Internal leakage from a heat-transfer tube in a shell-and-tube heat exchanger generates acoustic emission (AE) signals that differ from those under normal operating conditions. During propagation to externally mounted sensors, the signals are affected by coupling interfaces, propagation distance, thermal insulation, and ambient temperature. Mounting a sensor directly on the high-temperature shell places it at risk of thermal damage, whereas mounting it on the outer surface of the insulation weakens mechanical coupling. The sensor configuration must therefore balance acoustic transmission performance, thermal isolation, and installation constraints. Experiments were conducted on a shell-and-tube heat exchanger at a test pressure of 0.2 MPa under a normal no-leak condition and leakage conditions produced by 1, 5, and 10 mm long cracks in a single heat-transfer tube. AE signals were acquired using four configurations: an acoustic waveguide extending through the shell to contact the leaking tube; an acoustic waveguide extending through a 10 cm thick glass-fiber insulation layer to contact the shell; a sensor mounted directly on the shell; and a sensor mounted on the outer surface of the insulation. A one-dimensional convolutional neural network (1D-CNN), a long short-term memory (LSTM) network, and a convolutional neural network (CNN) were used for two-stage leak detection and severity classification. Performance was evaluated using acquisition-level decisions and window-level stability. For the waveguide-to-tube configuration, all three models achieved 100% leak-detection accuracy in four validation acquisitions; the severity-classification accuracies of the 1D-CNN, LSTM, and CNN were 100%, 75%, and 100%, respectively. For the waveguide-to-shell configuration, eight validation acquisitions covered four conditions and measurement locations near and far from the leak. All three models achieved 100% accuracy in both leak detection and severity classification, the lowest mean window-level accuracy was 92.50%, and all six model decisions for the near and far measurement locations under the normal no-leak condition were correct. With the sensor mounted directly on the shell, the LSTM network returned indeterminate results for both measurement locations under the normal no-leak condition. In the single acquisition for the 10 mm crack with the sensor mounted on the insulation surface, none of the three models detected the leak. Considering detection performance, cross-model consistency, thermal isolation, and installation constraints, the waveguide-to-shell configuration requires neither access to the equipment interior nor modification of the pressure boundary. It positions the sensor away from the high-temperature shell while maintaining a continuous metallic acoustic transmission path and is suitable for online monitoring of internal leakage in insulated heat exchangers.

acoustic emission,acoustic waveguide,internal leakage diagnosis,insulated heat exchanger,neural network