An intensive computation source has become increasingly important in recent years to meet the time-critical and low-latency needs of Industrial Internet of Things (IIoT) systems. Existing IIoT-based devices are built with limited computational resources, delivering results in a limited fashion when used in highly resource-intensive applications. Since then, a novel concept known as Edge Computing (EC) has been introduced to reduce network latency and alleviate strain on cloud data centers using an EC server located at the network's periphery. The EC server only managed to gather a small number of resources compared to the resource cloud. Without prior context about task deadline and load, an EC server could not optimally handle latency-sensitive and computation-intensive tasks. Additionally, the EC server did not significantly improve overhead minimization when sending data to and from the remote cloud and the user's device. Parallel to the development of EC, nonorthogonal multiple access (NOMA) has been identified as a technique with the potential to substantially increase spectrum efficiency. In this paper, a NOMA-based EC framework for IIoT system is examined, in which multiple task nodes transfer their task via NOMA to multiple edge servers in proximity for execution. As such, this paper aims to develop a joint optimization model for making decisions about task offloading and allocating resources in Industrial edge computing. An adaptive resource allocation decision model (ARADM) based on deep reinforcement learning (DRL) and heuristically modified long short-term memory (H-LSTM) using hybrid Cat and Mouse Dingo Optimization (HCMDO) is proposed to allocate the task optimally. We formulate joint optimization by considering multi-constraint objective function with communication, computation, and cache parameters using HCMDO. Further, these optimal parameters are used in training an H-LSTM along with benchmark dataset. The outcome of the H-LSTM network utilized in DRL to improve convergence speed, accuracy and stability by predicting optimal cost and load. The goal is to minimize service delay, energy consumption, balance load and maximize resource utilization. The experimental results validated the developed model and its ability to improve the quality of resource allocation in Industrial edge computing.