Channel estimation (CE) is essential for wireless links but becomes progressively onerous as Fifth Generation (5G) Multi-Input Multi-Output (MIMO) systems and extensive fading expand the search space and increase latency. This study redefines CE support as the process of learning to deduce channel type and signal-tonoise ratio (SNR) directly from per-tone Orthogonal Frequency-Division Multiplexing (OFDM) observations,with blind channel state information (CSI). We trained a dual deep model that combined Convolutional Neural Networks (CNNs) with Bidirectional Recurrent Neural Networks (BRNNs). We used a lookup table (LUT) label for channel type (class indices instead of per-tap values) and ordinal supervision for SNR (0–20 dB,5-dB steps). The method was tested on Single-Input Single-Output (SISO),the 2×2 Alamouti space-time code,and 4×4 Quasi-Orthogonal Space-Time Block Coding (QO-STBC) in six standard situations: Nakagami fading,Log-Normal shadowing,Multipath fading,Gaussian,Rayleigh fading,and Rician fading. Channel identification was nearly perfect,and the SNR was robust,with most SNR errors being in adjacent bins indicating stable behaviour. The model reached 99.68% validation accuracy with 8.14 × 10−5 bit error rate (BER) and reduced complexity of 1.78 × 108 for high order of subcarriers The method’s novelty lies in accurate,low-complexity CE support from raw symbols and its demonstrated impact on end-to-end BER pilotless CE and SNR estimation to select equalizer without CSI reconstruction.
