Aiming at the low efficiency of defect analysis for intelligent cockpit applications, this paper proposes CockpitInsight, a four-stage automatic analysis framework based on Large Language Models (LLM). The framework comprises multimodal defect understanding, adaptive log filtering, context augmentation, and Chain-of-Thought root cause reasoning. The adaptive time-window filtering algorithm employs a core, extended and background three-zone layered strategy to compress 170 000 raw log lines to approximately 700 lines (99.6% compression ratio), resolving the contradiction between massive logs and LLM context window limitations. The LLM-summary-enhanced semantic retrieval method generates structured summaries for historical defects to bridge expression gaps, achieving high-quality cross-project semantic matching on an index of 1 054 defects. Validation data shows an end-to-end analysis time of 70 s (including video analysis), with generated reports containing complete reasoning chains and log evidence references, demonstrating the engineering feasibility of LLM-driven multimodal defect analysis in the intelligent cockpit domain.

To address the limitation that conventional voice testing methods fail to effectively evaluate the response capability of intelligent cockpit voice interaction systems, this paper presents an intelligent voice testing system driven by a Large Language Model (LLM). The system first employs the LLMs to perform semantic generalization on base voice commands and, in combination with a multimodal large model, synthesizes test audio that covers multiple timbres and speaking rates, thereby overcoming the constraints inherent in conventional fixed-form test case design. A parameterized mapping model between test cases and execution programs is then constructed, which decouples the test logic from the execution environment and substantially reduces the dependence of execution-program development on hardware resources. In the result verification stage, data from three independent dimensions (including audio feedback recognition, visual parsing of the head-unit display, and controller log analysis) are fused through a weighted decision-fusion model, thereby enhancing the objectivity and reliability of the verdicts. Tests are conducted on the cockpit voice-based music control function for validation, and the results demonstrate that the proposed system performs excellently in terms of test pass rate, decision accuracy, semantic consistency, and generalization diversity, confirming the effectiveness of this method in improving test coverage, efficiency, and confidence.

To address the issues of fragmented entry points, lengthy operational paths, and inadequate adaptability of static interfaces for complex functions across diverse driving scenarios in intelligent cockpits, this paper proposes a resource allocation design method for automotive Human-Machine Interfaces (HMI) oriented toward complex functional task contexts. A closed-loop framework comprising “task context representation, user behavior modeling, functional relevance calculation, strategy scoring, and interface resource adjustment” is established. Firstly, a representation of complex functional task contexts is constructed by integrating multidimensional data, including vehicle status, task procedures, interaction history, spatiotemporal context, and user identity. Secondly, a user behavior model is developed by extracting long-term preferences from anonymized infotainment system telemetry data and incorporating short-term behavioral trends. Subsequently, the priority of candidate functions is calculated by synthesizing task-context relevance and user-behavior relevance. Finally, a collaborative adjustment scheme for the structural and content layers is generated, incorporating knowledge base constraints, strategy feasibility, and disruption cost considerations. The proposed method is validated through a controlled experiment on an infotainment system testbench, using seat comfort and spatial adjustment functions as the test case. The results demonstrate that, compared with conventional static designs, the proposed method significantly reduces function search time and task completion time, decreases the number of operational steps and erroneous inputs, and yields superior performance in terms of design efficiency, solution rationality, and user satisfaction. Consequently, it effectively enhances interaction efficiency and user experience in driving scenarios.

An adaptive regenerative braking control method based on multi-source information fusion is proposed to address the fixed recovery intensity of conventional electric vehicles, insufficient adaptability to complex conditions and degraded deceleration consistency when battery recovery capability is limited. Navigation speed limits, preceding and multiple dynamic targets, road grade, road-surface state and available battery recovery power are integrated in a perception-decision-execution architecture. Signal validity judgment, dynamic regenerative torque calculation and electro-hydraulic compensation are then designed. The hydraulic compensation is activated only when the available battery recovery power cannot meet the deceleration demand, thereby improving redundancy and deceleration consistency. Road tests on the Voyah Dreamer vehicle model show that the proposed method improves energy consumption by 4.02%, reduces braking frequency by 44.64%, and limits deceleration fluctuation within ±0.02 g under multi-speed-limit and mildly congested conditions. The method improves energy economy and drivability while maintaining vehicle safety.

Aiming at the vehicle energy management problem caused by multiple power sources of hybrid electric vehicles, as well as the differentiated requirements of different users for power source utilization, 4 energy modes, namely forced pure electric, pure electric priority, intelligent hybrid and charge-sustaining priority, are developed. Detailed control strategies for the 4 modes are formulated. The differences among the modes in terms of vehicle pure electric driving range, charge-depleting fuel consumption and acceleration performance are investigated. Meanwhile, the actual application of each energy mode is statistically analyzed based on big data of sold vehicles. The results show that the 4 energy modes differ in pure electric driving range, while the fuel consumption in the charge-sustaining stage remains consistent. Except for the forced pure electric mode, the acceleration performance of different energy modes is identical, and the pure electric priority mode accounts for the highest proportion in actual application.

