Recently, guided wave-based techniques have garnered increased attention from researchers in the field of Structural Health Monitoring (SHM) for damage detection and quantification. Extracting features that are sensitive to changes in structural conditions has become a critical step. Guided waves, being one of the most widely used techniques, are highly responsive to structural changes caused by various types of damage. However, due to disturbances from environmental and operational conditions (EOCs) and waveforms reflected from boundaries and damage in complex systems, manually extracting features related to the system state is challenging. To address this, various neural networks, capable of automatically deriving features, have been developed and applied to system identification tasks. Among these, Convolutional Autoencoders (CAEs) have proven effective as damage detectors, projecting data efficiently into a latent space. However, existing CAE-based models for damage diagnosis often overlook the influence of EOCs, which can significantly impact model performance. In this work, a scheme that provides accurate damage level classification while estimating EOCs, such as loading conditions, making the model adaptable to varying external factors is introduced. By restructuring and combining the CAE with feedforward neural networks (FFNNs), the proposed scheme also enables signal reconstruction, a potentially valuable feature when data is limited. The method has been validated on an aluminum plate with various damage levels under different loading conditions, achieving near-perfect state classification and low-error signal reconstruction. The framework has also been tested on incomplete datasets to verify its robustness.