Stacked Sparse Autoencoder Research Articles

Image enhancement methods for inspection of planar and non-planar FRP structures using a noise-based microwave NDT inspection system

Microwave nondestructive testing (MNDT) includes inspection techniques that assess a particular material’s health status using low-power and contactless inspection systems. In near-field microwave inspections, imaging results are heavily influenced by the standoff distance parameter, i.e., the physical separation between the microwave sensor and the sample under test (SUT). Variations in the standoff distance during an inspection tend to cause defect masking of disbonds and delaminations in fiber-reinforced polymer (FRP) materials, causing defects to go undetected frequently. An MNDT near-field inspection system using noise waveforms is used to identify engineered internal defects within carbon fiber-reinforced polymer (CFRP) samples. Tactics utilizing Principal Component Analysis (PCA), Stacked Sparse Autoencoders (SSAEs), and an autoencoder network trained in a manner for anomaly detection are used to minimize the effects of standoff distance, reduce defect masking, and increase the ability to identify hidden defects. The samples tested are constructed to possess planar and non-planar geometries, such that the viability of the data-driven image enhancement and standoff distance correction methods are demonstrated with respect to a wide variety of in situ inspection applications.

Leveraging electrocardiography signals for deep learning-driven cardiovascular disease classification model

Electrocardiography (ECG) is the most non-invasive diagnostic tool for cardiovascular diseases (CVDs). Automatic analysis of ECG signals assists in accurately and rapidly detecting life-threatening arrhythmias like atrioventricular blockage, atrial fibrillation, ventricular tachycardia, etc. The ECG recognition models need to utilize algorithms to detect various kinds of waveforms in the ECG and identify complicated relationships over time. However, the high variability of wave morphology among patients and noise are challenging issues. Physicians frequently utilize automated ECG abnormality recognition models to classify long-term ECG signals. Recently, deep learning (DL) models can be used to achieve enhanced ECG recognition accuracy in the healthcare decision making system. In this aspect, this study introduces an automated DL enabled ECG signal recognition (ADL-ECGSR) technique for CVD detection and classification. The ADL-ECGSR technique employs three most important subprocesses: pre-processed, feature extraction, parameter tuning, and classification. Besides, the ADL-ECGSR technique involves the design of a bidirectional long short-term memory (BiLSTM) based feature extractor, and the Adamax optimizer is utilized to optimize the trained method of the BiLSTM model. Finally, the dragonfly algorithm (DFA) with a stacked sparse autoencoder (SSAE) module is applied to recognize and classify EEG signals. An extensive range of simulations occur on benchmark PTB-XL datasets to validate the enhanced ECG recognition efficiency. The comparative analysis of the ADL-ECGSR methodology showed a remarkable performance of 91.24 % on the existing methods.

Optimizing MobileNetV2 for improved accuracy in early gastric cancer detection based on dynamic pelican optimizer

This paper presents an innovative framework for the automated diagnosis of gastric cancer using artificial intelligence. The proposed approach utilizes a customized deep learning model called MobileNetV2, which is optimized using a Dynamic variant of the Pelican Optimization Algorithm (DPOA). By combining these advanced techniques, it is feasible to achieve highly accurate results when applied to a dataset of endoscopic gastric images. To evaluate the performance of the model based on the benchmark, its data is divided into training (80 %) and testing (20 %) sets. The MobileNetV2/DPOA model demonstrated an impressive accuracy of 97.73 %, precision of 97.88 %, specificity of 97.72 %, sensitivity of 96.35 %, Matthews Correlation Coefficient (MCC) of 96.58 %, and F1-score of 98.41 %. These results surpassed those obtained by other well-known models, such as Convolutional Neural Networks (CNN), Mask Region-Based Convolutional Neural Networks (Mask R–CNN), U-Net, Deep Stacked Sparse Autoencoder Neural Networks (SANNs), and DeepLab v3+, in terms of most quantitative metrics. Despite the promising outcomes, it is important to note that further research is needed. Specifically, larger and more diverse datasets as well as exhaustive clinical validation are necessary to validate the effectiveness of the proposed method. By implementing this innovative approach in the detection of gastric cancer, it is possible to enhance the speed and accuracy of diagnosis, leading to improved patient care and better allocation of healthcare resources.

SCAE-Stacked Convolutional Autoencoder for Fault Diagnosis of a Hydraulic Piston Pump with Limited Data Samples.

