Construction of science and technology achievement transfer and transformation platform based on deep learning and data mining technology

Abstract

The transfer of scientific and technological achievements is an inevitable stage in the application of science and technology to the process of productivity. This process is accompanied by various influencing factors. How to eliminate the influence of adverse influence factors on the transformation of technology into productivity is crucial to the development of social productive forces. Based on this, from the perspective of deep learning, this study builds a technology transfer transformation platform through deep learning combined with data mining technology and analyzes the method in detail. On this basis, this paper takes a city as an example to analyze the platform of scientific and technological achievements transfer. In addition, by collecting existing data as system input and data mining analysis, this paper summarizes the advantages, disadvantages, opportunities and threats of the city’s enterprises in the transformation of results and proposes corresponding countermeasures. The example verification shows that the method proposed in this study has certain practical effects and can provide theoretical reference for subsequent related research.

Keywords

Deep learning data mining transfer of scientific and technological achievements convolutional neural network system analysis

1 Introduction

Since the 1980 s, with the rapid development of information technology and biotechnology, a large number of scientific research institutions have emerged. At present, developed and developing countries have gradually formed a relatively complete national science and technology system. The United States is a technology superpower, the United Kingdom, France and Germany are traditional European technology powers, Japan, South Korea, Canada and Australia are emerging technological advanced countries, Russia, India and China are in a catch-up position. Most institutions are more specialized national research institutions. However, there are also some research institutions that perform other functions in addition to scientific research, such as the US Department of Energy, the National Aeronautics and Space Administration, the National Research Council of Canada, the Australian Geosciences Bureau, etc., and they are also national government departments [1].

National research institutions are roughly divided into two types. One type of institution is a relatively large, comprehensive and substantive large-scale national research institution, represented by Germany, France, Japan, Russia, and China. The other type of institutions are relatively dispersed institutions. The United States and the United Kingdom are the main representatives. There are many national scientific research institutions with professional and departmental management, which are represented by national laboratories and research councils. In today’s world, national scientific research institutions, together with universities and enterprise R&D institutions, constitute the “troika” that promotes the development of science and technology. Through the interaction of market, talents and scientific and technological achievements, the three parties of production, learning and research form an orderly division of labor and mutual cooperation, which constitutes a complete collaborative innovation value chain [2]. Among them, the company focuses on technological innovation and application of results, and at the same time develops the market. The university focuses on knowledge transfer and high-quality personnel training, while focusing on free and flexible scientific exploration. The national scientific research institutions are facing the national strategic needs, serving the national and local economic development, providing enterprises with continuous high-tech achievements and realizing the transfer and transformation of scientific and technological achievements. At the same time, it plays an important role in research work on national security and public health [3]. It is an important task for national scientific research institutions to carry out technology transfer and technical services, attach importance to strengthening the links between science and technology and the economy, and provide knowledge, technology and talents for national and local scientific and technological progress as well as economic and social development. Experience tells us that by setting up specialized institutions or using technology transfer intermediaries to promote the transfer of scientific research results to enterprises and providing technical support and consulting services for enterprise R&D by undertaking research tasks entrusted by enterprises, cooperating with R&D institutions, and jointly establishing innovation alliances of industry, universities and research institutes are relatively successful choice. For example, the National Laboratory of the US Department of Energy has established a technology transfer office.

Generally speaking, as demand traction gradually becomes the most important driving force for the development of contemporary science and technology, scientific and technological achievements are no longer the product of closed-door construction of research institutes. Due to the involvement of enterprises, it provides a rich imagination and a broad stage for the development of science and technology. Technological innovations are constantly emerging and rapidly applied, making scientific and technological activities break through the barriers of research institutes and universities. The power of the market and the rules of competition have penetrated into every corner of scientific and technological activities. Today, the development of science and technology is more dependent on the support and benign interaction of the business community and requires a solid material foundation as a guarantee.