To accelerate the commercial application of solid-state batteries, the paper investigates the thermal runaway characteristics of pouch-type semi-solid battery cells and small modules. The study reveals that compared to conventional liquid electrolyte batteries, semi-solid batteries exhibit more violent reactions during thermal runaway, characterized by intense combustion accompanied by a large amount of high-temperature and high-pressure gas and internal active material ejection, with maximum temperatures exceeding 1 300 ℃. Based on these characteristics, the paper proposes a battery module solution integrating “directional venting” and “multi-level thermal insulation” safety design combined with an efficient cooling system. This solution effectively suppresses the propagation of thermal runaway after cell failure, and the directional venting design at the module level significantly reduces the complexity of thermal safety design for battery systems and enhances the thermal safety reliability of battery packs.

Against the backdrop of rapid iteration in intelligent electric vehicles, design language has emerged as a core carrier for brand identification and user emotional interaction. However, the dynamic coupling between design language and users' perceptual cognition lacks large-scale data support and objective quantitative tools. Taking BYD and Tesla as typical Eastern-Western comparative cases, this study constructs a 2020—2026 multimodal dataset consisting of 125 standardized vehicle images and 6 596 valid user comments, generating over 140 000 image-text pairs. With image-text features extracted via Chinese-CLIP, combined with Kansei engineering factor analysis and Analytic Hierarchy Process (AHP), this paper quantitatively analyzes the temporal evolution of design language and the trend of perceptual imagery matching. The results reveal obvious brand divergence in users' perceptual vocabulary: BYD is dominated by “stability” and “spaciousness”, while Tesla centers on “technological sense” and “futuristic sense”. The image-text matching degree of both brands declines slowly, with a sharper drop for Tesla, indicating diversified user aesthetics amid market maturity. At the factor level, BYD shows differentiated factor trends: its perception of technological futurism continuously weakens, whereas the imagery of luxury and refinement rebounds after bottoming out. Tesla's “technological futurism” factor remains high yet fluctuates, uncovering distinct evolutionary logics between Eastern and Western automotive brands.

To improve the safety standard system for automotive artificial intelligence in China and address the issue of insufficient compliance adaptability in the industry, the paper systematically reviews the released and under-development international standards for automotive artificial intelligence based on the automotive artificial intelligence safety framework of ISO/PAS 8800, and compares the adoption and independent development of relevant standards in China, as well as existing gaps. It also summarizes the evolution of global policies and regulations governing automotive artificial intelligence. China's existing standards for automotive artificial intelligence safety remain to be further improved in system construction, forward-looking deployment and connection with relevant laws and regulations. Combined with the top-level plans of automotive aritificial intelligence standardization advancement center and the SAFER AI initiative, the paper proposes a construction pathway for China's automotive artificial intelligence safety standard system, centered on risk classification, full-lifecycle management, vehicle-grade adaptation, international alignment, and regulatory coordination. This study provides theoretical references and practical guidance for the improvement of China's automotive artificial intelligence safety standard system, industrial compliance and international standard coordination.

To address the adaptability bottlenecks of traditional automotive-grade chip mission profile frameworks in the context of high-voltage electrification and the large-scale deployment of wide-bandgap devices in new energy vehicles, this paper conducts a systematic review of the theoretical framework, standardization paradigms, load parameter discrepancies, reliability assessment models, and engineering application boundaries of automotive chip mission profiles, based on AEC-Q100 Rev-J. The fundamental differences in chip loading characteristics between conventional fuel vehicles and new energy vehicles are comparatively analyzed across multiple dimensions. The study reveals that the thermal load duration of chips in new energy vehicles is approximately 10 times longer than that in fuel vehicles, with operating voltages reaching 800 V and power device switching frequencies increasing by approximately 20 times. Traditional automotive chip reliability assessments are inadequate for new energy vehicles, failing to address multi-physics coupling, long-term aging, and the unique failure mechanisms of wide-bandgap devices. Currently, the reliability certification of automotive-grade chips is shifting from "general compliance demonstration" to "application suitability demonstration". In the future, constructing refined, dynamic, and multi-scale mission profiles, along with improving the suitability assessment system for wide-bandgap devices, will become critical directions requiring breakthrough in this field.

To clarify the current state of the automotive AI chip industry, address technological pain points, and identify the future development trends of the industry, while supporting the domestic automotive AI chip industry's independent and high-quality development, this focuses on the core application scenarios of AI chips in intelligent connected vehicles, such as smart driving perception, decision-making, control, and smart cockpit interaction. It systematically reviews the core requirements of intelligent connected vehicles for AI chips, including high computing power, low power consumption, high security, and high reliability. At the same time, it analyzes the technology layouts and product iterations of mainstream domestic and international automotive AI chip companies like Qualcomm, NVIDIA, Tesla, Horizon, and SiEngine Technology, and, based on the current industry situation, summarizes the main technical bottlenecks and solutions for automotive AI chips in areas like memory bandwidth and packaging processes. On this basis, it assesses the future development trends of automotive AI chips in terms of domain fusion, heterogeneous architecture integration, computing power growth, the implementation of computing-storage integration technology, and advanced packaging. The goal is to provide a reference and inspiration for the technological research and development, product iteration, industrial layout and relevant research of vehicle-mounted AI chips.