The chip on PCBA is developing towards small size and high density, which make it much difficult to detect micro solder bump defects inside the package. To address the problems of difficulty and low efficiency in locating internal faults of ICs on industrial high-density integrated PCBA, a chip-on-board defect detection method combining the infrared thermal imaging and the deep learning algorithm is proposed, which realizes intelligent defect detection of ICs on PCBA suitable for industrial production scenarios. Taking the real DDR memory chip on FPGA as the target, the infrared defect detection model is formulated, and the test bench is established to conduct experimental research on the fault detection of solder bumps in the chip. The designed program realizes the chip data storage and readout. The infrared image sequence is collected to analyze the temperature evolution of different defect types in the process of DDR chip reading and writing. The thermal signals of different measurement areas are extracted for defects that are difficult to intuitively distinguish by infrared images. With the hyperparameter optimization, the CNN classification model realizes efficient and accurate detection of different defect types, including address, data, and bank address solder joint fault. Furthermore, after transfer learning, the other 9 different solder joint defects of the chip are accurately identified, and the accuracy is over 95% and over 92% under the conditions of 10 and 20 dB Gaussian white noise, respectively. It provides an efficient and effective method for microelectronics packaging and reliability analysis on industrial high-density integrated PCBA.