2 Related work

In the traditional TF. IDF-based topic semantic mining, the typical models are LSI, pLSI and LDA [4]. Bai Y et al. studied the basic problem that the sequence of user search words in the information retrieval matched the sequence of document words. Moreover, the method of improving the precision of retrieving relevant documents through the high-order “semantic structure” implicit in the document is proposed systematically, which has caused the upsurge of topic model research [5]. For the same document retrieval matching problem, Yin H aims to improve the lack of probabilistic interpretation of the final result of the LSI model. While inheriting the advantages of LSI’s automatic document indexing, document dimensionality reduction and construction of semantic space, it uses the potential hierarchical model to provide probabilistic mixed composition decomposition, with the optimization of likelihood function as the result [6]. The LDA proposed by Lou N is a three-layer Bayesian model that can be used for classification, novelty detection, summarization, similarity and correlation judgment [7]. The problem areas it deals with include document modeling, document classification, and collaborative filtering.

At present, in the context of the explosive growth of electronic media data such as network information, deep learning has been expanded in various fields. The reason is that deep learning as a sub-domain of data-intensive machine learning can directly learn from a large amount of data, samples and experience to master the logic and knowledge in the data, and finally let the machine automatically understand the completion of certain tasks [8]. That is to say, deep learning writes rules ahead of the traditional feature-based grammar natural language processing method and implements complex processes by learning the knowledge in the data to achieve the purpose of accomplishing certain tasks. However, unlike the general machine learning architecture, the deep neural network inspired by the human brain nervous system has more hidden layers. An important basis for supporting this phenomenon is that the brain nervous system does have a rich hierarchical structure. A well-known example is the Hubel-Wiesel model of the Nobel Prize in Medicine and Physiology [9], which reveals the mechanism of visual nerves and becomes the source of deep learning. The most famous case of deep learning is the project developed by Liu X [10]. This has caused great shock to the industry, and deep learning application research has been widely carried out. The international data center predicts that the global driving revenue of AI will increase from 8 billion in 2016 to 47 billion in 2020 [11]. At the same time, it is worth noting that Professor AndrewNG was also one of the co-authors of LatentDirichlet Allocation, which laid the foundation for the development of the semantic mining theme model in 2003. It can be seen that the combination of topic models and deep learning has a long history [12]. Another case is the simultaneous interpretation system that Microsoft publicly displayed in November 2012 in Tianjin, China. The English pronunciation input at the front desk and the simultaneous translation into Chinese in the background, this interpretation system based on deep learning has attracted great attention in the context of the shortage of simultaneous interpretation talents [13]. There is also a typical case in which AlphaGo, based on deep learning and intensive learning, defeated South Korea’s nine-part Li Shishi in 2017, which caused great discussion. Specifically, in the semantic mining application of this key field of natural language processing, the research on semantic similarity calculation based on deep learning has been growing in the last five years, partly due to the driving of the annual international semantic evaluation SemEval competition [14]. In the field of agricultural science and technology information, the improvement of semantic mining model needs to be practiced, not only in model derivation, parameter estimation algorithm improvement, but also in the field of large-scale agricultural science and technology document semantic mining distributed parallel computing in the context of big data. In the field of deep learning, the development of artificial intelligence is still moving forward. Combined with the corresponding neural network architecture (It is not limited to the above-mentioned CNN, R_N’Ns, deep learning network DNN, deep trusted network DBN and convolution deep confidence network CDBN, etc.), learning data-specific characterization is applied to solve some practical problems [15]. In an article published in 2015, Yang D et al. elaborated on the theory and model of the 2015 SemEval competition for the semantic similarity calculation of twitter information. In this SemEval competition, the ASOBEK system developed by EyeciogluandKeller to calculate semantic similarity uses the support vector machine (sVM). This classifier has simple real word overlap function and character language model features, and the calculation effect is good [16]. Liu K et al. developed and utilized the MITRE system based on the extended string-matching feature recurrent neural network in this competition and achieved good results. Other players developed a semantic similarity system, using supervised learning models with features including language model overlap, term alignment, edit distance, and sentence embedding cosine cosine similarity calculations, etc., and also achieved good results [17].

