Abstract: Industrial end-product qualities, eg, the composition fraction and molecular weight etc, are usually measured by using corresponding analyzers with considerable delay.The analyzer system, moreover, is expensive, unreliable and difficult to maintainAn adaptive Kernel leaning (AKL) network was proposed to build the soft sensor model for industrial analyzer and meanwhile to monitor its potential faults.The network utilized Kernel function and geometric angle to build an adaptive network topologyTwo forms of learning strategies for the AKL network were obtained and their corresponding recursive algorithms are developed, respectivelyNumerical simulations for analyzer of the Tennessee Eastman (TE) process showed that the soft composition analyzer developed by using the proposed AKL networks could achieve satisfying estimation precision under both normal and fault-existing operating conditions.