Neural Network Calibration Research Articles

A Multi-Task Convolutional Neural Network Relative Radiometric Calibration Based on Temporal Information

Due to the continuous degradation of onboard satellite instruments over time, satellite images undergo degradation, necessitating calibration for tasks reliant on satellite data. The previous relative radiometric calibration methods are mainly categorized into traditional methods and deep learning methods. The traditional methods involve complex computations for each calibration, while deep-learning-based approaches tend to oversimplify the calibration process, utilizing generic computer vision models without tailored structures for calibration tasks. In this paper, we address the unique challenges of calibration by introducing a novel approach: a multi-task convolutional neural network calibration model leveraging temporal information. This pioneering method is the first to integrate temporal dynamics into the architecture of neural network calibration models. Extensive experiments conducted on the FY3A/B/C VIRR datasets showcase the superior performance of our approach compared to the existing state-of-the-art traditional and deep learning methods. Furthermore, tests with various backbones confirm the broad applicability of our framework across different convolutional neural networks.

Machine learning modeling for fuel cell-battery hybrid power system dynamics in a Toyota Mirai 2 vehicle under various drive cycles

Electrification is considered essential for the decarbonization of mobility sector, and understanding and modeling the complex behavior of modern fuel cell-battery electric-electric hybrid power systems is challenging, especially for product development and diagnostics requiring quick turnaround and fast computation. In this study, a novel modeling approach is developed, utilizing supervised machine learning algorithms, to replicate the dynamic characteristics of the fuel cell-battery hybrid power system in a 2021 Toyota Mirai 2nd generation (Mirai 2) vehicle under various drive cycles. The entire data for this study is collected by instrumenting the Mirai vehicle with in-house data acquisition devices and tapping into the Mirai controller area network bus during chassis dynamometer tests. A multi-input - multi-output, feed-forward artificial neural network architecture is designed to predict not only the fuel cell attributes, such as average minimum cell voltage, coolant and cathode air outlet temperatures, but also the battery hybrid system attributes, including lithium-ion battery pack voltage and temperature with the help of 15 system operating parameters. Over 21,0000 data points on various drive cycles having combinations of transient and near steady-state driving conditions are collected, out of which around 15,000 points are used for training the network and 6,000 for the evaluation of the model performance. Various data filtration techniques and neural network calibration processes are explored to condition the data and understand the impact on model performance. The calibrated neural network accurately predicts the hybrid power system dynamics with an R-squared value greater than 0.98, demonstrating the potential of machine learning algorithms for system development and diagnostics.

Enhancing the accuracy of low-cost thermocouple devices through deep-wavelet neural network calibration

Data collection using thermocouple sensors in low-cost data acquisition is prone to noise interference, which could reduce the data quality. Noise sources such as cold junction compensators, electromagnetic interference, and Johnson noise can significantly affect the reliability and accuracy of conventional measurements. This study aims to improve the quality of thermocouple sensor readings on low-cost data acquisition using calibration method based on deep learning and the denoising process using a wavelet transform. This taken approach successfully increase the accuracy value of 97.67% with a mean absolute error (MAE) of 0.2. The precision also increases of 262.7% as indicated by the result of signal-to-noise ratio (SNR) with a value of 105.29 dB. Comparative analysis was carried out against National Instruments® device and it was found that deep-wavelet method had a lower and higher of MAE and SNRdB values of 16.67% and 0.8% respectively. This study shows that the denoising-calibration method with deep-wavelet can improve the accuracy and reliability of data from low-cost thermocouple devices.

Enhancing accuracy of surface wind sensors in wind tunnel Testing: A Physics-Guided neural network calibration approach

Irwin’s surface wind sensor is widely used in wind tunnel testing for urban and environmental aerodynamics studies. However, the conventional physics-based calibration of this sensor could result in reduced measurement accuracy in regions with low flow velocities and high turbulence intensity. To address this issue, this study proposes a novel physics-guided neural network (PGNN) calibration approach, which couples a physics-based calibration model, derived from extended Taylor series expansions of measured wind speed, with an adaptive, data-driven general regression neural network. Sensors are calibrated within the turbulent boundary layer of an empty flat plate, considering both mean and standard deviation of wind velocity measured by high-accuracy thermal anemometry. The accuracy of calibrated sensors is then assessed using a 1:400 benchmark urban model. Experimental results show significant improvement in measurement accuracy, reducing mean absolute percentage error for wind speed standard deviation from 92.3 % with the current model to 9.8 % using PGNN.