In the past few years of deep learning and semantic model development, training deep learning algorithms in large document corpora has achieved almost the same effect as hand-designed algorithms. In 2015, Xu J used by Adrian Sanborn et al. of Stanford University calculated the semantic similarity of two different segments of a text and developed a semantic similarity calculation system to evaluate the semantic similarity of two arbitrary sentences [18]. Moreover, the application of deep learning technology in two major areas is not completely separate. In the past two years, researchers have also tried to use the computer-oriented technology CNN to provide RINI-Ns with many new ideas and experiences in the field of natural language processing, so that it has achieved better results in real-world applications. For CNN technology research, Feng S et al. used a kernel-based algorithm similar to that used in traditional machine learning to train convolutional neural networks. It has finally achieved breakthroughs in performance and accuracy, which is superior to the results of traditional machine learning methods [19]. For the first time, Sprigle S et al. proposed the CDBN (Convolutional Deep Confidence Network) method. The idea is to reduce the error of the total network structure layer by training the entire deep architecture. The final result shows that it has achieved superior results in extracting a small number of features to represent large image objects [20]. In the same year, Kimmons R et al. proposed a method based on convolutional neural network to process serialized data. This method achieves better results for extracting unlabeled video data with respect to naturally occurring time-series correlation information [21]. Unlike the above-mentioned scholars’ supervised methods, Ranzato et al. tried unsupervised training based on convolutional neural networks. It trains multiple sparse features by training nonlinear arithmetic functions, multi-layer convolution filtering, and feature sub-sampling methods. Schematic diagram of CNN convolution process as show in Fig. 1.

Fig. 1

Schematic diagram of CNN convolution process.

3 Deep learning related theory

3.1 Convolutional neural networks

In convolutional neural networks, local connections, parameter sharing, and pooling (also called down-sampling) are generally used to solve the difficult training problems in traditional neural networks. As shown in Fig. 2, the left side is a full connection diagram, and the right side is a partial connection diagram of pixels separated by 2 pixels. If the full connection method and the image have 1000 × 1000 pixels, the number of neurons in the next hidden layer is 10⁶. If full connection is used, there will be 10¹² weight parameters. If a local connection is used instead, each neuron is only connected to a partial image of 10 × 10 size in the image, and the network parameters are reduced to 10⁸. Because deep networks are often used in deep learning, even if neurons use local connections, deep network models can still obtain global features of images under the prominent connections and neural interactions of neurons.

Fig. 2

Schematic diagram of full connection and local connection.

Pooling refers to further compressing the convolutional feature map obtained by convolution, sampling on the convolutional feature map with a certain window and step size and taking the average or maximum value of the adjacent convolutional feature map as the input to the next layer. Taking the commonly used 2 × 2 pooling window as an example, we can select the maximum or average of the adjacent 4 pixels in the convolutional feature map as the output, which is the maximum pooling and averaging pooling. Pooling not only greatly reduces the amount of data in the convolutional feature map, but also reduces the risk of overfitting and retains more key feature information. The convolutional neural network combines the three methods of local connection, weight sharing and pooling, and obtains the displacement, scale and deformation invariance on the original image, and improves the operation speed and precision in image processing.

The first hidden layer of LeNet-5 is the convolutional layer, and each convolution kernel has a size of 5 × 5. The second hidden layer is the pooling layer, which further compresses the convolutional feature map of the upper layer. The third hidden layer is the convolutional layer, which consists of 16 sets of filters. The fourth hidden layer is the pooling layer, which further compresses the data. The fifth layer is a fully connected layer consisting of three full connections, which are finally sorted by a classifier. In actual use, we can flexibly adjust the parameters such as the size and number of the convolutional layer and the pooling layer. Schematic diagram of LeNet-5 structure as show in Fig. 3, Schematic diagram of AlexNet structure as show in Fig. 4.

Fig. 3

Schematic diagram of LeNet-5 structure.

Fig. 4

Schematic diagram of AlexNet structure.