Deep learning (DL) models require enormous amounts of data to produce reliable diagnosis results. The superiority of DL models over traditional machine learning (ML) methods in terms of feature extraction, feature dimension reduction, and diagnosis performance has been shown in various studies of fault diagnosis systems. However, data acquisition can sometimes be compromised by sensor issues, resulting in limited data samples. In this study, we propose a novel DL model based on a stacked convolutional autoencoder (SCAE) to address the challenge of limited data. The innovation of the SCAE model lies in its ability to enhance gradient information flow and extract richer hierarchical features, leading to superior diagnostic performance even with limited and noisy data samples. This article describes the development of a fault diagnosis method for a hydraulic piston pump using time-frequency visual pattern recognition. The proposed SCAE model has been evaluated on limited data samples of a hydraulic piston pump. The findings of the experiment demonstrate that the suggested approach can achieve excellent diagnostic performance with over 99.5% accuracy. Additionally, the SCAE model has outperformed traditional DL models such as deep neural networks (DNN), standard stacked sparse autoencoders (SSAE), and convolutional neural networks (CNN) in terms of diagnosis performance. Furthermore, the proposed model demonstrates robust performance under noisy data conditions, further highlighting its effectiveness and reliability.

Research on tap breakage monitoring method for tapping process based on SSAELSTM fusion network

Aiming at the problems of high tap breakage rate during tapping of small-diameter internal threads and the difficulty of feature extraction and fusion in characterising tap breakage, this paper proposes a tap breakage monitoring method based on stack sparse autoencoder long-short-term memory (SSAELSTM) fusion network. Specifically, based on change rule of torque in the tapping process and the analysis of the processing mechanism, the normal label and the breakage label are defined, and 36 groups of feature in the multi-feature domain of the torque signal are extracted, constituting a labelled set of feature. Additionally, a stacked sparse autoencoder network (SSAE) is employed to fuse the feature that serve as inputs to the long short-term memory network (LSTM) for recognition, and the classification errors of LSTM are back-propagated to SSAE to optimize the weights of the LSTM and SSAE networks. Finally, 31 SSAELSTM fusion network models with different structures are constructed, and the final tap breakage monitoring model is determined by comparing the performance indexes. Compared with the features before fusion, the constructed model fusion effectively reduces the amount of data and improves the correlation with the labels. Compared with other network models, it was found that SSAELSTM fusion network is superior to other models in both accuracy and robustness, providing breakage warning for the tapping process, and laying the foundation for subsequent decision-making to avoid tap breakage.

Early identification of out-of-roundness damage wheels in railway freight vehicles using a wayside system and a stacked sparse autoencoder

Early identification of wheel defects can prevent serious damage to railways, considerably lowering maintenance costs for both railway administrations and rolling stock operators. Within this context, an unsupervised methodology based on artificial intelligence techniques is presented, which allows the detection and classification of out-of-roundness damage wheels, such as wheel flats and polygonal wheels, based on dynamic responses induced on the track by crossing freight railway vehicles. The methodology involves the following steps: (i) data collection and pre-processing, (ii) feature extraction (iii) data fusion and (iv) feature discrimination. In the first phase, an FFT algorithm is applied to the acceleration track responses. Then, features are extracted after training a Stacked Sparse Autoencoder, in which the main features of the responses are obtained after a compression stage using an encoder network. This lower dimensional layer forces the model to learn a compression of the input data. Then, these extracted features are merged using the Mahalanobis distance, which enhances the sensitivity to the damage recognition. Posteriorly, an outlier analysis is performed to distinguish a healthy wheel from a defective one and a cluster analysis to discriminate the two types of out-of-roundness (OOR) damage and classify the severity of each type of damage.

Deep Learning-Based Active Jamming Suppression for Radar Main Lobe

Due to the development of digital radio frequency memory (DRFM), active jamming against the main lobe of the radar has become mainstream in electronic warfare. The jamming infiltrates the radar receiver via the main lobe, covering up the target echo information. This greatly affects the detection, tracking, and localization of targets by radar. In this study, we consider jamming suppression based on the independence of RF features. First, two stacked sparse auto-encoders (SSAEs) are built to extract the RF characteristics and signal features carried out by the actual radar signal for subsequent jamming suppression. This method can effectively separate RF features from signal features, making the extracted RF features more efficient and accurate. Then, an SSAE-based jamming suppression auto-encoder (JSAE) is proposed; the mixed signal, including the radar signal, jamming signal, and noise, is input to JSAE for dimensionality reduction. Therefore, the radar signal and RF features, extracted by the two SSAEs in the previous step, are used to constrain the features of the reduced mixed signal. Moreover, we integrate the feature level and signal level to jointly achieve jamming suppression. The original radar signal is used to assist the radar signal reconstructed by the decoder. By first filtering out interference-related features and then reconstructing the signal, we can achieve better jamming suppression performance. Finally, the effectiveness of the proposed method is verified by simulating the actual collected data.