SimCalib: Graph Neural Network Calibration Based on Similarity between Nodes

Graph neural networks (GNNs) have exhibited impressive performance in modeling graph data as exemplified in various applications. Recently, the GNN calibration problem has attracted increasing attention, especially in cost-sensitive scenarios. Previous work has gained empirical insights on the issue, and devised effective approaches for it, but theoretical supports still fall short. In this work, we shed light on the relationship between GNN calibration and nodewise similarity via theoretical analysis. A novel calibration framework, named SimCalib, is accordingly proposed to consider similarity between nodes at global and local levels. At the global level, the Mahalanobis distance between the current node and class prototypes is integrated to implicitly consider similarity between the current node and all nodes in the same class. At the local level, the similarity of node representation movement dynamics, quantified by nodewise homophily and relative degree, is considered. Informed about the application of nodewise movement patterns in analyzing nodewise behavior on the over-smoothing problem, we empirically present a possible relationship between over-smoothing and GNN calibration problem. Experimentally, we discover a correlation between nodewise similarity and model calibration improvement, in alignment with our theoretical results. Additionally, we conduct extensive experiments investigating different design factors and demonstrate the effectiveness of our proposed SimCalib framework for GNN calibration by achieving state-of-the-art performance on 14 out of 16 benchmarks.

Asphalt Pavement Damage Detection through Deep Learning Technique and Cost-Effective Equipment: A Case Study in Urban Roads Crossed by Tramway Lines

Asphalt pavements are subject to regular inspection and maintenance activities over time. Many techniques have been suggested to evaluate pavement surface conditions, but most of these are either labour-intensive tasks or require costly instruments. This article describes a robust intelligent pavement distress inspection system that uses cost-effective equipment and the ‘you only look once’ detection algorithm (YOLOv3). A dataset for flexible pavement distress detection with around 13,135 images and 30,989 bounding boxes of damage was used during the neural network training, calibration, and validation phases. During the testing phase, the model achieved a mean average precision of up to 80%, depending on the type of pavement distress. The performance metrics (loss, precision, recall, and RMSE) that were applied to estimate the object detection accuracy demonstrate that the technique can distinguish between different types of asphalt pavement damage with remarkable accuracy and precision. Moreover, the confusion matrix obtained in the validation process shows a distress classification sensitivity of up to 98.7%. The suggested technique was successfully implemented in an inspection car. Measurements conducted on urban roads crossed by tramway lines in the city of Palermo proved the real-time ability and great efficacy of the detection system, with potentially remarkable advances in asphalt pavement examination efficacy due to the high rates of correct distress detection.

Neural Network calibration method for VARANS models to simulate wave-coastal structures interaction

This study develops a calibration method for the porous media to properly model the interaction between waves and coastal structures using VARANS models. The proposed method estimates the porosity, np, and the optimum values of the Forchheimer coefficients, α and β, that best represent the wave-structure interaction for a complete set of laboratory tests. Physical tests were conducted in a 2D wave flume for a homogeneous mound breakwater under regular wave conditions. Numerical tests were carried out using the IH-2VOF model to simulate the corresponding physical tests and incident wave conditions (HI, T). The numerical tests covered a wide range of Forchheimer coefficients found in the literature, α and β, and the porosity, np, with a total of 555 numerical tests. The results of 375 numerical tests using IH-2VOF were used to train a Neural Network (NN) model with five input variables (HI, T, np,α and β) and one output variable (KR2). The NN model explained more than 90% (R2 > 0.90 and RMSE <5%) of the variance of the squared coefficient of reflection, KR2. This NN model was used to estimate the KR2 in a wide range of np,α and β, and the error (εa) between the physical measurements with regular waves and the NN estimations of KR2 was calculated. The results of εa as function of np,α and β showed that for a given porosity, np, it was difficult to obtain a pair of α and β values that gave a common low error if few physical tests are used for calibration. Then to calibrate properly a VARANS model it seems necessary to check the results obtained for each combination of α and β with many laboratory {HI, T} tests. The minimum root-mean-square error of KR2 (εrms) was calculated to find the optimum values of porosity and Forchheimer coefficients: np = 0.44,α = 200 and β = 2.825 for the tested structure. Blind tests were conducted with the remaining 180 numerical tests using IH-2VOF to validate the proposed method for VARANS models. In this study, eight or more physical tests were required to find adequate values of np,α and β for VARANS models related to the best performance of wave-porous structure interaction.