Google Net continues the idea of using deeper networks to enhance the expressive power of the web, thereby increasing the accuracy of the model. A network with too many layers is likely to cause two problems. On the one hand, it will increase the complexity of the model and cause over-fitting. On the other hand, a large number of parameters that need to be trained and optimized will consume a lot of computing resources. As shown in Fig. 5, GoogleNet uses the Inception structure to turn the fully-connected network into a sparse connection. Based on the sparseness of the network structure, GoogleNet takes full advantage of the high computational performance of dense matrices. Since then, there have been improved versions of GoogleNet, V2, V3, and V4. In subsequent experiments, we used Inception V3 as the feature extraction network. As shown in Fig. 6, ResNet directly bypasses the transformation of subsequent layers by adding some jump connections between non-adjacent network layers, so that network learning forms a residual function, which greatly optimizes the training problem of deep networks.

Fig. 5

Schematic diagram of Inception structure.

Fig. 6

Schematic diagram of the residual network structure.

3.2 CNN training process

In the process of forward propagation excitation of convolutional neural networks, it is assumed that the input layer is the first layer and the output layer is the Lth layer. The output of the intermediate hidden layer is as shown in Equations (1) and (2). $u^{l} = W^{l} x^{l - 1} + b^{l}$ (1) $x^{l} = f (u^{l})$ (2)

In the formula, l denotes the first layer, W^l and b^l denote the layer 1 weight matrix and offset parameters, x^l-1 denotes the output of the previous layer, u^l denotes the result of the current layer, and f (·) denotes the activation function used.

For a training set containing class C and N samples, the loss function is defined as shown in Equation (3). $E^{N} = \frac{1}{2} \sum_{n = 1}^{N} \sum k = 1 c (t_{k}^{n} - y_{k}^{n})^{2}$ (3)

Therefore, the training error of a specific single sample is as shown in Equation (4). $E^{n} = \frac{1}{2} \sum_{k = 1}^{c} (t_{k}^{n} - y_{k}^{n}) 2$ (4)

In the process of backpropagation, based on the gradient descent strategy, the parameters of the whole network are adjusted in the negative gradient direction of the data, and the training error of the whole data is continuously transmitted back from the output layer to optimize the parameters of the whole network. The sensitivity is as shown in Equation (5). $\frac{\partial E}{\partial b} = \frac{\partial E}{\partial u} \frac{\partial u}{\partial b} = δ$ (5)

The training error of the output layer is derived for its input, and the sensitivity of the output layer is as shown in Equation (6). $δ^{L} = f^{'} (u^{L}) \circ (y^{n} - t^{n})$ (6)

In the formula, f′ (·) represents the derivative of the activation function, yⁿ and tⁿ represent the model prediction value and the real value of the nth sample, respectively, and the symbol ∘ represents element-by-element multiplication.

The error of the output layer is back-propagated to the remaining hidden layers, and the sensitivity of each layer is as shown in Equation (7). $δ^{l} = {(W^{L + 1})}^{T} δ^{l + 1} \circ f^{'} (u^{l})$ (7)

Therefore, the partial derivative of the training error to the remaining hidden layer parameters is as shown in Equation (8). $\frac{\partial E}{\partial w^{l}} = x^{l - 1} {(δ^{l})}^{T}$ (8)

The update of the weight matrix parameters is as shown in Equation (9). $Δ W^{l} = - η \frac{\partial E}{\partial w^{l}}$ (9)

In the formula, η represents the learning rate. Through continuous training, we can get the loss function with the smallest local error and get the optimal network model.

3.3 LSTM

The Long Short-Term Memory (LSTM) is a special Recurrent Neural Network (RNN), first proposed by Hochreiter & Schmidhuber in 1997. LSTM is suitable for processing data with long delays and long intervals in the sequence, and has achieved great success in speech recognition, text translation, and image description.

