Traditional Backpropagation Algorithm Research Articles

Interface Reliability Modeling of Coaxial Through Silicon Via Based on WOA-BP Neural Network

Abstract Due to the complex structure and thermal mismatch of coaxial through silicon via (TSV), cracks easily occur under thermal load, leading to interface delamination or spalling failure. The reliability issue of coaxial TSV is important for its application in three-dimensional packaging, so it is of great significance to predict the crack trend and evaluate the reliability of coaxial TSV. In this paper, an algorithm model with the combination of whale optimization algorithm (WOA) and back propagation (BP) neural network for the reliability prediction of coaxial TSV is proposed. Based on finite element method (FEM), the training and validation datasets of the energy release rates (ERR) of the crack at the critical interface are calculated to construct the deep learning neural network. Six key structure parameters affecting the reliability of coaxial TSV are selected as the input values of the BP neural network. The maximum relative error of whale optimization algorithm optimized back propagation (WOA-BP) neural network model is 0.88%, which is better than the prediction results of the traditional BP and genetic algorithm (GA) optimized BP models. The WOA-BP neural network model was also compared with BP and GA-BP neural network models with four error metric models. It is verified that WOA-BP neural network model has the best prediction performance. The proposed model can be used to achieve improved prediction accuracy for the interface reliability of coaxial TSV under complex structural conditions since it has higher accuracy and stronger robustness.

Thermoelectric Cooler Modeling and Optimization via Surrogate Modeling Using Implicit Physics-Constrained Neural Networks

Thermoelectric cooler (TEC) is a promising active cooling device to remove the localized hot spots precisely in VLSI chips. In this paper, we use a novel implicit physics-constrained neural networks (called IPCNN) to build a surrogate model for the single TEC device with the reduction from 3-D to 1-D. First, the surrogate model represented by the deep neural networks (DNN) allows parameterization of key design and running parameters, such as current density, length, and thermal boundary conditions of the TEC. Second, the proposed method tries to partition the physics laws into two different groups, which then are enforced by supervised learning and physics-informed neural networks (PINN) framework sequentially. Such implicit PCNN scheme can lead to much faster training speed and better convergent accuracy for the unsupervised training. The existing plain PINN enforces all the physics laws via the loss functions and the network tends to have very slow training speed and a large convergent error for large problems. An extreme learning machine (ELM) is used for the networks in the first stage. Compared with fully-connected network (FCN) trained by the traditional back-propagation algorithm, ELM can be easily trained and converges much faster. Furthermore, by leveraging the differential nature of the DNN model, we can directly estimate the derivative of the cooling heat flux with respect to current density instead of using a finite difference approximation. The calculated derivatives are used to find the optimal current density to achieve maximum cooling heat flux via Newton’s method. Last but not least, we propose a novel hybrid finite element neural network (FENN) method to perform thermal analysis of the VLSI chip system with the TEC device. The DNN model is embedded into COMSOL through the heat flux boundary conditions. Experimental results show that the ML-based method can achieve about 8.5× speedup with good accuracy than the COMSOL-based finite element method. Furthermore, the proposed IPCNN is more stable and accurate than the existing PINN. The proposed FENN can have a 5.1× speedup and 5.4× memory reduction over the traditional numerical method.

Linear Time Electromigration Analysis Based on Physics-Informed Sparse Regression

In this work, we propose a novel physics-informed sparse regression (PISR) framework to solve stress evolution (described by Korhonen’s equations) in general multi-segment wires using an unsupervised learning scheme. Unlike the existing physics-informed neural network framework (PINN), the PISR method trains the trainable weights through the Moore-Penrose generalized inverse algorithm used in extreme learning machine (ELM), which is extremely faster than the back-propagation algorithm. To improve the accuracy of PISR for complex multi-segment interconnects, we employ domain decomposition schemes in both space and time. For each subdomain, we use different trainable weights but the same shared neural network to represent each subsolution, which leads to more efficient memory usage. Furthermore, we propose to use sparse matrix techniques to accelerate the training speed of the PISR method and prove that the resulting PISR has linear time complexity for analyzing tree-structured interconnects. Finally, we divide the time into many time internals and apply an autoregressive model to simulate each time interval in sequence to further improve scalability and reduce memory cost so that the PISR method can perform EM analysis for large-scale multi-segment interconnects. Experimental results on different kinds of interconnect structures show that the proposed PISR method has the same accuracy level as the numerical methods. The results on NT T-junctions interconnect trees show that the proposed PISR method indeed demonstrates true linear time complexity. Furthermore, PISR can deliver 8.9×, 20.6×, and 1284× speedups over the recently proposed semi-analytic method (ASOV), finite difference method accelerated with model order reduction (FDM-MOR), FDM for the interconnect with NT= 5000, respectively. Furthermore, we show that PISR also achieves an 818× speedup in training over the plain PINN method based on the traditional back-propagation algorithm.

