Abstract: Injection molding is among the most important and widely adopted manufacturing technologies in the polymer processing industry, in which the speed of injection plays a decisive role in the quality of plastic products. However, in the actual molding process, the injection molding machinery may be subjected to external disturbance, such that the desired injection speed is not efficiently maintained, ultimately affecting product quality and increasing production waste. It is thus essential to design a control scheme that contributes to mitigating the effects of external non-repetitive disturbance on system performance. In this study, we developed a model-free adaptive iterative learning control (MFAILC) scheme that is combined with a fading factor. We initially describe the injection molding process in detail, which covers mold closing, melt injection, pressure holding and plasticization stages. After that, we established a non-linear state-space model for the injection section of the molding process. Given that this model has strong non-linearity and uncertainty, it is difficult to design a controller using traditional strategies. Thus, to address this problem, we developed a compact-form dynamic linearization (CFDL) model with iterative characteristics, which contains a pseudo-partial derivative (PPD) that incorporates all of the non-linearities of the injection system and reflects the relationship between the variations in system input and output. Having established this model, we then designed a speed control strategy and a PPD learning law based on the iterative axis by minimizing an optimal quadratic index function that balances tracking accuracy and the smoothness of control input. To prevent the PPD from learning divergence, a resetting algorithm of the PPD is then introduced. These three elements thus form an integrated MFAILC scheme. To mitigate the adverse effects of external measurement disturbance, a fading factor was subsequently integrated into the proposed MFAILC scheme, which adaptively reduces the weight of historical disturbance-related errors in subsequent iterations. Rigorous convergence analysis is conducted in the sense of expectation, proving that both the mean value and variance of the output tracking error of the MFAILC algorithm converge to zero as the number of iterations increases, thereby verifying the robustness of the algorithm from the perspectives of both deterministic convergence and statistical characteristics. Finally, using actual injection molding process parameters, we conducted a series of simulation experiments based on the MATLAB platform. The results obtained revealed that the proposed MFAILC strategy with a fading factor achieves full convergence of the injection speed trajectory with the desired curve after fifty times iteration, with a maximum tracking error of only 0.00039. Comparative experiments, in which we used MFAILC without a fading factor (maximum error of 0.0008) and P-type ILC with a fading factor, indicated that the proposed method outperforms the contrast schemes with respect to both convergence speed and disturbance suppression capacity. Collectively, the procedures developed in this study provide a reliable data-driven control solution for the regulation of high-precision injection speed and will facilitate the deployment of the MFAILC strategy in practical industrial procedures.