The internal structure of the LSTM is shown in Fig. 7. The core of the LSTM is the state of the cell. The information does not decay when the information propagates inside the LSTM. At the same time, the LSTM controls the information-to-cell propagation capability through a unique gate structure. The gate structure of LSTM is implemented by a Sigmoid function, and the output of the Sigmoid function is between 0 and 1, reflecting the degree of retention. In the formula, x^t represents the input of the current cell, and h_t-1 represents the output of the cell at the previous moment. i_t, f_t and o_t represent the input gate, the forgetting gate and the output gate in the cell, respectively. As shown in Equation (13), the cell first generates the candidate vector g_t, and then combines the forgetting gate, the input gate and the state c_t-1 of the cell at the previous moment to generate a new cell state c_t. The forgetting gate and the input gate play a role in screening the previous moment state and current input. Finally, according to the output gate, the current time cell output h_t is determined. $i_{t} = σ (W_{xi} x_{t} + W_{hi} h_{t - 1} + b_{i})$ (10) $f_{t} = σ (W_{xf} x_{t} + W_{hf} h_{t - 1} + b_{f})$ (11) $o_{t} = σ (W_{xo} x_{t} + W_{ho} h_{t - 1} + b_{o})$ (12) $g_{t} = tanh (W_{xc} x_{t} + W_{hc} h_{t - 1} + b_{c})$ (13) $c_{t} = f_{t} ⊙ c_{t - 1} + i_{t} ⊙ g_{t}$ (14) $h_{t} = o_{t} ⊙ tanh (c_{t})$ (15)

Fig. 7

LSTM internal structure.

4 Data feature recognition method based on deep learning

Video has more information than pictures. The spatial scene information contained in a single frame in the technology video is generally referred to as spatial information, and the target motion information carried between adjacent frames is referred to as time information. How to better explore the spatio-temporal information in technology video has always been a hot spot in the field of video behavior recognition research. Deep learning enables end-to-end training directly from raw technology video and eliminates the need to manually design features and provides efficient feature representation for video behavior recognition. Convolutional neural networks can extract image spatial features. Based on the cyclic neural network, deep learning can more effectively model the timing information between video frames. Based on the theory of deep learning, researchers have made a series of progress in the field of video behavior recognition. Existing deep learning-based behavior recognition methods mainly include video frame based recognition method, dual channel based recognition method, LSTM network based recognition method and 3D convolution based recognition method.

4.1 Video frame-based identification method

As shown in Fig. 8, a single video frame is input into the CNN network for identification. The second picture is LateFusion. At intervals, some video frames are skipped, and the features of the sampled frames are fused at the fully connected layer. On the one hand, the amount of information on the network has been improved, and on the other hand, the timing information between video frames has been extended to some extent. The third picture is Early Fusion, which inputs an adjacent video frame into the CNN, greatly retaining the timing information between adjacent frames in the video clip. The fourth picture shows Slow Fusion. This fusion method draws on the advantages of the first two fusion methods. Each layer of convolution pays attention to mining timing information between video frames. The experimental results show that this method of slow fusion can best utilize the spatio-temporal information of video. It achieved the best recognition accuracy in four ways, but the experimental results were much worse than the traditional manual model.

Fig. 8

Schematic diagram of single frame-based behavior recognition.

4.2 Dual channel-based identification method

5 LSTM-based identification method

Video content is rich in timing information, so the convolutional neural network alone cannot fully utilize the time domain information of the video. Although the video single frame-based identification method and the dual channel-based identification method utilize timing information of adjacent video frames to some extent, more time series information is needed for video with large interval and time span. The output of the LSTM is determined by the current input and the output of the previous time. It can represent the sequence information of the sequence and is widely used in processing timing problems, such as image description, speech recognition, document digest, etc. Dual channel-based identification method as show in Fig. 9.

Fig. 9

Dual channel-based identification method.

When using LSTM in practice, the convolutional neural network (Google Net, VGG, etc.) is generally used to extract the spatial features of the image, and then the spatial features are input into the LSTM network according to the original timing to characterize the timing information of the video, which fully utilizes the spatial information and time series information of the video to identify the behavior. Donahue proposes a video behavior recognition architecture that uses CNN and LSTM together. First, the video frame sequence is sent to the CNN, and the extracted sequence space feature is used as the LSTM input.

