Blind source separation (BSS) is a mathematical technique that addresses the challenge of separating a mixture of signals into its original, independent sources without prior knowledge of the mixing process. The main objective of this research is to propose a BSS algorithm focused on identifying mutually uncorrelated sources that drive the network, which consists of discrete-time linear time-invariant (LTI) interconnected modules. This study first improves an existing BSS method, which uses a finite impulse response (FIR) model in the prediction error method (PEM) framework to model the input-output transfer function matrix and make it applicable in real-time scenarios. Implemented on field data from Hifi Engineering Inc., this algorithm has demonstrated reliability in online estimating sources over several commercial deployments. However, this method, refined by choosing an automatic regularization factor, still reveals certain limitations: the optimal order of the FIR model remains indeterminate, affecting accuracy and computational efficiency. Moreover, precisely predicting delays between measurement locations and adjusting various parameters requires significant manual intervention, posing barriers to scalability and automation. Next, we propose a novel method within a subspace framework to address these constraints, using the known network topology while minimizing prediction error. Given the network's structure, this approach first integrates elements of PEM with subspace identification techniques, allowing us to identify parameterized state-space matrices through sequential optimization processes aimed at reducing prediction error, which is defined as the difference between the estimated and parameterized Kalman filter predictor. This technique functions within a steady-state Kalman framework, employing data equations to accurately predict the dynamics of the multi-input multi-output (MIMO) network. The performance of the structured BSS method is evaluated by comparing the results of estimating the independent sources obtained through structured state matrices against those generated by conventional subspace methods with unstructured state matrices. The simulation results indicate that structured BSS, given the network topology, offers superior accuracy and robustness in source estimation. This progress in BSS shows how having prior knowledge about the network structure can help separate sources more efficiently and reliably in complex networks.