VR-based computer maintenance practical training platform development design and application research

The controller in the NTM is equivalent to the CPU in the computer, its external memory is equivalent to the memory of the computer, and its read/write terminal is equivalent to the IO device of the computer. The controller modifies and reads the memory block through the read/write terminal. During the operation of the computer, the CPU addresses the data information according to the control signal from the controller and determines the location in the memory to read and write the data information. Unlike the actual machine, there is no derivable concept for computer operations on memory, whereas in the NTM, all read and write operations on the memory block matrix are derivable. k vector is the weight vector to be calculated, the result of NTM addressing satisfies the normalisation condition of Eq. (1), and the result of reading on the external memory matrix satisfies the relationship of Eq. (2).

1 ∑i=1i=mki=1 2 r=∑i=1mki∗vi

The addressing of the NTM, that is, the process of calculating the weights corresponding to all the vectors in the external matrix, includes addressing by content and addressing by position. Addressing by content requires first calculating the similarity between the target vectors given by the controller and all the vectors in the external matrix of the NTM, while addressing by position makes some adjustments to the results of the addressing by the content process according to their position coordinates in the memory matrix. Addressing by content and addressing by position jointly realise the addressing of the NTM and form the effect of attention mechanism. The general addressing formula of the NTM is as follows: 3 wtc(i)=exp⁡(βt∗K[kt,Mt(i)])∑j=0N−1exp(βt∗K[kt,Mt(j)])

The key for content addressing part of the NTM is very similar to the role of the attention mechanism. Addressing by content requires a normalised weight value for all vectors in the memory, and when reading and modifying the external memory matrix in the NTM, it needs to be modified to varying degrees depending on the result of the addressing. This process attention mechanism is very similar. In the addressing process of the NTM, the output vector k of the controller needs to make a similarity judgment with each vector in the memory matrix. The general implementation method is to calculate the distance between the vector and each vector in the memory matrix as the similarity result. This similarity is calculated by the following Eq. (4): 4 k[u,v]=u∗v‖u‖∗‖v‖

The part of computing similarity can be computed using the Luong attention mechanism, which is widely used in sequence applications and has achieved effective enhancements in several areas of natural language processing. This attention mechanism belongs to one of the implementations of the global attention mechanism, in which the hidden state vector of the target is compared with all the output states of the encoder during the computation of the intermediate semantic vector, and a weight vector of length equal to the time step of the encoder is computed. The weights of this attention are calculated as given in the following Eq. (5): 5 at(s)=exp⁡(score(ht,h^s))∑s′exp(score(ht,h^s′))

There are many ways to calculate the score here, and the current mainstream calculation is given in the following Eq. (6): 6 score(ht,h^s)={htTh^sdothtTWah^sgeneralvaTtanh⁡(Wa[ht,h^s])concat

In this study, the general approach is mainly used to calculate similarity. The corresponding attention weight calculation method, in the NTM improved the addressing mechanism, corresponding to the content-based addressing intermediate vector wtc calculation method as given in the following Eq. (7): 7 wtc(s)=exp⁡(βt∗score(ht,h^s))∑s′exp(βt∗score(ht,h^s′))

In the decoder part enhanced with the external memory matrix, after addressing by content, it is still necessary to use addressing by position to further obtain more accurate different weights of each vector in the external memory matrix. This part mainly includes three processes: linear interpolation, transfer and sharpening, which are calculated as follows: 8 wtg=gtwtc+(1−gt)wt−1 9 w^t(s)=∑j=0N−1wtg(j)st(s−j)

The output of the NTM controller controls the workflow of the entire NTM-VR model. The controller is implemented as a neural network, which means it can be a recurrent neural network or a fully connected or convolutional network, and the neural network controller interacts with the whole system input and output. The read and write sides of the NTM-VR model calculate the weights of each vector in the external memory matrix in the current state based on the control signals from the controller and then perform read and write operations on the external memory based on the weights. The NTM-VR model needs to calculate a set of weight coefficients for both the read and write operations of the external memory matrix, and this process of finding the weight coefficients is the addressing process of the NTM-VR model.

10 wt(s)=w^t(s)γt∑j=0N−1w^t(j)γt

The moment in the previous formula is t, and w_t is the vector of addressing results. Eq. (8) represents a linear interpolation with the result of the previous addressing, where gt∈[0,1] . Eq. (9) represents an one-dimensional convolution, that is, to change the addressing process may produce the offset. Eq. (10) represents the sharpening process, which can give more weight to some important positions in the addressing result, thus making the variance of the weight distribution larger, where γt≥1 .