A Fusion Positioning Method for Indoor Geomagnetic/Light Intensity/Pedestrian Dead Reckoning Based on Dual-Layer Tent-Atom Search Optimization-Back Propagation.

Indoor positioning using smartphones has garnered significant research attention. Geomagnetic and sensor data offer convenient methods for achieving this goal. However, conventional geomagnetic indoor positioning encounters several limitations, including low spatial resolution, poor accuracy, and stability issues. To address these challenges, we propose a fusion positioning approach. This approach integrates geomagnetic data, light intensity measurements, and inertial navigation data, utilizing a hierarchical optimization strategy. We employ a Tent-ASO-BP model that enhances the traditional Back Propagation (BP) algorithm through the integration of chaos mapping and Atom Search Optimization (ASO). In the offline phase, we construct a dual-resolution fingerprint database using Radial Basis Function (RBF) interpolation. This database amalgamates geomagnetic and light intensity data. The fused positioning results are obtained via the first layer of the Tent-ASO-BP model. We add a second Tent-ASO-BP layer and use an improved Pedestrian Dead Reckoning (PDR) method to derive the walking trajectory from smartphone sensors. In PDR, we apply the Biased Kalman Filter-Wavelet Transform (BKF-WT) for optimal heading estimation and set a time threshold to mitigate the effects of false peaks and valleys. The second-layer model combines geomagnetic and light intensity fusion coordinates with PDR coordinates. The experimental results demonstrate that our proposed positioning method not only effectively reduces positioning errors but also improves robustness across different application scenarios.

A multi-layer photonic spiking neural network with a modified backpropagation algorithm for nonlinear classification

In this work, we construct a fully connected multi-layer photonic spiking neural network (PSNN) based on excitable vertical-cavity surface-emitting lasers with an embedded saturable absorber (VCSELs-SA) equipped with the modified supervised backpropagation (BP) algorithm. Different from the traditional BP algorithm, our proposed modified BP algorithm is based on optical spikes, and its activation function exploits the excitable characteristic of VCSELs-SA. It is demonstrated that this BP PSNN can be competent in linearly inseparable problems by correctly solving the exclusive or (XOR) problem. Thus, a three-layer fully connected BP PSNN is designed to perform the prediction of Human immunodeficiency virus-1 protease (HIV-1 PR) cleavage sites and achieves 91.95% classification accuracy with the modified BP algorithm, which could be a useful tool in helping to find effective inhibitors of HIV PR to fight against acquired immunodeficiency syndrome (AIDS). These results demonstrate the validity of our modified BP algorithm in dealing with nonlinear classification problems in PSNNs, which is helpful for supervised learning in the photonic neuromorphic computing architecture for future neuromorphic applications.

Variance-capturing forward-forward autoencoder (VFFAE): A forward learning neural network for fault detection and isolation of process data

Data-driven models have emerged as popular choices for fault detection and isolation (FDI) in process industries. However, real-time updating of these models due to streaming data requires significant computational resources, is tedious and therefore pauses difficulty in fault detection. To address this problem, in this study, we have developed a novel forward-learning neural network framework that can efficiently update data-driven models in real time for high-frequency data without compromising the accuracy. The neural network parameters are updated using a suitably constructed forward-forward learning algorithm instead of the traditional backpropagation algorithm. Firstly, we develop a variance-capturing forward-forward autoencoder (VFFAE) for FDI. Further, we showcase that the previously trained VFFAE model can be quickly adapted to incoming data which demonstrate the efficacy of the proposed framework. We have three process case studies to validate the proposed approach, namely, the Tennesse-Eastman dataset, nuclear power flux dataset, and wastewater plant dataset, to validate the proposed approach. Our findings demonstrate that within the initial 90 s, the model underwent 90 updates using a forward-forward approach and only 10 updates using backpropagation-based methods without compromising accuracy. This highlights the model’s capacity to effectively handle streaming data during the modeling process.