The design of underwater tactile force sensor with differential pressure structure and backpropagation neural network calibration

The underwater tactile force measurement was prone to cross-sensitivity, causing the difficulty in distinguishing tactile force signal with the underwater complex environment of water pressure influence. For this problem, an underwater tactile force sensor whose sensing core was based on Microelectromechanical Systems (MEMS) was designed with differential pressure typed structure. The hollow hemispherical flexible contacts located at the upper and lower end, and the hollow cylindrical shell in the middle part composed the structure of the capsule-shaped sensor. The upper flexible contact could sense the compound signal composed of water pressure and tactile force, at the same time, the lower flexible contact could measure the water pressure information. The deformation signal of the upper and lower flexible contacts could be transformed to the force sensor core’s upper and lower surfaces with silicon oil filled in the inner hollow part of the sensor. The tactile force signal could be obtained with water pressure eliminated through vector superposition method under the influence of static pressure of water. The structure and manufacture technology were introduced, and the Backpropagation (BP) neural network data regression algorithm was designed for the cross sensitivity. The experiments are conducted to demonstrate the effectiveness of the differential pressure structure in eliminating the influence of water static pressure. The results indicated that the BP neural network data regression algorithm successfully produced real tactile force signals, which is highly beneficial for the intelligent operation of underwater dexterous hand. Additionally, the sensor has an accuracy of 5%.

NCTV: Neural Clamping Toolkit and Visualization for Neural Network Calibration

With the advancement of deep learning technology, neural networks have demonstrated their excellent ability to provide accurate predictions in many tasks. However, a lack of consideration for neural network calibration will not gain trust from humans, even for high-accuracy models. In this regard, the gap between the confidence of the model's predictions and the actual correctness likelihood must be bridged to derive a well-calibrated model. In this paper, we introduce the Neural Clamping Toolkit, the first open-source framework designed to help developers employ state-of-the-art model-agnostic calibrated models. Furthermore, we provide animations and interactive sections in the demonstration to familiarize researchers with calibration in neural networks. A Colab tutorial on utilizing our toolkit is also introduced.

On the impact of deep neural network calibration on adaptive edge offloading for image classification

Edge devices can offload deep neural network (DNN) inference to the cloud to overcome energy or processing constraints. Nevertheless, offloading adds communication delay, which increases the overall inference time. An alternative is to use adaptive offloading based on early-exit DNNs. Early-exit DNNs have branches inserted at the output of given intermediate layers. These side branches provide confidence estimates. If the confidence level of the decision produced is sufficient, the inference is made by the side branch. Otherwise, the edge offloads the inference decision to the cloud, which implements the remaining DNN layers. Thus, the offloading decision depends on reliable confidence levels provided by the side branches at the device. This article provides an extensive calibration study on different datasets and early-exit DNNs for the image classification task. Our study shows that early-exit DNNs are often miscalibrated, overestimating their prediction confidence and making unreliable offloading decisions. To evaluate the impact of calibration on accuracy and latency, we introduce two novel application-level metrics and evaluate well-known DNN models in a realistic edge computing scenario. The results demonstrated that calibrating early-exit DNNs improves the probabilities of meeting accuracy and latency requirements.

A subranging nonuniform sampling memristive neural network-based analog-to-digital converter

This work presents a novel 4-bit subranging nonuniform sampling (NUS) memristive neural network-based analog-to-digital converter (ADC) with improved performance trade-off among speed, power, area, and accuracy. The proposed design preserves the memristive neural network calibration and utilizes a trainable memristor weight to adapt to device mismatch and increase accuracy. Rather than conventional binary searching, we adopt quaternary searching in the ADC to realize subranging architecture’s coarse and fine bits determination. A level-crossing nonuniform sampling (NUS) is introduced to the proposed ADC to enhance the ENOB under the same resolutions, power, and area consumption. Area and power consumption are reduced through circuit sharing between different stages of bit determination. The proposed 4-bit ADC achieves a highest ENOB of 5.96 and 5.6 at cut-off frequency (128 MHz) with power consumption of 0.515 mW and a figure of merit (FoM) of 82.95 fJ/conv.

Simultaneous Quantitative Analysis of Salmeterol and Fluticasone in Inhalation Spray Using HPLC and a Fast Spectrophotometric Technique Combined with a Time Series Neural Network and Multivariate Calibration Methods.

