Optimization Research on Interactive Methods of Ideological and Political Education in Colleges and Universities under Intelligent Teaching Environment

This system utilizes teaching models that include perspective characters, multimedia blackboards, showcases, and venue room structures, which can be drawn as 3D models using 3ds Max. After that it is exported to. FBX format, then this file is imported into Unity3D for function integration, during which it is necessary to write logic scripts and attribute settings in Unity3D, write scripts for development, and finally export the fully functional model as an apk format file [18]. The process of developing the model is shown in Figure 1. 1)

Scenario and learning tool modeling

3D learning scene is the core of this system modeling. It mainly includes the venue room, multimedia blackboard, showcase and so on. First use 3ds Max software for all objects object three-dimensional modeling, the process can be used to edit some of the existing model examples, such as rectangular, sphere, etc., and then need to model the surface rendering process or mapping. One of the multimedia blackboards needs to reduce its surface reflection effect. The reflection is too strong, so the screen display will be white. The subsequent video rendering texture will further enhance its visual effect. For space reasons, the detailed modeling process will not be described here.

Model building and optimization process needs to be corrected on the axis of the model (i.e., are rotated around the ¥ axis), more conducive to the unified combination of all models, each large model in the construction process there is also a parent-child model of the relationship, which is the need to fix the parent-child relationship. Model construction and optimization are shown in Figure 2.

Figure 1.

Model development process

Figure 2.

Model build and optimization

Figure 3.

Automatic online modeling process

In this paper, the raw audio data is processed using Perceptual Masking Deep Neural Network (PM-DNN) based algorithm.The main principle of PM-DNN noise reduction algorithm is to elaborate the time-frequency masking associated between speech and noise.This algorithm is capable of high-performance noise reduction for processing speech in complex environments. The noise reduction process is divided into two parts: training and enhancement [20]. In the training phase, noise common to offline classroom environments is added to the pure speech, and the mapping relationship between the classroom noise and the pure speech used for training is obtained after neural network training. In the testing phase the classroom audio test set with noise is input to the neural network model for training to get the pure classroom speech after noise reduction.The model of PM-DNN noise reduction algorithm is shown in Fig. 6.

Figure 6.

PM-DNN noise reduction algorithm model

The specific noise reduction process is as follows: 1)

Use PM-DNN network to train pure speech features and output separated speech amplitude spectrum S˜ and ambient noise amplitude spectrum N˜ :

After calculating the final output S^ of the network, the weight parameters of the network are trained by pure audio signals. The trained objective function consists of two parts shown in Eq. (2): 2J=ω‖ S^−S ‖22+v‖ S˜−S ‖22 where ω denotes the weight of the error between S^ and the pure speech signal S, and v denotes the weight of the error between the speech amplitude spectrum S˜ and the pure speech signal S, while guaranteeing ω + v = 1.

After training and validation, the PM-DNN algorithm has good noise reduction performance in both high signal-to-noise ratio and low signal-to-noise ratio noise environments.

In the process of speech recognition, in order to further improve the efficiency of speech recognition and facilitate the subsequent extraction of the characteristic parameters of the speech signal, we need to add a preprocessing step to remove redundancy before recognition. The pre-processing steps mainly include pre-emphasis processing, frame-splitting and windowing processing, and endpoint detection processing. 1)

Pre-emphasis

We use pre-emphasis to enhance the high-frequency part and eliminate the radiation effect. The principle of pre-emphasis is to input the audio signal into the high-pass filter, output the high-frequency component, and finally obtain the enhanced audio signal.

The transfer function of the high-pass filter is shown in equation (3): 3H(z)=1−αz−1

Where, a is the pre-emphasis coefficient, which is generally taken as 0.9~1.0 depending on the specific scenario. The pre-emphasis coefficient in this experiment is taken as 0.965. Defining the audio signal at the nnd moment as φ(n), the computational function of the first-order high-pass filter is shown in the following equation: 4ϕ(n)=φ(n)−αφ(n−1)

Figure 7.