A Millimeter-Wave 3D Imaging Algorithm for MIMO Synthetic Aperture Radar.

Multiple-input-multiple-output synthetic aperture radar (MIMO-SAR) is being studied and applied in more and more scenarios. However, there is still a certain distance away from real-time imaging using advanced algorithms. The traditional backpropagation algorithm (BPA) multi-accumulation integration is unsuitable for dealing with large-size scanning data, and the wavenumber domain algorithm requires the array to satisfy Nyquist sampling law in azimuth to avoid aliasing in imaging reconstruction. Based on these issues, a novel 3D imaging method is proposed for MIMO-SAR. An appropriate transformation and inverse Fourier transform (FT) is carried out for the frequency domain; thus, accumulation in the wavenumber domain is not required, which is easy to implement. The computational complexity of the algorithm is much lower than BPA and has better generalizability than the wavenumber domain algorithm. Coherence factor (CF) is also introduced to achieve sidelobe suppression. Proof-of-principle experiments were also carried out in the 92.5 GHz band based on the MIMO-SAR prototype system. Both simulation and experimental results of different distributed targets show good performance of imaging and do not lose the quality of image reconstruction.

A method of using back propagation neural network to estimate orbital lifetime of LEO satellites

The “IADC space debris mitigation guidelines” requires that the post-mission orbit lifetime of low Earth orbiting (LEO) satellites should not be longer than 25 years, which is the key to debris mitigation. To analyze the degree of global compliance to this 25-year guideline, a fast and relatively accurate orbit life estimation algorithm is required. The traditional method uses numerical orbit propagation to accomplish the task, although its accuracy is high when a high-precision force model is used, the calculation is time-consuming. This paper proposes a new method for orbit lifetime estimation without relying on orbit propagation. A back propagation (BP) neural network is used to directly fit the relationship from the satellite parameters to the orbit lifetime. The space objects that have reentered the Earth’s atmosphere are selected from the catalog database of the US Space Surveillance Network (SSN), and are used to generate samples for network training. The genetic algorithm is combined with the traditional back propagation algorithm for the training of the neural network. The results show that for the specific application background of identifying whether a satellite complies with the 25-year guideline, the orbit lifetime estimation accuracy of the BP network meets the demand to some extent. Moreover, because it does not involve orbit propagation, the BP network has a high computational efficiency in estimating orbit lifetime, showing a good application prospect.

Novel Cuckoo Search-Based Metaheuristic Approach for Deep Learning Prediction of Depression

Depression is a common illness worldwide with doubtless severe implications. Due to the absence of early identification and treatment for depression, millions of individuals worldwide suffer from mental illnesses. It might be difficult to identify those who are experiencing mental health illnesses and to provide them with the early help that they need. Additionally, depression may be associated with thoughts of suicide. Currently, there are no clinically specific diagnostic biomarkers that can identify the severity and type of depression. In this research paper, the novel particle swarm-cuckoo search (PS-CS) optimization algorithm is proposed instead of the traditional backpropagation algorithm for training deep neural networks. The backpropagation algorithm is widely used for supervised learning in deep neural networks, but it has limitations in terms of convergence speed and the possibility of getting trapped in local optima. These problems were addressed by using a deep neural network architecture for depression detection tasks along with the PS-CS optimization technique. The PS-CS algorithm combines the strengths of both particle swarm optimization and cuckoo search algorithms, which allows for a more efficient and effective optimization of the network parameters. We also evaluated how well the suggested methods performed against the most widely used classification models, including (K-nearest neighbor) KNN, (support vector regression) SVR, and decision trees, as well as the most widely used deep learning models, including residual neural network (ResNet), visual geometry group (VGG), and simple neural network (LeNet). The findings show that the suggested method, PS-CS, in conjunction with the CNN model, outperformed all other models, achieving the maximum accuracy of 99.5%. Other models, such as the KNN, decision trees, and logistic regression, achieved lower accuracies ranging from 69% to 97%.

