A variety of machine learning methods have good performance in fault diagnosis of induction motors under trained electric load. However, existing methods have low accuracy when diagnosing faults under untrained electric loads. In fact, it is impossible to train a system by using infinite number of electric loads. To solve this problem, a novel fault diagnosis framework including a training framework and an adaptation framework is proposed in this article. The system only needs to be trained by the training framework and rated load data, then, it can diagnose any other untrained loads by the adaptation framework. In the training framework, most training methods cannot automatically change their network structures to achieve a global maximum accuracy without overfitting. To address this issue, a broad learning (BL) with a particle swarm optimization is proposed. In the adaptation framework, most features (statistical feature and sample entropy) from untrained load are different from trained load. This degrades the diagnostic accuracy. To overcome this problem, an adaptive factor for statistical feature is, therefore, proposed to process the winding current data from untrained loads to be close to the data of trained load. At the same time, adaptive coefficient is proposed to adjust the sample entropy (SampEn) obtained from acoustic signal to ensure that the values of SampEn between untrained loads and trained load are similar. Even though features from untrained loads can be adjusted by the adaptation framework, the activation functions of BL trained by rated load are still different from those for untrained loads. To solve this issue, an improved scale exponential linear unit-broad learning with scale coefficient is, therefore, proposed to adapt the differences of the activation functions between the trained and untrained loads for enhancing the classification accuracy. Experimental results show that the proposed fault diagnostic framework is accurate under untrained loads.