Deep self-taught learning framework for intrusion detection in cloud computing environment

<p>Cloud has become a target-rich environment for malicious attacks by cyber intruders. Security is a major concern and remains an obstacle to the adoption of cloud computing. The intrusion detection system (IDS) is regarded as defense-in-depth. Unfortunately, most machine learning approaches designed for cloud intrusion detection require large amounts of labeled attack samples, but in real practice, they are limited. Therefore, the key impetus of this work is to introduce self-taught learning (STL) combining stacked sparse autoencoder (SSAE) with long short-term memory (LSTM) as a candidate solution to learn the robust feature representation and efficiently improve the performance of IDS with respect to false alarm rate (FAR) and detection rate (DR). Accordingly, the proposed approach as a first step employs SSAE to achieve dimensional reduction by learning the discriminative features from network traffic. The approach adopts LSTM to recognize the intrusion with the features encoded by SSAE. To evaluate the detective performance of our model, a comprehensive set of experiments are conducted on NSL-KDD. Also, ablation experiments are conducted to show the contribution of each component of our approach. Further, the comparative analysis shows the efficacy of our approach against the existing approaches with an accuracy of 86.31%.</p>

Status Recognition of Marine Centrifugal Pumps Based on a Stacked Sparse Auto-Encoder

Marine centrifugal pumps (MCPs) are widely used in ships, so it is important to identify their status accurately for their maintenance. Due to the influence of load, friction, and other non-linear factors, the vibration signal of an MCP shows non-linear and non-stationary characteristics, and it is difficult to extract the state characteristics contained in the vibration signal. To solve the difficulty of feature extraction of non-linear non-stationary vibration signals generated by MCPs, a novel MCP frequency domain signal feature extraction method based on a stacked sparse auto-encoder (SSAE) is proposed. The characteristic parameters of MCP frequency domain signals are extracted via the SSAE model for classification training, and different statuses of MCPs are identified. The vibration signals in different MCP statuses were collected for feature extraction and classification training, and the MCP status recognition accuracy based on the time domain feature and fuzzy entropy feature was compared. According to the test data, the accuracy of MCP status recognition based on the time domain feature is 71.2%, the accuracy of MCP status recognition based on the fuzzy entropy feature is 87.7%, and the accuracy of MCP status recognition based on the proposed method is 100%. These results show that the proposed method can accurately identify each status of an MCP under test conditions.

Empowering EEG motor imagery classification with deep transfer learning approach

AbstractThe brain‐computer interface (BCI) enables individuals with impairments to interact with the real world without relying on the neuromuscular pathway. BCI leverages artificial intelligence (AI) models for control. It can capture brain activity patterns associated with mental processes and convert them into commands for actuators. One promising application of BCI is in rehabilitation centres. When compared to traditional methods integrated with machine learning (ML) approaches that were initially explored in the past decade, there is a relatively limited number of studies utilizing deep learning (DL) approaches in BCI applications. A novel approach has been developed for the automatic recognition of motor imagery (MI) tasks. This article introduces the improved sparrow search algorithm with deep transfer learning‐based EEG motor imagery classification (ISSADTL‐EEGMIC) algorithm for BCIs. The presented ISSADTL‐EEGMIC algorithm primarily focuses on the automated classification of motor imagery tasks. To achieve this, the ISSADTL‐EEGMIC method first preprocesses EEG signals using the wavelet packet decomposition (WPD) technique. Additionally, it employs the neural architecture search network (NASNet) technique for feature extraction. Moreover, the ISSADTL‐EEGMIC model performs the classification process using the stacked sparse autoencoder (SSAE) model. Furthermore, hyperparameter optimization of the SSAE model is performed using the ISSA technique, which helps achieve maximum classification performance. The results from the simulation of the ISSADTL‐EEGMIC algorithm have been examined using two benchmark datasets. The experimental findings suggest that it outperforms recent approaches in terms of performance.