Chromatographic methods have been used for the simultaneous determination of salmeterol (SMT) and fluticasone (FLU). The methods take a lot of analysis time, need sample pre-treatment and large amounts of solvents, and are costly. The aim of this paper was to propose a simple, quick, and low-cost method for the determination of SMT and FLU using a time series neural network and multivariate calibration methods, including partial least-squares (PLS) and principal component regression (PCR). The simultaneous spectrophotometric determination of SMT and FLU in binary mixtures and anti-asthma spray was performed by applying multivariate calibration methods and an intelligent approach. The coefficient of determination (R2) of the time series neural network was obtained: 1 and 0.9997 for SMT and FLU, respectively. The mean recovery of PLS was found to be 99.29% and 99.84% for SMT and FLU, respectively. Also, the mean recovery related to the PCR method was 102.05% and 103.72% for SMT and FLU, respectively. The inhalation spray was assessed using HPLC and the results were compared with chemometrics methods via an analysis of variance (ANOVA) test. Intelligent and multivariate calibration methods were proposed. Simultaneous spectrophotometric determination of salmeterol and fluticasone was studied in an anti-asthma spray. HPLC as a reference method was performed and compared with chemometrics methods. Rapid, simple, low-cost, and accurate are the benefits of the proposed approaches.

Neural Network Calibration of Star Trackers

To maintain ideal performance, star trackers must be able to predict the direction of incident starlight to within a few arcseconds across the entire instrument field of view. Parametric camera models are commonly employed to calculate star vectors from camera images and correct for aberrations in the instrument optics. This conventional approach can be quite effective, but systematic errors can be difficult to eliminate and the proper selection of calibration basis functions is often difficult to determine. This study explores using supervised machine learning approaches such as Radial Basis Function networks (RBFNs) and Support Vector Machines (SVMs) for star tracker calibration as an alternative to conventional aberration formulations. These networks can be formulated either as a correction to a low-order camera model or complete replacement for the whole model. When applied to the instrument calibration of a dozen Sinclair Interplanetary ST-16RT2 sensors the RBFN formulation offers 27% reduction in the calibration residuals and almost 12% reduction in the validation residuals over conventional formulations.

A 2.5-GS/s Four-Way-Interleaved Ringamp-Based Pipelined-SAR ADC with Digital Background Calibration in 28-nm CMOS

A 2.5-GS/s 12-bit four-way time-interleaved pipelined-SAR ADC is presented in 28-nm CMOS. A bias-enhanced ring amplifier is utilized as the residue amplifier to achieve high bandwidth and excellent power efficiency compared with a traditional operational amplifier. A high linearity front-end is proposed to alleviate the non-linearity of the diode for ESD protection in the input PAD. The embedded input buffer can suppress the kickback noise at high input frequencies. A blind background calibration based on digital-mixing is used to correct the mismatches between channels. Additionally, an optional neural network calibration is also provided. The prototype ADC achieves a low-frequency SNDR/SFDR of 51.0/68.0 dB, translating a competitive FoMw of 0.48 pJ/conv.-step at 250 MHz input running at 2.5 GS/s.

Artificial neural network and multivariate calibration methods assisted UV spectrophotometric technique for the simultaneous determination of metformin and Pioglitazone in anti-diabetic tablet dosage form

In this paper, for the simultaneous measurement of Metformin (MET) and Pioglitazone (PIO) in Actoplus MET as antidiabetic tablet, chemometrics and spectrophotometry methods without the need for separation steps was used. The applied chemometrics methods were artificial neural network (ANN), partial least squares regression (PLS), and principal component regression (PCR). The ANN consisting of two, four and six layers with 2, 4, 6, 8, and 10 neurons was trained using a feed forward back-propagation (FFBP) learning. The algorithms used were Levenberg-Marquardt (LM) and gradient descent with momentum and adaptive learning rate back propagation (GDX). The mean square error (MSE) of the LM algorithm was obtained 3.18 ​× ​10−30 and 1.58 ​× ​10−30 for MET and PIO, respectively, which represented that the LM algorithm performed better than the GDX algorithm. In the PLS method, lower root mean square error (RMSE) (MET ​= ​0.0558, PIO ​= ​0.3981) showed better performance compared PCR method (MET ​= ​0.0559, PIO ​= ​0.4048). Finally, the results of the proposed methods and high performance liquid chromatography (HPLC) as a reference approach were compared with one-way ANOVA test at 95% confidence level, which did not show a significant difference between the data.

Modelling the functional dependency between root and shoot compartments to predict the impact of the environment on the architecture of the whole plant: methodology for model fitting on simulated data using Deep Learning techniques

AbstractTree structural and biomass growth studies mainly focus on the shoot compartment. Tree roots usually have to be taken apart due to the difficulties involved in measuring and observing this compartment, particularly root growth. In the context of climate change, the study of tree structural plasticity has become crucial and both shoot and root systems need to be considered simultaneously as they play a joint role in adapting traits to climate change (water availability for roots and light or carbon availability for shoots). We developed a botanically accurate whole-plant model and its simulator (RoCoCau) with a linkable external module (TOY) to represent shoot and root compartment dependencies and hence tree structural plasticity in different air and soil environments. This paper describes a new deep neural network calibration trained on simulated data sets computed from a set of more than 360 000 random TOY parameter values and random climate values. These data sets were used for training and for validation. For this purpose, we chose VoxNet, a convolutional neural network designed to classify 3D objects represented as a voxelized scene. We recommend further improvements for VoxNet inputs, outputs and training. We were able to teach the network to predict the value of environment data well (mean error &amp;lt; 2 %), and to predict the value of TOY parameters for plants under water stress conditions (mean error &amp;lt; 5 % for all parameters), and for any environmental growing conditions (mean error &amp;lt; 20 %).

