In this paper, we propose a deep convolutional neural network for camera based wildfire detection. We train the neural network via transfer learning and use window based analysis strategy to increase the fire detection rate. To achieve computational efficiency, we calculate frequency response of the kernels in convolutional and dense layers and eliminate those filters with low energy impulse response. Moreover, to reduce the storage for edge devices, we compare the convolutional kernels in Fourier domain and discard similar filters using the cosine similarity measure in the frequency domain. We test the performance of the neural network with a variety of wildfire video clips and the pruned system performs as good as the regular network in daytime wild fire detection, and it also works well on some night wild fire video clips.

Graph convolutional networks (GCNs) as the emerging neural networks have shown great success in Prognostics and Health Management because they can not only extract node features but can also mine relationship between nodes in the graph data. However, the most existing GCNs-based methods are still limited by graph quality, variable working conditions, and limited data, making them difficult to obtain remarkable performance. Therefore, it is proposed in this paper a two stage importance-aware subgraph convolutional network based on multi-source sensors named I²SGCN to address the above-mentioned limitations. In the real-world scenarios, it is found that the diagnostic performance of the most existing GCNs is commonly bounded by the graph quality because it is hard to get high quality through a single sensor. Therefore, we leveraged multi-source sensors to construct graphs that contain more fault-based information of mechanical equipment. Then, we discovered that unsupervised domain adaptation (UDA) methods only use single stage to achieve cross-domain fault diagnosis and ignore more refined feature extraction, which can make the representations contained in the features inadequate. Hence, it is proposed the two-stage fault diagnosis in the whole framework to achieve UDA. In the first stage, the multiple-instance learning is adopted to obtain the importance factor of each sensor towards preliminary fault diagnosis. In the second stage, it is proposed I²SGCN to achieve refined cross-domain fault diagnosis. Moreover, we observed that deficient and limited data may cause label bias and biased training, leading to reduced generalization capacity of the proposed method. Therefore, we constructed the feature-based graph and importance-based graph to jointly mine more effective relationship and then presented a subgraph learning strategy, which not only enriches sufficient and complementary features but also regularizes the training. Comprehensive experiments conducted on four case studies demonstrate the effectiveness and superiority of the proposed method for cross-domain fault diagnosis, which outperforms the state-of-the art methods.

Physically unclonable functions (PUFs) are a class of hardware-specific security primitives based on secret keys extracted from integrated circuits, which can protect important information against cyberattacks and reverse engineering. Here, we put forward an emerging type of PUF in the electromagnetic domain by virtue of the self-dual absorber-emitter singularity that uniquely exists in the non-Hermitian parity-time ( PT )–symmetric structures. At this self-dual singular point, the reconfigurable emissive and absorptive properties with order-of-magnitude differences in scattered power can respond sensitively to admittance or phase perturbations caused by, for example, manufacturing imperfectness. Consequently, the entropy sourced from inevitable manufacturing variations can be amplified, yielding excellent PUF security metrics in terms of randomness and uniqueness. We show that this electromagnetic PUF can be robust against machine learning–assisted attacks based on the Fourier regression and generative adversarial network. Moreover, the proposed PUF concept is wavelength-scalable in radio frequency, terahertz, infrared, and optical systems, paving a promising avenue toward applications of cryptography and encryption.

In this article, we present a computationally efficient algorithm to detect Volatile Organic Compounds (VOC) leaking out of components used in chemical processes in petrochemical refineries and chemical plants. A leaking VOC plume from a damaged component appears as a dynamic dark cloud in infrared videos. We describe a two-stage deep neural network structure, taking advantage of both spatial and temporal structure of the dynamic texture regions created by the leaking VOC plume. We first detect moving pixels which are darker then their neighboring pixels. We extract one-dimensional (1-D) signals representing the temporal history of such pixels from video and feed the 1-D signals to a 1-D convolutional neural network. If those pixels are near the edge of a VOC plume, their 1-D temporal signals exhibit high-frequency behavior. The neural network generates high probability estimates for such pixels. If 1-D neural network generates high confidence values, final decision is reached using a deep convolutional neural network (CNN) which processes image frames. The overall structure is computationally efficient because the spatio-temporal CNN does not process all of the image frames of the captured video. Experimental results are presented.

We propose a co-design approach for compute-in-memory inference for deep neural networks (DNN). We use multiplication-free function approximators based on l ₁ norm along with a co-adapted processing array and compute flow. Using the approach, we overcame many deficiencies in the current art of in-SRAM DNN processing such as the need for digital-to-analog converters (DACs) at each operating SRAM row/column, the need for high precision analog-to-digital converters (ADCs), limited support for multi-bit precision weights, and limited vector-scale parallelism. Our co-adapted implementation seamlessly extends to multi-bit precision weights, it doesn't require DACs, and it easily extends to higher vector-scale parallelism. We also propose an SRAM-immersed successive approximation ADC (SA-ADC), where we exploit the parasitic capacitance of bit lines of SRAM array as a capacitive DAC. Since the dominant area overhead in SA-ADC comes due to its capacitive DAC, by exploiting the intrinsic parasitic of SRAM array, our approach allows low area implementation of within-SRAM SA-ADC. Our 8×62 SRAM macro, which requires a 5-bit ADC, achieves ~105 tera operations per second per Watt (TOPS/W) with 8-bit input/weight processing at 45 nm CMOS. Our 8×30 SRAM macro, which requires a 4-bit ADC, achieves ~84 TOPS/W. SRAM macros that require lower ADC precision are more tolerant of process variability, however, have lower TOPS/W as well. We evaluated the accuracy and performance of our proposed network for MNIST, CIFAR10, and CIFAR100 datasets. We chose a network configuration which adaptively mixes multiplication-free and regular operators. The network configurations utilize the multiplication-free operator for more than 85% operations from the total. The selected configurations are 98.6% accurate for MNIST, 90.2% for CIFAR10, and 66.9% for CIFAR100. Since most of the operations in the considered configurations are based on proposed SRAM macros, our compute-in-memory's efficiency benefits broadly translate to the system-level.