Since the hysteresis and strong nonlinearity exist in the pneumatic system, it is difficult to achieve effective control of the pneumatic gripping force based on the air pressure signal directly, and using modeling to estimate the gripping force is an effective way to achieve low-cost control without force sensors. A convolutional neural network(CNN)-based optimized long short-term memory neural network (LSTM) was proposed for the low-cost pneumatic gripping force estimation method. According to the characteristic that the gripping force of the end pneumatic gripper of the industrial robot is related to historical inputs/outputs of the pneumatic circuit, a LSTM network with memory characteristics was used to build a sensorless pneumatic/gripping force estimation model. Aiming at the problem of large modeling error by using LSTM network directly, CNN was used to extract the nonlinear relationship of air pressure and gripping force in the input information to optimize the LSTM network structure, which could improve the ability of the model to describe the multi-valued correspondence/nonlinear hysteresis relationship between air pressure and gripping force, and achieve effective estimation of gripping force of pneumatic gripper. The experimental results show that compared with the LSTM prediction models, the root mean square error of the modeling estimation and validation estimation for the proposed model is reduced by 77.14% and 70.83%, and the maximum error is reduced by 79.80% and 78.84%, respectively, which proves the effectiveness of the proposed estimation method.