A battle royale optimization with feature fusion-based automated fruit disease grading and classification

<abstract> <p>Fruit Disease Detection (FDD) using Computer Vision (CV) techniques is a powerful strategy to accomplish precision agriculture. Because, these techniques assist the farmers in identifying and treating the diseased fruits before it spreads to other plants, thus resulting in better crop yield and quality. Further, it also helps in reducing the usage of pesticides and other chemicals so that the farmers can streamline their efforts with high accuracy and avoid unwanted treatments. FDD and Deep Learning (DL)-based classification involve the deployment of Artificial Intelligence (AI), mainly the DL approach, to identify and classify different types of diseases that affect the fruit crops. The DL approach, especially the Convolutional Neural Network (CNN), has been trained to classify the fruit images as diseased or healthy, based on the presence or absence of the disease symptoms. In this background, the current study developed a new Battle Royale Optimization with a Feature Fusion Based Fruit Disease Grading and Classification (BROFF-FDGC) technique. In the presented BROFF-FDGC technique, the Bilateral Filtering (BF) approach is primarily employed for the noise removal process. Besides, a fusion of DL models, namely Inception v3, NASNet, and Xception models, is used for the feature extraction process with Bayesian Optimization (BO) algorithm as a hyperparameter optimizer. Moreover, the BROFF-FDGC technique employed the Stacked Sparse Autoencoder (SSAE) algorithm for fruit disease classification. Furthermore, the BRO technique is also employed for optimum hyperparameter tuning of the SSAE technique. The proposed BROFF-FDGC system was simulated extensively for validation using the test database and the outcomes established the enhanced performance of the proposed system. The obtained outcomes emphasize the superior performance of the BROFF-FDGC approach than the existing methodologies.</p> </abstract>

Improved network intrusion classification with attention-assisted bidirectional LSTM and optimized sparse contractive autoencoders

Accurately identifying network intrusion cannot only help individuals and enterprises better deal with network security problems, but also maintain the Internet environment. This work proposes a new hybrid classification method named SABD for network intrusion detection. SABD integrates Stacked sparse contractive autoencoders (SSCA), Attention-based Bidirectional long-term and short-term memory (LSTM), and Decision fusion. Specifically, SSCA is used for extracting features, which are sent to the attention-based bidirectional LSTM for the classification. Besides, an improved optimization algorithm named genetic simulated-annealing-based particle swarm optimization is designed to optimize hyperparameters of SSCA. Finally, the decision fusion algorithm is adopted to integrate classification results of multiple classifiers and yield the final results. Based on experimental results from four different types of data sets, the proposed SABD outperforms its most advanced peers in classification accuracy.

Fault diagnosis of brushless DC motor based on Stack Sparse Autoencoder

Because of their simple structure, long service life, high efficiency, etc., brushless DC (BLDC) motors have been widely applied in many fields. In some applications, high requirements for BLDC continuous use of motors, so often to BLDC running state monitoring of motors, realizes the early fault diagnosis, to improve the reliability, safety, and prolonged use. To solve this problem, a BLDC fault diagnosis method based on Fast Fourier Transform (FFT), Stacked Sparse Auto-encoder (SSAE), and soft classifier was proposed. In the laboratory, a BLDC model R-3525 was used as the experimental equipment, and the data of 9 kinds of fault conditions such as bearing damage, cage damage, inner and outer ring damage, and the normal condition of the motor were collected. The collected data are transformed into 28 × 28 two-dimensional data sets by FFT. The feature expressions of various faults are adaptively learned from many data, and the intelligent diagnosis of the motor is realized by the feature function expression. The experimental results demonstrate that, in comparison to the single neural network approach, the stacked method can significantly enhance the precision of fault classification and achieve fault diagnosis for brushless DC motors. This method carries valuable implications for other varieties of BLDC motors.