The decoder is another NTM-VR model parameter module with weights separate from the NTM-VR model. The advantage of this model is that the input data and output data are in different data distributions, and their respective data distributions can be represented by their respective NTM-VR model neurons separately, and the network is easier to converge and can achieve better training results, but the total number of weights is twice that of the encoder or decoder when the dimensions of their hidden states are the same, which may affect the working efficiency of the system in some scenarios.

11 hienc=RNN(Wenc,Xi) 12 hidec=RNN(Wdec,Yi−1,hiattn)

This subsection shares the weights of the decoder and the encoder, as in Eq. (13), so that the parameters of the whole NTM-VR model are reduced by half by sharing the weights of the NTM-VR model, but the processes of encoding and decoding in the Seq2Seq algorithm remain unchanged.

13 Wenc=Wdec

After merging the weights, the model eliminates the attention layer in Seq2Seq and provides a new attention mechanism based on the NTM mechanism for a model with shared encoder and decoder weights. The NTM exists an m-row and n-column memory matrix externally. In the encoding phase, the elements of the matrix in the NTM are modified by encoding generated control information to achieve the role of generating memory, and in the decoding phase, the elements of the matrix in the current NTM are read and modified by the control information generated by decoding. Decoding modifies the memory content while also getting the memory generated from the encoding and decoding stages, and the formula of the process of generating attention mechanism through the NTM in the decoding stage is given in Eq. (14).

14 hiattn=NTM(Mm∗n,hidec)

Each weight in the neural network is a parameter value of the floating point type during the participation in the calculation, and the gradient of the quantised value is calculated by the reverse transfer of the gradient. But at this time, this gradient value is used to update the previously saved up floating point type of parameter value. The process of the reverse transfer of gradient is shown in Eq. (15).

15 gw=gwb

Following this process, both forward inference and gradient back propagation of the model can be completed. When the model is fully trained, a final quantisation operation can be performed using the latest floating point parameters to form the quantised neural network and weights with the following Eqs (16) and (17): 16 xint=round(xS)+z 17 xQ=clamp(0,Nlevel−1,xint)

The replacement of value types in the quantisation process of the network model requires a uniform transformation of the original range of values into a new range of integer-type values, and there is a conversion process of scales and zeros in this process. The quantisation process is shown in Eq. (16), where s represents the scale, which is the range of variation of floating point numbers corresponding to one unit of the integer type, and z represents the zero point in the quantisation process, the threshold range before the quantisation of the weight value, which is a symmetric range about the z value. This symmetrical range can be simplified by the following Eq. (18): 18 clamp(a,b,x)={ax≤axa≤x≤bbx≥b

The corresponding process of finding the floating point number corresponding to each quantised value during the training process is called inverse quantisation, in which can be carried out using the following Eq. (19): 19 xfloat=(xQ−z)s

The quantisation method using zeros makes the quantised results symmetrically distributed, but the introduction of zeros increases the computational effort in the quantisation process exponentially. The design can be quantised using a quantisation process with a fixed value of zero. This quantisation process can be simplified using the following Eqs (20) and (21): 20 xint=round(xs) 21 xQ=clamp(0,Nlevel−1,xint)

This fixed zero point corresponds to the inverse quantisation, as given in the following Eq. (22): 22 xfloat=(xQ−z)s

For the sequence generation model and sequence classification model, the quantisation method in this subsection is used to compress the weights of the NTM-VR model in the feature extraction network, which can further obtain a more lightweight network structure.

In the current computer maintenance training platform teaching in colleges and universities, the teaching and teaching content of the practical training courses as well as the actual social work cannot meet the needs of course teaching and actual social work mainly because the computer network courses emphasise theoretical teaching and ignore the practical training, and because of the low utilisation rate of laboratory equipment. The maintenance training platform teaching content is too old, which is away from the actual needs of society. The content of experimental training cannot be perfectly combined with the theoretical course, and the mode of practical training is too single. Therefore, it is urgent to build a reasonable structure and perfect function of the practical training teaching platform to meet the needs of practical training and better cultivate network technology talents in line with the needs of society. And the use of traditional computer network experimental rooms to carry out network experiments, often using direct plugging and unplugging line connection equipment experimental methods, has several major drawbacks: first of all, a lot of students’ experimental time is wasted in plugging and unplugging lines, rather than in the experimental content, which results in poor learning. Secondly, a large number of students plug and unplug lines at the same time; the scene is chaotic; the teacher’s energy is wasted in maintaining order in the classroom, so they cannot concentrate on counselling students to perform experiments; thus, the teaching effect is poor. Thirdly, the loss of equipment in the experimental process is very serious. The use of simulation technology in the teaching of the computer maintenance training platform can effectively avoid the aforementioned problems.