Improving performance prediction of evacuated tube solar collector through convolutional neural network method

The Convolutional Neutral Network (CNN) method is used for predicting the performance of an all-glass straight-through evacuated tube solar collector based on operational data collected over several days. The input layer of the CNN-model includes the mass flow rate, inlet temperature, solar irradiance, wind speed and outdoor temperature. The temperature difference of the collector fluid inlet and outlet, collected heat and efficiency of the collector are the main output. The CNN-model outperforms a Multiple Linear Regression (MLR) and Back Propagation (BP) neural network model by having the best prediction accuracy with the lowest Root Mean Squared Error (RMSE = 0.00577), the lowest Mean Absolute Error (MAE = 0.00357) and the highest value of Coefficient of Determination (R2 = 0.9534) for predicting the collector efficiency, and the same also applied to the temperature difference and useful solar heat. The BP-model performed better than the MLR model. MLR is therefore not recommended for performance evaluation of vacuum tube solar collectors due its poor capability to handle non-linear problems leading to lower accuracy. The prediction precision of CNN outperformed BP. A traditional BP algorithm employs a large number of weights, it is calculation intensive, and is more complex than the CNN algorithm. The CNN-model showed that solar collector performance is more affected by the solar irradiance and mass flow rate than by the wind speed and outdoor temperature.

Optimization of Performance Traditional Back-propagation with Cyclical Rule for Forecasting Model

The traditional Back-propagation algorithm has several weaknesses, including long training times and significant iterations to achieve convergence. This study aims to optimize traditional Back-propagation using the cyclical rule method to cover these weaknesses. Optimization is done by changing the training function and standard Back-propagation parameters using the training function and cyclical rule parameters. After that, a comparison of the two results will be carried out. This study uses quantitative method of time-series data on coronavirus cases sourced from the Worldometer website, then analyzed using three forecasting models with five input layers, one hidden layer (5, 10, and 15 neurons) and one output layer. The results showed that the 5-10-1 model with the training function and cyclical rule parameters and the tansig and purelin activation functions could perform well in optimization, including faster training time and smaller iterations (epochs), MSE training performance, and better tests. Low and high accuracy (92%) with an error rate of 0.01. So it was concluded that the training function and cyclical rule parameters with the tansig and purelin activation functions were able to optimize the traditional Back-propagation method, and the 5-10-1 model could be used for forecasting active cases of the coronavirus in Asia

Aspergillus niger</em> Fungus Detection using Transfer Learning Technique and Modified Backpropagation Algorithm with Inertia and Legendre Polynomial

Looking at the loss due to health problems from fungal diseases in one hand and the benefits from its industrial/agricultural use, rapid automated fungal species identification is the need of the hour. Hence, proposed a fast identification of fungal species by a 15 minutes staining procedure followed by an artificial-intelligence-based image classification technique. In this modern era, deep architectures have shown a significant performance on computer vision problems. Instead of developing a new model from scratch, the pre-trained convolutional neural network models are available to obtain the appropriate features from input samples using the transfer-learning technique. This work utilizes the transfer-learning approach for feature extraction and classification performed using the proposed modified third-term Backpropagation (BP) algorithm. This proposed algorithm contains Inertia as a third factor in the weight updation rule expanded in the form of the Legendre polynomial to overcome the limitations of the traditional Backpropagation algorithm. The effectiveness of the proposed classifier compared to the results of the existing cutting-edge algorithms namely, Backpropagation algorithm, Backpropagation algorithm using Momentum, and softmax classifier. Compare to the existing models, the proposed model scored a high testing accuracy of 97.27%.

Constrained neural classifier training method for flaw detection in industrial pipes using particle swarm optimisation

A novel method for constrained training of multi-class artificial neural network classifiers is proposed in this work. The traditional training procedure is usually based on mean square error minimisation and thus, all classes of interest are considered as having the same relevance for system performance. This is not always the case for real-world applications in which the class relevance may be unbalanced. In this paper, cost functions designed to introduce classification performance constraints for specific classes are presented and particle swarm optimisation is used as global optimisation method. The proposed method is applied to a non-destructive evaluation decision support problem using pulsed eddy currents signals. Experimental results obtained from thermally insulated industrial pipes indicate the efficiency of the proposed method in comparison to neural networks trained from the traditional back-propagation algorithm.