Robustness of neural network calibration model for accurate spatial positioning.

The present study devotes to a systematical exploration for the robustness of neural network-based camera calibration method in the circumstance of three-dimensional (3D) spatial positioning via machine vision technique. By analyzing the error propagation route in the calibration-reconstruction process, a dimensionless error attenuation coefficient is proposed to measure the robustness of a calibration model with respect to input calibration error. Using this metric, the robustness of the neural network (NN) model under different optical configurations, i.e., input noise level, optical distortion and camera viewing angle, are analyzed in detail via synthetic simulation. Due to its generalized fitting capacity, the NN model is found to be superior to conventional pinhole model and polynomial model in terms of model robustness. To take full advantage of this feature, the NN model is further deployed to the scenarios of asymmetric camera layout and multiple camera joint calibration. Both synthetic simulation and experiment test demonstrate that the NN model can significantly improve the robustness and the accuracy of 3D spatial positioning in these non-normal scenarios.

On a Neural Network to Extract Implied Information from American Options

Extracting implied information, like volatility and dividend, from observed option prices is a challenging task when dealing with American options, because of the complex-shaped early-exercise regions and the computational costs to solve the corresponding mathematical problem repeatedly. We will employ a data-driven machine learning approach to estimate the Black-Scholes implied volatility and the dividend yield for American options in a fast and robust way. To determine the implied volatility, the inverse function is approximated by an artificial neural network on the effective computational domain of interest, which decouples the offline (training) and online (prediction) stages and thus eliminates the need for an iterative process. In the case of an unknown dividend yield, we formulate the inverse problem as a calibration problem and determine simultaneously the implied volatility and dividend yield. For this, a generic and robust calibration framework, the Calibration Neural Network (CaNN), is introduced to estimate multiple parameters. It is shown that machine learning can be used as an efficient numerical technique to extract implied information from American options, particularly when considering multiple early-exercise regions due to negative interest rates.

A Modified Online Intelligent Method to Calibrate Radar and Camera Sensors for Data Fusing

This paper introduces a modified online intelligent calibration method to calibrate radar and camera sensors in traffic scenarios. Radar and camera sensor calibration is to construct a one-to-one mapping between the radar coordinate system and the image pixel coordinate system. Based on the currently widely used artificial intelligence method (deep neural networks), camera can accurately recognize the target type. For the research and application of radar sensor and camera sensor fusion, the most important thing is to synchronize the sensor sampling time of the radar and the camera and space calibration. This paper emphasizes on spatial calibration. At present, the common calibration methods are the calibration method based on the internal and external parameters of the camera, the four-point calibration method, and the neural network calibration method. However, these three calibration methods have significant advantages and disadvantages. In the field of transportation, the disadvantages are even more prominent. This paper proposes a modified online intelligent calibration method for the application field of radar and camera sensors in traffic scenarios. We have proved through simulation experiments that this method is less time-consuming and manual operation while ensuring high-precision calibration performance.

Rugged and Compact Three-Axis Force/Torque Sensor for Wearable Robots

In the field of robotics, sensors are crucial in enabling the interaction between robots and their users. To ensure this interaction, sensors mainly measure the user’s strength, and based on this, wearable robots are controlled. In this paper, we propose a novel three-axis force/torque sensor for wearable robots that is compact and has a high load capacity. The bolt and nut combination of the proposed sensor is designed to measure high-load weights, and the simple structure of this combination allows the sensor to be compact and light. Additionally, to measure the three-axis force/torque, we design three capacitance-sensing cells. These cells are arranged in parallel to measure the difference in capacitance between the positive and negative electrodes. From the capacitance change measured by these sensing cells, force/torque information is converted through deep neural network calibration. The sensing point can also be confirmed using the geometric and kinematic relation of the sensor. The proposed sensor is manufactured through a simple and inexpensive process using cheap and simply structured components. The performance of the sensor, such as its repeatability and capacity, is evaluated using several experimental setups. In addition, the sensor is applied to a wearable robot to measure the force of an artificial muscle.