It is a challenge to identify food waste sources in all-weather industrial environments, as variable lighting conditions can compromise the effectiveness of visual recognition models. This study proposes and validates a robust, interpretable, and adaptive multimodal logic fusion method in which sensor dominance is dynamically assigned based on real-time illuminance intensity. The method comprises two foundational components: (1) a lightweight MobileNetV3 + EMA model for image recognition; and (2) an audio model employing Fast Fourier Transform (FFT) for feature extraction and Support Vector Machine (SVM) for classification. The key contribution of this system lies in its environment-aware conditional logic. The image model MobileNetV3 + EMA achieves an accuracy of 99.46% within the optimal brightness range (120–240 cd m⁻²), significantly outperforming the audio model. However, its performance degrades significantly outside the optimal range, while the audio model maintains an illumination-independent accuracy of 0.80, a recall of 0.78, and an F1 score of 0.80. When light intensity falls below the threshold of 84 cd m⁻², the audio recognition results take precedence. This strategy ensures robust classification accuracy under variable environmental conditions, preventing model failure. Validated on an independent test set, the fusion method achieves an overall accuracy of 90.25%, providing an interpretable and resilient solution for real-world industrial deployment.

Finite element analysis (FEA) is a key tool for predicting the mechanical behavior of additively manufactured parts. However, its predictive accuracy relies on the underlying constitutive model and, in particular, on precise model parameters. The main purpose of this study is to develop and validate a condition-grouped multimodal deep learning framework for efficiently identifying the ten parameters of the Johnson–Cook (JC) yield and fracture models for Ti-6Al-4V (TC4) alloy fabricated by laser powder bed fusion (L-PBF) using batch finite element simulation data. Using nearly 20,000 simulation datasets and response images covering different temperatures, strain rates, and stress triaxialities, convolutional neural networks (CNNs), long short-term memory (LSTM), and multilayer perceptron (MLP) models are trained to capture both sequential and physical features of the stress–strain responses. Grouping the data by test condition decouples the coupled model parameters and improves prediction accuracy. The identified yield parameters give R² values consistently above 0.94, and the stress–strain curves reconstructed from the identified parameters agree well with experimental data (R = 0.976). The failure parameters reproduce the main fracture characteristics in simulation, but the post-peak behavior remains subject to uncertainty. The proposed method therefore allows the corresponding JC parameters to be determined directly from stress–strain curves, providing a fast route for calibrating constitutive models of TC4 components.