Strawberry maturity detection plays an essential role in modern strawberry yield estimation and robot-assisted picking and sorting. Due to the small size and complex growth environment of strawberries, there are still problems with existing recognition systems’ accuracy and maturity classifications. This article proposes a strawberry maturity recognition algorithm based on an improved YOLOv5s model named YOLOv5s-BiCE. This algorithm model is a replacement of the upsampling algorithm with a CARAFE module structure. It is an improvement on the previous model in terms of its content-aware processing; it also widens the field of vision and maintains a high level of efficiency, resulting in improved object detection capabilities. This article also introduces a double attention mechanism named Biformed for small-target detection, optimizing computing allocation, and enhancing content perception flexibility. Via multi-scale feature fusion, we utilized double attention mechanisms to reduce the number of redundant computations. Additionally, the Focal_EIOU optimization method was introduced to improve its accuracy and address issues related to uneven sample classification in the loss function. The YOLOv5s-BiCE algorithm was better at recognizing strawberry maturity compared to the original YOLOv5s model. It achieved a 2.8% increase in the mean average precision and a 7.4% increase in accuracy for the strawberry maturity dataset. The improved algorithm outperformed other networks, like YOLOv4-tiny, YOLOv4-lite-e, YOLOv4-lite-s, YOLOv7, and Fast RCNN, with recognition accuracy improvements of 3.3%, 4.7%, 4.2%, 1.5%, and 2.2%, respectively. In addition, we developed a corresponding detection app and combined the algorithm with DeepSort to apply it to patrol robots. It was found that the detection algorithm exhibits a fast real-time detection speed, can support intelligent estimations of strawberry yield, and can assist picking robots.

Maize is susceptible to various diseases throughout its growth cycle, which can significantly reduce yields. The accurate identification of maize diseases with similar symptomatic manifestations is particularly challenging under field conditions due to heterogeneous lighting and variable weather conditions. This paper proposes a novel detection model named SCFM-DETR, which is based on an improved Real-Time DEtection TRansformer (RT-DETR) to achieve robust identification of maize diseases in complex environments. SimAM-StarNet is employed as the backbone for feature extraction in this model, reducing the number of parameters and improving multiscale feature fusion, thereby diminishing the impact of background noise. Furthermore, the original RepC3 module is replaced with a newly designed CGLU-FasterBlock-MANet (CFM) module, which enhances adaptive feature fusion for finer discriminative capability. The experimental results demonstrate that the SCFM-DETR model achieves an average precision of 96.7% and a recall of 95.8% on a maize disease dataset, exceeding the corresponding metrics of the baseline RT-DETR-R18 model by 3.1% and 6.0%. Additionally, the model reduces the number of parameters and computational load by 47% and 49%, respectively, making it highly suitable for deployment in computationally limited agricultural settings. This work offers a high-accuracy, lightweight framework that facilitates intelligent crop disease monitoring and supports the advancement of smart agriculture.

Streaming incident diagnosis requires integrating long telemetry streams—logs, traces, alerts, metrics, tickets, and operator actions—under a context budget that is much smaller than the event history. Existing long-context mechanisms improve token capacity, but they do not directly specify which delayed failure dependencies should be preserved when the stream is noisy, drifting, and only partially observed. We present Causal-Dependency State Space Prompting (CSSP), a bounded-memory framework that treats diagnosis as online failure-dependency state estimation rather than raw prompt extension. CSSP combines (i) an event canonicalizer that normalizes templates, typed entities, and high-cardinality identifiers, (ii) a selective state-space memory trained with weak dependency supervision derived from request identifiers, entity co-reference, and template-transition rules, and (iii) a learned state-token interface that exposes the latent state to the same frozen 7B instruction-following reasoner used by all large language model baselines. We explicitly soften the causal claim: the supervision does not prove interventional causality, but encourages the memory to preserve operational dependency evidence relevant to diagnosis. A constrained verifier checks evidence alignment, tool consistency, and unsafe actions with one bounded revision step. Across BGL, HDFS, and OpenStack, CSSP obtains the best mean F 1 with five-run confidence intervals, remains robust to log corruption and noisy weak links, and explains the smaller OpenStack gain through incident-type analysis. These results support dependency-aware state tokenization as a practical interface for reliable bounded-memory incident reasoning.

Anomaly detection is vital for maintaining the reliability of smart agricultural control systems, especially under conditions where sensor faults may disrupt automated irrigation and environmental monitoring. To address the limitations of traditional manual inspections and purely supervised methods, this study proposes DE-TabNet—a semi-supervised anomaly detection model based on bidirectional encoding. The model integrates temporal sequence learning via a BiGRU-VAE network and feature dependency extraction through a sparse autoencoder, forming an unsupervised feature fusion network. This representation is subsequently fine-tuned using a limited set of labelled data through a TabNet-based supervised module. Experimental evaluations on a real-world smart drip irrigation dataset demonstrated that DE-TabNet significantly outperforms state-of-the-art methods, including the DAGMM, iForest, VAE-LSTM, OCSVM, and Semi-TabNet methods. Specifically, DE-TabNet achieved up to 86.9% accuracy and an 80.8% F1 score, corresponding to improvements of 11.6% and 3.82%, respectively, over those of the baseline models. Additional validation on public datasets confirmed the model’s strong generalization ability, highlighting its practical value for robust anomaly detection in agricultural and other time series sensor systems.