The overlapping frame of the speech signal

The short-time zero crossing rate refers to the number of times a frame of audio signal completes alternating positive and negative values through the horizontal axis. The expression for the short time zero rate Z_n of audio signal x_n (m) is: 8Zn=12∑m=0N−1| sgn[ xn(m) ]−sgn[ xn(m−1) ] | where N denotes the frame length and sgn [] is the sign function whose expression is sgn[ x ]={ 1,x≥00,x<0

After completing the preprocessing step, the feature extraction operation begins. The extraction and screening of feature parameters is an important part of performing speaker identification. The extraction methods of base frequency, resonance peak, and Mel frequency cepstrum coefficient are described in detail below. 1)

Extracting the base frequency

In this paper, the cepstrum method is used to extract the keynote frequency, and the definition equation of the cepstrum is as follows: 9c(n)=12π∫−ππln| z(s(n)) |ejwndw

The process of extracting gene frequencies is shown in Figure 8.

Figure 8.

Process of extracting gene frequency

Figure 9.

The process of extracting the resonance peak

Figure 10.

MEL’s descent spectrum extraction process

The formula for converting the actual frequency of the input audio signal to the Mel frequency is as follows: 10Mel(f)=2595*log10(1+f700) where f denotes the actual frequency of the input signal.

After calculating the Mel frequency, the discrete Fourier transform is performed on the audio signal to facilitate the observation of the energy distribution of the audio signal. The larger the energy spectrum, the more eigenvalues required to recover the signal I. Let the speech time domain signal function be q(n), then the spectrum function Q(k) is calculated as: 11Q(k)=DFT(q(n))

The energy spectrum is filtered using a triangular filter bank, which reduces the amount of computation. At the same time, the resonance phenomenon is more pronounced, reducing the influence of other waveforms on the audio signal and making the eigenvalues easier to obtain, assuming that there is M triangular filter, and the range of values of M is 22, 26, and they are used to merge into a filter bank with the center frequency set to f(m), and m is taken as (1,2,⋯,M).

The frequency response of the triangle filter is expressed as follows: 12Hm(k)={ 0,k<f(m−1)2(k−f(m−1))(f(m+1)−f(m−1))(f(m)−f(m−1)),f(m−1)≤k≤f(m)2(f(m+1)−k)(f(m+1)−f(m−1))(f(m)−f(m−1)),f(m)≤k≤f(m+1)0,k>f(m+1) where ∑m=0M−1Hm(k)=1

After processing using the triangular filter bank, the logarithmic energy of each triangular filter is calculated using the following equation: 13s(m)=ln(∑m=0M−1| Xa(k) |2Hm(k)),0≤m≤M

After calculating the logarithmic energy, the Mel frequency cepstrum coefficients are obtained by calculating the inverse discrete cos -function transform values: 14Cn=∑k=1Mcos(πn(k−0.5)M)ln(s(m)),n=1,2,…,N

Where N represents the order of Mel frequency cepstrum coefficient, the value range is 12~16.

Because the first-order Mel frequency cepstrum coefficient can express limited characteristics, it cannot fully represent the characteristics of the original audio signal, so the Mel frequency cepstrum coefficient is actually only suitable for processing static speech. The formula for the difference parameter is shown in equation (15): 15dt={ Ct+1−Ct,t<Q∑q=1Qq(Ct+q−Ct−q)2∑q=1Qq2Ct−Ct−1,t≥P−QOther where Q is used to denote the time difference of the first order derivative and P refers to the order of the cepstrum coefficients. C_t denotes the tth cepstrum coefficient, and d_t represents the t th order difference parameter.

