Tool condition monitoring (TCM) is crucial for manufacturing systems to maximize productivity, maintain part quality, and reduce waste and cost. Available TCM systems mainly depend on data-driven classical machine learning methods to analyze different sensors’ feedback signals for tool condition prediction. Despite their applicability for high process variability and part complexity, they require long development lead time and extensive expert efforts for signal feature definition, extraction, and fusion to accurately detect the tool condition. Additionally, they substantially depend on sensors whose nature is intrusive to the cutting process. Therefore, this research presents a generalized, nonintrusive multisignal fusion approach for real-time tool wear detection in milling that redefines process learning directly from raw signals. In this two-stage approach, the signals’ intrinsic mode functions (IMFs) are extracted, optimized, and directly fused in a deep long short-term memory (LSTM) recurrent neural network (RNN) for tool condition prediction. The IMF extraction and optimization mask the effect of the cutting conditions to accentuate the tool condition effect. Therefore, it generalizes and minimizes the learning process to cover a wider range of unlearned process parameters. Embedded feature architecting of the LSTM-RNN is applied to the optimized IMFs for signal fusion and tool condition prediction to standardize the learning process and significantly minimize the lead time. Spindle motor current, voltage, and power signals are used to avoid process intrusion. A systematic study is carried out to define the optimum LSTM-RNN architecture. Extensive experimental validation results have demonstrated tool wear detection accuracy >95 % at different ranges of unlearned cutting conditions.