Abstract
The fault diagnosis of heat pump systems is of great significance to reduce system energy consumption and ensure stable operation. Intelligent algorithms based on deep learning have a good diagnostic effect on complex thermal systems such as heat pump systems, but they also face the application problem of poor model generalization, which makes them unable to adapt to complex and actual dynamic scenarios. This paper focuses on the specific problems of the heat pump system's varying severity of fault conditions that cause the actual diagnosis accuracy of the model to decrease. Then, a new self-adaptive diagnosis method was proposed based on a residual data training framework combined with data scaling strategy, which adapts varying severity diagnosis under the condition that the training data derives from a single severity level. This paper analyzes the thermodynamic validity of the residual data training framework and, correspondingly, proposes a benchmark model based on the attention mechanism to greatly improve the quality of the residual data. Analyzing the thermodynamic laws of different fault severity and proposing a local scaling strategy for residual data allows for the residual sample instance of a single severity to be transferred to the residual sample space of multiple severity. Finally, verification of the ASHRAE RP-1043 data set revealed that the fault diagnosis method based on residual data can significantly improve the diagnosis accuracy and generalization. Furthermore, the higher accuracy of the benchmark model results in more obvious performance improvement. At the same time, the local data scaling strategy can significantly improve the upward and downward compatibility of the diagnostic model. It proves that the method proposed in this paper can effectively reduce the sensitivity of the diagnosis model to the severity of the fault, and it must be known that this self-adaptive method can also solve the problems of other actual diagnosis scenarios, such as different systems, varying environments, and varying loads.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have