Convolutional neural networks not only have the advantages of traditional neural networks, such as adaptability and fault tolerance, but also have the advantages of automatic feature extraction and weight sharing. Convolutional neural networks calculate the output value through forward propagation, and the bias and weights in the calculation process can be adjusted through backward propagation. Convolutional and pooling layers are the most basic structures of convolutional neural networks, and at the same time, all convolutional neural network models are improved on this basic structure.

The two phases of feature extraction and feature mapping constitute the convolutional layer. In the feature extraction stage, the feature maps of each layer of each neuron are interconnected, and the convolutional filter is used to perform the convolution operation to extract that local feature. The feature extraction stage completes the extraction of local features and then feature mapping is required. This is generally mapped to a numerical value using an activation function. The most commonly used activation function is the Sigmoid function, which is derived from the type S curve in biology, and its function is to introduce nonlinearity and map the input value to a range from 0 to 1. The Sigmoid function is shown in (16) below. 16f(x)=1/(1+ex)

Direct classification or regression prediction after calculation through the convolutional layer is prone to overfitting, and a pooling layer is generally introduced to reduce the probability of overfitting. The purpose of the pooling layer is to collect the main feature information, which can reduce the dimensionality of the training data to reduce the probability of overfitting, but also to enhance the network’s adaptability to changes in the size of the feature data. The general form of pooling layer is shown below: 17X=f(down(X)*w+b) where down(x) denotes the subsampling function, which generally sums and averages all the pixel points of different input data n*n. The subsampling layer makes the output data shrink n*n while each calculation corresponds to a weight w and a bias b of its own.

LSTM recurrent neural network has a processing device called cell, cell is used to determine whether the learning results are useful for the overall results. cell contains three gate units, input gate, output gate and forget gate, which are used to filter the computational results, the structure of the gate unit of the LSTM is shown in Fig. 11, the cell will be filtered by the forget gate, which is mainly used to filter out the LSTM recurrent neural network each operation does not meet the conditions of the computational results, and only retain the beneficial results of the LSTM recurrent neural network. Current research shows that LSTM recurrent neural network can solve the long order dependence problem, and the generalization of this recurrent neural network is very high.

Figure 11.

LSTM unit structure

Suppose h_t−1 is the output value of the LSTM recurrent neural network at moment t–1, c_t−1 is the state of the gating unit at moment t–1, and x_t is the sequence input value at moment t. h_t and c_t are the output value of the LSTM recurrent neural network and the state of the gating unit at moment t, respectively. Then the forgetting gate, input gate and output gate in the LSTM recurrent neural network at this time are calculated as follows: 18ft=θ(Wf⋅ht−1+Wf⋅xt+qf) 19it=θ(Wi⋅ht−1+Wi⋅xt+qi) 20ot=θ(Wo⋅ht−1+Wo⋅xt+qo)

Where f_t, W_f and q_f are the result of forgetting gate calculation, weight matrix and bias term respectively. i_t, W_i and q_i are the input gate calculation result, weight matrix and bias term respectively. i_o, W_o and q_o are the input gate calculation results, weight matrix and bias terms, respectively. The final output of the LSTM is determined by the output results of the output gates and the state of the cell together, as shown in (21) to (23) below. 21ct˜=tanh(Wc⋅ht−1+Wc⋅xt+qc) 22ct=ft∘ct−1+it∘c˜t 23ht=ot∘tanh(ct)

Where c denotes the input cell state at moment t. W_c denotes the input unit state weight matrix. q_c denotes the input unit state bias term. tanh denotes tanh the activation function. Denotes multiplication by elements.

The introduction of the attention mechanism can improve the structural deficiencies in the design of recurrent neural network models. The attention mechanism is a model proposed by Treisman and Gelade to simulate the attention mechanism of the human brain, which can be simply understood as a combinatorial function by calculating the probability distribution of attention to highlight the effect of a key input on the output. The attention mechanism considers all inputs, but instead of giving the same weight to each input, it pays more attention to certain specific inputs.