Fusing deep and handcrafted features for intelligent recognition of uptake patterns on thyroid scintigraphy

Thyroid scintigraphy is an important investigation for the clinical diagnosis of thyroid diseases. Thyroid diseases often present characteristic abnormal patterns in scintigraphic images. In this study, we developed an automated recognition model for thyroid uptake patterns. These patterns were classified into six categories, and they are diffusely increased, diffusely decreased, heterogeneous, focally increased, focally decreased, and normal. This study is the first report on such data for automated thyroid pattern interpretation to the best of our knowledge. A thyroid uptake pattern recognition network (TPRNet) was developed, using deep and handcrafted features of thyroid patterns to perform classification. The method can be trained using the traditional back-propagation algorithm. An in-house dataset that contains 4263 thyroid scintigraphic images was constructed to train and validate the TPRNet. Furthermore, an external test dataset that includes 1318 images was constructed to test the TPRNet. Experimental results show that the proposed method is effective for recognizing thyroid patterns in scintigraphic images. It outperforms compared methods, notably in the external test dataset, demonstrating the good generalization ability of the TPRNet. The proposed method is also compared against four physicians’ judgments on recognizing thyroid patterns, resulting in a performance that is comparable to that of experienced doctors, showing that it could be used in clinical practice.

An adaptive backpropagation algorithm for long-term electricity load forecasting.

Artificial Neural Networks (ANNs) have been widely used to determine future demand for power in the short, medium, and long terms. However, research has identified that ANNs could cause inaccurate predictions of load when used for long-term forecasting. This inaccuracy is attributed to insufficient training data and increased accumulated errors, especially in long-term estimations. This study develops an improved ANN model with an Adaptive Backpropagation Algorithm (ABPA) for best practice in the forecasting long-term load demand of electricity. The ABPA includes proposing new forecasting formulations that adjust/adapt forecast values, so it takes into consideration the deviation between trained and future input datasets' different behaviours. The architecture of the Multi-Layer Perceptron (MLP) model, along with its traditional Backpropagation Algorithm (BPA), is used as a baseline for the proposed development. The forecasting formula is further improved by introducing adjustment factors to smooth out behavioural differences between the trained and new/future datasets. A computational study based on actual monthly electricity consumption inputs from 2011 to 2020, provided by the Iraqi Ministry of Electricity, is conducted to verify the proposed adaptive algorithm's performance. Different types of energy consumption and the electricity cut period (unsatisfied demand) factor are also considered in this study as vital factors. The developed ANN model, including its proposed ABPA, is then compared with traditional and popular prediction techniques such as regression and other advanced machine learning approaches, including Recurrent Neural Networks (RNNs), to justify its superiority amongst them. The results reveal that the most accurate long-term forecasts with the minimum Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) values of (1.195.650) and (0.045), respectively, are successfully achieved by applying the proposed ABPA. It can be concluded that the proposed ABPA, including the adjustment factor, enables traditional ANN techniques to be efficiently used for long-term forecasting of electricity load demand.

Study on an Intelligent Prediction Method of Ticket Price in a Subway System with Public-Private Partnership

The accurate and rapid prediction of ticket prices for a public-private partnership (PPP) subway system, which is an important research topic in the field of civil engineering management, is of critical importance to ensure its smooth operation. To effectively cope with the effects of multiple influencing factors and strong nonlinearity among them, the mean impact value (MIV) method and the back-propagation (BP) feed-forward neural network improved by the sparrow search algorithm (SSA) are used in this study to develop an intelligent prediction model. First, we considered the relationship of the supply and the subway system service, which is a typical quasi-public product, and analyzed the relevant factors affecting its price adjustment. Then, we developed an intelligent method for the prediction of ticket prices based on the SSA-BP. This model not only makes full use of the powerful nonlinear modeling ability of the BP algorithm, but also takes advantage of the strong optimization ability and fast convergence speed of the SSA. Finally, this study screened out the key input factors by adopting the MIV method to simplify the structure of the BP algorithm and achieve a high prediction accuracy. In this study, Beijing Subway Line 4, Wuhan Metro Line 2, and Chengdu Metro Line 1 were selected as case study sites. The results showed that the linear correlations between influencing factors and ticket price for the PPP subway system service were weak, which indicated the need for using nonlinear analysis methods such as the BP algorithm. Compared with other prediction methods (the price adjustment method based on PPP contract, the traditional BP algorithm, the BP neural network improved by the genetic algorithm, the BP algorithm improved by the particle swarm optimization, and the support vector machine), the model proposed in this paper showed better prediction accuracy and calculation stability.