Reconstruction-based stacked sparse auto-encoder for nonlinear industrial process fault diagnosis

The reconstruction-based (RB) approach can effectively suppress the misdiagnosis problem due to the smearing effect in fault isolation. However, the current exploration of the RB approach for large-scale nonlinear systems is still limited. Therefore, this paper proposes a reliable and effective fault diagnosis method based on a reconstruction-based stacked sparse autoencoder (RBSSAE) for high-dimensional industrial systems. In RBSSAE, a reconstruction-based index achieved by the Steffensen iterative method is developed to check whether the given variable(s) are responsible for the faults efficiently. However, the number of possible faulty variable combinations grows exponentially with the system dimension or actual abnormal variables, causing an unbearable computational burden. Hence, the proposed RBSSAE utilizes a sequential floating forward selection approach to rapidly isolate the most decisive variable combination, meeting a requirement of online fault diagnosis. Finally, the effectiveness of the RBSSAE is verified on a numerical example and a real industrial case.

Optimal Deep Learning-Based Recognition Model for EEG Enabled Brain-Computer Interfaces Using Motor-Imagery

Abstract Brain-Computer Interfaces (BCIs) facilitate the translation of brain activity into actionable commands and act as a crucial link between the human brain and the external environment. Electroencephalography (EEG)-based BCIs, which focus on motor imagery, have emerged as an important area of study in this domain. They are used in neurorehabilitation, neuroprosthetics, and gaming, among other applications. Optimal Deep Learning-Based Recognition for EEG Signal Motor Imagery (ODLR-EEGSM) is a novel approach presented in this article that aims to improve the recognition of motor imagery from EEG signals. The proposed method includes several crucial stages to improve the precision and effectiveness of EEG-based motor imagery recognition. The pre-processing phase starts with the Variation Mode Decomposition (VMD) technique, which is used to improve EEG signals. The EEG signals are decomposed into different oscillatory modes by VMD, laying the groundwork for subsequent feature extraction. Feature extraction is a crucial component of the ODLR-EEGSM method. In this study, we use Stacked Sparse Auto Encoder (SSAE) models to identify significant patterns in the pre-processed EEG data. Our approach is based on the classification model using Deep Wavelet Neural Network (DWNN) optimized with Chaotic Dragonfly Algorithm (CDFA). CDFA optimizes the weight and bias values of the DWNN, significantly improving the classification accuracy of motor imagery. To evaluate the efficacy of the ODLR-EEGSM method, we use benchmark datasets to perform rigorous performance validation. The results show that our approach outperforms current methods in the classification of EEG motor imagery, confirming its promising performance. This study has the potential to make brain-computer interface applications in various fields more accurate and efficient, and pave the way for brain-controlled interactions with external systems and devices.

An industrial process fault diagnosis method based on independent slow feature analysis and stacked sparse autoencoder network

Deep learning, with its powerful multilayer nonlinear representation of deep neural networks, enables models trained based on deep learning to describe the true distribution of data more accurately and thus have better generalization capabilities. However, the training data are usually assumed to obey independent identical distribution. Industrial process data nowadays usually have complex characteristics such as non-Gaussian, nonlinear, dynamic, strong correlation, and high-dimensional, which make it difficult to meet the assumptions of deep learning. Using industrial process data directly would result in low model accuracy and bias in the model output. This paper proposes a fault diagnosis method based on independent slow feature analysis (ISFA) and stacked sparse autoencoder (SSAE) network. The statistical properties possessed by ISFA are used to perform preliminary feature extraction of industrial process data, and then the extracted preliminary features are input to the SSAE network for fine-grained feature extraction. The proposed two-step feature extraction method not only makes the data input to the SSAE network as close as possible to obey independent identical distribution and provides convincing fault diagnosis results, but also makes full use of the powerful feature extraction capability of the SSAE network, which makes the proposed method also significantly improve the detection accuracy of incipient faults. The feasibility and superiority of the proposed method is verified through the TE process.

Intelligent Resource Allocation Using an Artificial Ecosystem Optimizer with Deep Learning on UAV Networks

An Unmanned Aerial Vehicle (UAV)-based cellular network over a millimeter wave (mmWave) frequency band addresses the necessities of flexible coverage and high data rate in the next-generation network. But, the use of a wide range of antennas and higher propagation loss in mmWave networks results in high power utilization and UAVs are limited by low-capacity onboard batteries. To cut down the energy cost of UAV-aided mmWave networks, Energy Harvesting (EH) is a promising solution. But, it is a challenge to sustain strong connectivity in UAV-based terrestrial cellular networks due to the random nature of renewable energy. With this motivation, this article introduces an intelligent resource allocation using an artificial ecosystem optimizer with a deep learning (IRA-AEODL) technique on UAV networks. The presented IRA-AEODL technique aims to effectually allot the resources in wireless UAV networks. In this case, the IRA-AEODL technique focuses on the maximization of system utility over all users, combined user association, energy scheduling, and trajectory design. To optimally allocate the UAV policies, the stacked sparse autoencoder (SSAE) model is used in the UAV networks. For the hyperparameter tuning process, the AEO algorithm is used for enhancing the performance of the SSAE model. The experimental results of the IRA-AEODL technique are examined under different aspects and the outcomes stated the improved performance of the IRA-AEODL approach over recent state of art approaches.

