<p>This paper studies the multi-scale feature extraction technology in the field of angular contact ball bearing fault detection. Traditional fault detection methods are difficult to effectively reveal early fault signs when dealing with non-stationary signals or transient impact characteristics. To solve this problem, this paper presents an engineering integration framework for multi-scale fault diagnosis, with key novelties lying in the multi-domain fusion of order analysis, time-frequency domain processing technology and dynamic threshold determination, combined with transmission performance correlation. To verify the effectiveness of this framework, this paper conducts bench experiments covering various test scenarios such as multi-speed operation and gradual acceleration. The experimental results show that the organic combination of order analysis and time-frequency analysis can efficiently extract the fault features of bearings. The recognition success rate of the dynamic threshold criterion is as high as over 90%, while the misjudgment rate is controlled below 3%. By comparing the bench test data with the NVH (noise, vibration and acoustic roughness) performance of the entire vehicle and introducing the correlation index <i>R</i>² for quantitative evaluation, the high applicability and reliability of the proposed framework under complex vehicle working conditions were verified.</p>