Awesome back-propagation machine learning paradigm

For a better future in machine learning (ML), it is necessary to modify our current concepts to get the fastest ML. Many designers had attempted to find the optimal learning rates in their applications through many algorithms over the past decades, but they have not yet achieved their target of highest speed of back-propagation (BP). This research proposes a novel BP rule called the Instant Learning Ratios-Machine Learning (ILR-ML) or (ILRML). Unlike the traditional BP algorithms, the ILR-ML offers its learning without the concepts of the learning rate(s). The ILR-ML has a new concept called the "Learning Ratio" and indicated by a sign (Δl). The ILR-ML performs the full BP algorithm with 100% accuracy per each learning iteration. The ILR-ML is more suitable for the online machine learning.

Extreme learning machine-based spherical harmonic for fast ionospheric delay modeling

Due to the growing real-time data of a global navigation satellite system (GNSS) applications and strict requirements for positioning users, especially under the condition of abnormal ionospheric activity, the fast and high-accuracy model is increasingly becoming important. In the past few years, the spherical harmonic (SH) model is one of the commonly-used methods, which has been widely used for global and regional modeling of ionospheric delay. In this paper, in order to further improve the accuracy of the SH model whilst without extra time consumption, we explored how the extreme learning machine (ELM) and the SH model can be combined for accurate and real-time modeling of ionospheric delay. Feasibility of our proposed method (denoted as ELM-SH model) was evaluated, where the available ionospheric data was firstly used to produce an SH model, while the ELM was deployed to compensate for the error of the SH model. Afterward, we tested the ELM-SH model based on the ionospheric data obtained from the Continuously Operating Reference Stations (CORS) in Jiangsu Province, China. Notably, experiment results demonstrate that compared with the conventional SH model, our proposed ELM-SH method can greatly improve the precision with over 37.09% on Root Mean Square Error (RMSE). The ELM-SH achieves competitive accuracy against the traditional backpropagation algorithm (BP) based method, but with a very short modeling time (within 10−2 s). This will be significant for real-time ionosphere modeling for those real-time GNSS positioning applications.

Time-series prediction model of PM2.5 concentration based on LSTM neural network

LSTM (long-short term memory neural network) is a neural network suitable for simulating and predicting time series changes. This paper designs a PM2.5 concentration prediction model based on the LSTM neural network. Use Python language programming implementation under the Tensor Flow framework, training the model using real weather data from 2015-01-01∼2019-01-01, and the obtained model can accurately predict the concentration change of future PM2.5. RMSE and DC were used as evaluation indexes to compare the LSTM neural network model with the traditional BP (Back Propagation) algorithm model. It turns out, LSTM model has the characteristics of high precision and stable prediction effect, proved that the time series prediction ability of the LSTM model is more superior. As an effective method for predicting PM2.5 concentration data, it can provide the basis for predicting pollutants.

Speech Emotion Analysis of Netizens Based on Bidirectional LSTM and PGCDBN

In recent years, the application of speech emotion recognition (SER) in the supervision of Internet public opinion has received increasing attention. This study proposes a new SER algorithm to analyze the public opinion information of network platforms. Firstly, we extract different spectrum features from speech signals and combine them into frame level speech features. Then, we select conditional deep confidence network (CDBN) which has the ability to learn sequential features as the final classification model. We apply particle swarm optimization (PSO) and genetic algorithm (GA) during the fine-tuning stage of the CDBN to obtain more suitable optimal weights of the whole network, and propose the PSO-GA-CDBN (PGCDBN) model. Compare with the traditional back propagation (BP) algorithm, our training method accelerates the convergence speed of the network and improves the robustness and recognition performance of the network. In our experiment, we used the Chinese Academy of Sciences’ Institute of automation (CASIA) Chinese emotional corpus and self-collected Chinese speech datasets, which were collected from Sina Weibo, Tik tok and other online social media platforms. Compare with the popular emotion classifiers such as support vector machine (SVM), deep residual network (ResNet), long short-term memory (LSTM) neural network, DBN, our proposed PGCDBN achieves the best recognition results from both datasets. In addition, we use bidirectional LSTM before PGCDBN to further process the extracted speech features, and the result of bidirectional LSTM has stronger speech signal expression ability. The average recognition accuracy of this new hybrid deep learning model algorithm in two datasets is 98.67%, which can be used for the supervision of netizens’ opinions.