and Assessment

4.1. Vehicle Testing Environments

In Section 4, the refined data generated through the anomaly detection model were applied to actual vehicle control to examine whether the driving discomfort experienced during the operation of the smart regenerative braking system could be alleviated.

As shown in Figure 4, the data from Driver 1 operating the IONIQ5 were refined using an anomaly detection model. The refined data express the relationship between the distance to the preceding vehicle and braking intensity as a second-order equation using polynomial regression. Additionally, the vehicle speed and inter-vehicle distance at the moment of braking were represented as a first-order equation. Based on these equations, software was implemented to perform smart regenerative braking control, allowing for the evaluation of whether personalized data could reduce driving discomfort. The evaluation was conducted by comparing the actual driving data of the driver with the personalized smart regenerative braking driving data.

4.2. Assessment of Testing Results

Figure 8 presents the results of applying personalized data to smart regenerative braking control and comparing it with actual driving data. In the section where the relative distance to the preceding vehicle is approximately 20 m and the driving speed is about 40 kph, the comparison shows braking intensities of −5.74 m/s³ and −4.78 m/s³, respectively, indicating similar levels. This demonstrates that the smart regenerative braking control with the personalized data operates with braking forces comparable to those observed during actual driving. These results confirm the validity of the personalized data refined through the anomaly detection algorithm.

5. Conclusions

Our research investigates an anomaly detection method for selectively extracting the deceleration patterns that best represent a driver’s driving characteristics from raw driving data. To account for the complex nature of multivariate data, deep learning-based anomaly detection methods such as EncDec-AD and Transformers were utilized, demonstrating their potential for generalization. A quantitative and qualitative comparison of the two approaches revealed that the Transformer model, by detecting anomalies more conservatively to encompass a broader range of normalcy, is more suitable for this study. Furthermore, data refined through the anomaly detection model was applied to a smart regenerative braking system, and real-world vehicle tests confirmed the capability of achieving braking performances similar to that of actual drivers. These advancements serve as a critical technology for flexible and driver-friendly smart regenerative systems, reducing driving discomfort and enhancing driving safety.

Future work will focus on extending the proposed anomaly detection approach to account for dynamic driving environments, integrating additional sensor data for enhanced context awareness and exploring personalized adaptation mechanisms to further optimize braking performance based on individual driving behaviours. Additionally, large-scale deployment and long-term testing will be conducted to evaluate the robustness and reliability of the system under diverse driving scenarios.