To address the current issues in relay life prediction where the correlation between degradation states is underutilized and manual parameter tuning is inefficient, this study proposes an electromagnetic relay life prediction method based on the grey wolf optimizer (GWO) and bidirectional long short-term memory (BiLSTM) network. First, an accelerated degradation experiment platform for electromagnetic relays is constructed to collect full service-life data, from which the mean time to failure is derived. The data is then sampled, partitioned, and normalized, with four key feature parameters—coil resistance, load current, contact resistance, and release time—selected as model inputs using Pearson correlation analysis. Subsequently, the root mean squared error (RMSE) between predicted and actual relay life is employed as the fitness function for the GWO algorithm, optimizing the number of hidden layer neurons, dropout rate, and initial learning rate of the BiLSTM model. The prediction model is reconstructed using the optimal parameter combination. Experimental validation demonstrates that the proposed GWO-BiLSTM model achieves average reductions of 56.7% in RMSE and 58.2% in MAPE compared to conventional back-propagation neural network (BPNN), gated recurrent unit (GRU), and long short-term memory (LSTM) models, effectively enhancing the prediction accuracy of electromagnetic relay service life.