Currently, the objective evaluation of the DCT vehicle drivability requires the accurate identification of the driver’s intention and vehicle state as well as the selection of the targeted evaluation indicators. The existing identification methods usually cannot divide the driver’s intentions in detail and make full use of the characteristics of time-series signals. Simultaneously, external kinematic sensors are more commonly used than the sensors of vehicle powertrain, which impacts the recognition effect. This paper proposes a new method for identifying the DCT vehicle driver’s starting intentions based on an LSTM neural network and multi-sensor data fusion. The DCT vehicle driver’s starting intentions are subdivided and defined based on human–vehicle interaction analysis and K-means clustering. The input of the model consists of 11-dimensional variables that include motion parameters of the vehicle collected by the external sensors and the powertrain parameters collected by onboard sensors. The method proposed in this paper first establishes a recognition window, which is utilized to extract the starting process samples from the DCT vehicle driving data. Second, the 11 variables of each sample are used as one set of multi-dimensional time-series signals, which are preprocessed through wavelet denoising. Finally, the LSTM network is used to identify the samples. The identification results indicate that the highest recognition accuracy of the proposed algorithm is 94.27%, which is approximately 5% higher than conventional methods, such as fully connected neural networks and support vector machines. Furthermore, the model with 11 input variables outperforms the model with fewer input variables. The effectiveness and superiority of the identification model have been demonstrated.

目前，对DCT车辆驾驶性能的客观评价需要准确识别驾驶员的意图和车辆状态，并选择有针对性的评价指标。现有的识别方法通常不能对驾驶员的意图进行详细的划分，也不能充分利用时序信号的特点。同时，外部运动传感器比车辆动力总成传感器更常用，这影响了识别效果。本文提出了一种基于LSTM神经网络和多传感器数据融合的DCT车辆驾驶员起步意图识别新方法。基于人车交互分析和K-means聚类对DCT车辆驾驶员的出发意图进行细分和定义。模型的输入由 11 维变量组成，包括外部传感器收集的车辆运动参数和车载传感器收集的动力总成参数。本文提出的方法首先建立一个识别窗口，用于从DCT车辆行驶数据中提取启动过程样本。其次，将每个样本的11个变量作为一组多维时间序列信号，通过小波去噪进行预处理。最后，使用 LSTM 网络来识别样本。识别结果表明，该算法的最高识别准确率为94.27%，比全连接神经网络和支持向量机等传统方法提高了约5%。此外，具有 11 个输入变量的模型优于具有较少输入变量的模型。证明了识别模型的有效性和优越性。