In view of the large communication volume required by traditional centralized methods in active distribution network fault section location, as well as the challenges faced by existing artificial intelligence methods such as complex models and resource constraints when deployed in a distributed manner, a two-stage collaborative intelligent fault section location method of distributed deployment and centralized diagnosis is proposed. This method first uses variational mode decomposition to extract high-dimensional local fault features from the current signals collected by each distributed measurement unit, and then completes the preliminary diagnosis at the terminal through a lightweight multi-layer perceptron model, only uploading the low-dimensional fault section probability vector. Subsequently, at the centralized fusion end, taking these probability vectors as meta-features, meta-learning data is generated with the aid of K-Fold cross-validation, and a collaborative neural network based on Stacking is constructed to optimize and fuse multi-source information end-to-end, in order to output the final fault section discrimination result. The simulation experiments show that the proposed method can effectively integrate distributed observation information and achieve high-precision fault section location under various fault conditions. Its performance is significantly better than that of a single local model and a simple fusion strategy, providing a new and effective way for the rapid and accurate location of distribution network faults.