In this section, 28 video cases of ideological and political education classes in a university are selected for analysis. In the ideological and political education in colleges and universities, the basic information statistics of class 1, class 2, class 3 and class 4 ideological and political video class cases are shown in Table 3. From the table, it can be seen that the average value of the conversion rate of teacher-student interaction behavior in the ideological and political classroom of colleges and universities is 0.3462, the average value of the conversion rate of teacher-student interaction behavior in the classroom of class 1 is 0.3643, the average value of the conversion rate of teacher-student interaction behavior in the classroom of class 2 is 0.3371, the average value of the conversion rate of teacher-student interaction behavior in the classroom of class 3 is 0.3371, and the average value of the conversion rate of teacher-student interaction behavior in the classroom of class 4 is 0.3471. It can be concluded that the conversion rate of classroom teacher-student interaction behavior is the highest in classroom 1.

Among the 28 video lesson examples selected, interactive lesson examples using the method proposed in this paper totaled 71%. It can be seen that in more than half of the actual teaching, the proportion of teacher-student activities is comparable, with more teacher-student interactions and more time for teachers to guide and comment on learning outcomes. The teaching concept of “teacher-led and student-led” advocated in the quality lessons is reflected. Through the leading role of the teacher to cultivate students’ independent learning spirit and realize classroom interaction. Through the use of the methods proposed in this paper for ideological and political education in colleges and universities, not only can cultivate students’ interest in learning, but also increase students’ confidence in learning, so as to achieve the spirit of mutual cooperation and to play the role of the main body, so as to truly achieve the students are the masters of learning.

According to the coding table of teacher-student interaction behaviors in the ideological and political education classroom in colleges and universities, coding was completed for the teacher-student interaction behaviors of the four classes, and the data of Class 1, Class 2, Class 3, and Class 4, which had been coded in NVivo software, were imported into GSEQ (the software for analyzing the teacherstudent interaction behaviors) according to the time series for the lagged sequence analysis, in which behavioral conversion sequences with Z-scores greater than 3 were selected from the residual table to show the significance level of the sequences. With the help of Gephi software, a schematic diagram of teacher-student interaction behavioral transition relationship was drawn, which more intuitively clarified the relationship between the behavioral sequences, analyzed the significant behavioral sequences of classroom teacher-student interactions, summarized the characteristics of teacherstudent interactions in ideological and political education classrooms in colleges and universities, and provided more useful references to the professional development of teachers and the level of classroom interactions.

1 class of ideological and political education classroom teacher-student interaction behavior by ideological and political education classroom teacher-student interaction behavior coding system coding after data processing, transformation, the use of software to analyze the data, generate behavioral conversion table after the formation of the adjusted residual table in Excel. Higher education ideological and political education classroom teacher-student interaction behavior conversion adjusted residuals as shown in Table 4 (Table 1-14 is the code names of 14 students).

From the table, it can be seen that the interactive behavioral transitions between teachers and students were mainly focused on teacher instructions (6) and teacher feedback (4).The most intensive behavior in the ideological and political classroom of class 1 during the occurrence of the interaction was teacher instructions (6). For example, after the teacher asked a question, he/she issued a behavior that instructed the students (1→6, Z=5.87, p<0.05). Teachers give praise or encouragement for students’ behavior, students feel a sense of achievement and confidence (4→3, Z=4.85, P<0.05). From the results of the analysis, it can be seen that the teacher-student interaction behavior is mainly based on question-response-feedback, and more interaction occurs in verbal behavior between the teacher and student.

The teacher-student behavior transition relationship in the ideological and political classroom of Class A is shown in Figure 19. The figure indicates that the majority of the continuity sequence occurs between teacher and student behavior, and the continuity relationship of student behavior is significantly lower than that of teacher behavior. It shows that the interaction in the ideological and political education classroom teaching in colleges and universities is carried out by teachers’ organization and guidance.

Figure 19.

Class A class ideological and political class