Abstract: The study recruited 17 patients diagnosed with knee osteoarthritis (KOA) (including 7 with bilateral and 10 with unilateral knee arthritis) and 12 healthy volunteers. Using a self-developed three-channel wearable microphone array system, acoustic signals from the knee joints of the subjects were synchronously collected during a "sit-stand-sit" movement task, resulting in a total of 438 knee joint acoustic samples. Multidimensional features such as time domain, frequency domain, and Mel Frequency Cepstral Coefficients (MFCC) were extracted to construct a feature dataset. The study employed the logistic regression machine learning algorithm to build a classification model and evaluated its performance through Leave-One-Group-Out Cross Validation (LOGO CV). The focus was on comparing the classification performance of multi-channel data fusion with that of individual single-channel data (patella, medial and lateral tibial platforms). The experimental results showed that multi-channel data fusion improved classification accuracy. Among them, the combination of "patella + lateral tibial platform" performed better in terms of all indicators with low fluctuation, achieving an accuracy of 94.23±8.41% and an AUC-ROC of 99.94±0.27%. Among the single channels, the patella site achieved the highest average accuracy and average AUC, reaching 92.73±9.35% and 98.79±3.56%, respectively. This study provides a new, low-cost, and highly reliable model for community screening and rehabilitation assessment of KOA.