Therefore, factoring in the needs of teachers and students and the existing training platform as a parameter, we used the NTM-VR model for simulation analysis. The analysis results are shown in Figure 1. Analysis of Figure 1 shows that students’ acceptance of the VR training platform is 67.4%, which is much higher than 32.6% acceptance for the traditional platform; and teachers’ acceptance of the VR training platform is as high as 71.5%. The calculation results show that the VR technology-based computer maintenance practical training platform can not only connect the real equipment integrated network practical training environment into the Internet but also connect each laboratory bench unit of the practical training platform. With the open design, students can break through the limitation of time and space to carry out practical training exercises using this module. Students can configure the equipment in the experimental environment by simply logging onto the access controller of the practical training system through the Internet and entering the name of the corresponding equipment to access the unit for experiments, which greatly enhances students’ participation and improves teachers’ teaching quality.

Fig. 1

The teaching resource library is designed to categorise scattered and messy teaching resources, store them in an orderly manner, and provide students and teachers with uploading, retrieving and downloading functions so that students can obtain teaching resources and practice topics conveniently through the Internet when they need them and realise the sharing of teaching resources. This will change the traditional teaching concept, teaching style and learning style, combining task-based learning and project-based teaching methods, turning students from being indoctrinated in learning to being inquirers in learning, highlighting the position of students as the main body and centre in the learning process. At the same time, the use of teaching resources library can also realise fragmented learning, students can carry out independent learning when they have free time, choose teaching resources according to their needs, selectively choose learning resources according to their learning situation and learning progress and realise personalised learning. Therefore, the learning intensity, learning outcomes and practical training platform approach are brought into the NTM-VR model for simulation analysis, and the results are shown in Figure 2. From Figure 2, it can be seen that under the uniform learning intensity, the learning outcome based on the VR computer repair practical training platform is 76%, which is much higher than the learning outcome based on the traditional computer repair practical training platform of 24%. This shows that VR technology connects the practical training platform to the real virtual, and at the same time, it can make the computer-related parts real in front of students so that they can master computer repair skills more quickly. Under the unified learning outcome, the learning intensity of the VR-based computer repair practical training platform is only 32%, which is much lower than the learning intensity of 68% based on the traditional computer repair practical training platform. This shows that VR technology brought into the classroom can drive students’ motivation and improve the acquisition of students’ learning skills while reducing the learning intensity. The aforementioned data calculation results show that the open computer network implementation platform uses a combination of virtual and real experimental environment and online teaching mode, introduces advanced teaching concepts and teaching modes, builds shareable computer network course practical training teaching resources and creates a teaching platform that fully supports online practical training. It can build an experimental teaching system that unites knowledge and ability. The goal of the computer maintenance training platform is not simply to master the network foundation of computers and practical training tools but also to develop computer network courses and computational thinking skills around ‘computing’.

Fig. 2

According to the aforementioned calculation results and field research, it is known that the computer maintenance practical training platform is developed and designed for the purpose of enriching the types of experiments for students because students can only carry out the networking practice in the network practical training course through the real practical training environment and cannot create a fault scenario. The virtual simulation technology can create a fault scenario for students and exercise their ability to apply their knowledge flexibly. Therefore, in order to make the VR computer maintenance practical, the training platform designed in this study must have the following four major functions for more realistic use:

Update the experimental teaching content, which requires computer network courses in the framework of the complete knowledge and ability of the curriculum system, a comprehensive update of teaching content consistent with the computer subject planning. To explore the ‘‘computational thinking skills, computer network technology application skills, multidisciplinary cross-fertilization’’ of the unity of knowledge and ability training system, and to establish a practical training system to adapt to it, from the training program to fully support the unity of knowledge and ability of the course, the establishment of a new curriculum system in line with the development of the times and social needs.