A novel method of health indicator construction and remaining useful life prediction based on deep learning

The construction of health indicators (HI) for traditional deep learning requires human training labels and poor interpretability. This paper proposes an HI construction method based on Stacked Sparse Autoencoder (SSAE) and combines SSAE with Long short-term memory (LSTM) network to predict the remaining useful life (RUL). Extracting features from a single domain may result in insufficient feature extraction and cannot comprehensively reflect the degradation status information of mechanical equipment. In order to solve the problem, this article extracts features from time domain, frequency domain, and time-frequency domain to construct a comprehensive original feature set. Based on monotonicity, trendiness, and robustness, the most sensitive features from the original feature set are selected and put into the SSAE network to construct HI for state partitioning, and then LSTM is used for RUL prediction. By comparing with the existing methods, it is proved that the prediction effect of the proposed method in this paper is satisfied.

Digital twin model for chiller fault diagnosis based on SSAE and transfer learning

The equipment of chiller systems is characterized by a complex mechanical structure and operating environments that vary widely, resulting in high failure rates, energy waste, and costly maintenance. Traditional fault diagnosis methods suffer from low levels of digitalization and intelligence. Therefore, timely and accurate detection and diagnosis of the operating status of chiller systems are critical. To address the above issues, this study proposes a digital twin (DT) model for chiller system fault diagnosis based on stacked sparse auto-encoder (SSAE) and transfer learning (TL). First, a chiller system digital twin mapping model is constructed utilizing digital twin technology. Then, the SSAE model is constructed to provide real-time defect diagnosis and fault result validation. Finally, to address the limited chiller system data in practical applications, the knowledge of fault diagnosis acquired in the source domain is transferred to the target domain via TL. Experimental results demonstrate that the SSAE model outperforms back propagation (BP), recurrent neural network (RNN) and long short-term memory (LSTM) model in term of accuracy. Furthermore, using TL achieves a diagnostic accuracy of over 90% for different degrees of fault severity, with a significant decrease in cross-entropy loss compared to no TL, confirming the effectiveness of the TL method.

Spatial and temporal prediction of secondary crashes combining stacked sparse auto-encoder and long short-term memory

Secondary crashes occur within the spatial and temporal impact area of primary crashes, resulting in traffic delays and safety problems. While most existing studies focus on the likelihood of secondary crashes, predicting the spatio-temporal location of secondary crashes could offer valuable insights for implementing prevention strategies. This includes guiding the deployment of emergency response measures and determining appropriate speed limits. The main objective of this study is to develop a prediction method for the spatial and temporal locations of secondary crashes. A hybrid deep learning model SSAE-LSTM is proposed by combining stacked sparse auto-encoder (SSAE) and long short-term memory network (LSTM). Traffic and crash data on the California I-880 highway covering the period of 2017–2021 are collected. The identification of secondary crashes is performed by the speed contour map method. The time and distance gaps between primary and secondary crashes are modeled using multiple 5-minute interval traffic variables as inputs. Multiple models are developed for benchmarking purposes, including PCA-LSTM, which incorporates principal component analysis (PCA) and LSTM, SSAE-SVM, which incorporates SSAE and support vector machine (SVM), and back propagation neural network (BPNN). The performance comparison indicates that the hybrid SSAE-LSTM model outperforms the other models in terms of both spatial and temporal prediction. In particular, SSAE4-LSTM1 (with 4 SSAE layers and 1 LSTM layer) demonstrates superior spatial prediction performance, while SSAE4-LSTM2 (with 4 SSAE layers and 2 LSTM layers) excels in temporal prediction. A joint spatio-temporal evaluation is also conducted to measure the overall accuracy of the optimal models over different permitted spatio-temporal ranges. Finally, practical suggestions are provided for secondary crash prevention.