The average value of each time of the LSTM unit is taken as the final output when predicting the behavior. Sharma introduces an attention mechanism based on LSTM. By dividing the convolutional feature map into k × k regions and scoring these regions, the model captures key parts of the learning video motion, which helps to learn the refined features of the video. When Sharma is actually used, CNN can be used to extract the convolutional features of the image. As shown in Fig. 10, the LSTM cells of the same layer represent timing extensions and can also be modeled using multiple layers of LSTM cells stacked. The top of the network is the output of the LSTM at different times. Generally, the output of the last moment is selected for prediction.

Fig. 10

LSTM-based identification method.

Although the traditional two-dimensional convolution can effectively extract the spatial features of the image, when it is applied to the field of video behavior and the video is used as an independent image extraction feature, a lot of timing information is inevitably lost. For the defect of two-dimensional convolution, J_i expands the two-dimensional convolution to three-dimensional and extracts the video features from the spatial and temporal dimensions by performing 3D convolution to extract the motion information between adjacent frames of the video. At the same time, multiple channels are generated from the input video frame, convolution and subsampling are performed in each channel, and the final feature combines information from all channels. The result of 3D convolution on the TRECVID data set is also superior to frame-based 2D convolution on most tasks, improving the accuracy of video behavior recognition.

6 Case analysis

This study takes a city as an example to analyze the platform of scientific and technological achievements. The city has jurisdiction over three counties, five districts and two parks, with a population of 5.9 million. In order to attract merchants to invest, the city attaches great importance to attracting investment. For this reason, it has set up major projects and the Central Enterprise Investment Promotion Office, which is responsible for the city’s investment promotion work. By the end of 2019, the total industrial output value of the above-scale industries was about 430 billion yuan, the ratio of the three major industries was about 1:4.1:3.4, the total investment in fixed assets was about 206 billion yuan, the total fiscal revenue was 44.4 billion yuan, and the total fiscal expenditure was 54 billion yuan. Moreover, the total volume of imports and exports has been rising year after year.

In recent years, the transformation of scientific and technological achievements has received more and more attention from the government, enterprises, scientific research units, and industry participants. Promoting the transformation of scientific and technological achievements of enterprises and accelerating the innovation and development of enterprises, “relying on the transformation of scientific and technological achievements, accelerating the adjustment of industrial structure, and cultivating new economic growth points” has become the consensus of all walks of life in the city, and the progress achieved is also obvious to all.

Beginning in 2010, in response to the provincial government’s call, in the context of very difficult fiscal expenditures, the SME Innovation Fund, the special fund for the transformation of scientific and technological achievements, and the Science and Technology Entrepreneurship Seed Fund have been established, which are used in conjunction with the provincial results transformation funds. In the 10 years from 2009 to 2019, the annual corporate science and technology funds increased from 1.5 million yuan to 4 million yuan, an increase of 167%. In recent years, the proportion of R&D expenditures to GDP in the whole society has also increased year by year. The city actively organized the declaration of enterprise results transformation projects, a total of 11 projects were approved, and the city received 89 million yuan of province’s unpaid funds. The organization and implementation of these projects not only stimulates the enthusiasm of enterprise innovation, but also strives for independent innovation, obtaining special funds for scientific and technological achievements, and obtaining high returns to become an inexhaustible motive force for enterprise development. They frequently docked universities, research institutes and research and development institutions, recruited scientific and technological talents and research and development achievements, and planned the innovation and industrialization of enterprises according to the criteria for reporting special funds (as shown in Table 1). At the same time, more and more universities and research institutes in and outside the province have begun to take the initiative to connect with the city’s enterprises, actively transforming their scientific and technological achievements or providing technical support for the project. The main indicators of industrial enterprises’ technology are shown in Table 2.

Table 1
R&D investment and investment intensity

Years R&D expenditure (100 million yuan) Percentage of GDP (%)

2014 9.79 0.75

2015 15.4 1.02

2016 20.3 1.19

2017 26.06 1.35

2015 31.04 1.46

Years	R&D expenditure (100 million yuan)	Percentage of GDP (%)
2014	9.79	0.75
2015	15.4	1.02
2016	20.3	1.19
2017	26.06	1.35
2015	31.04	1.46

Table 2

Main indicators of science and technology in industrial enterprises

Area		Unit	2014	2015	2016	2017	2018
Number of enterprises		Pieces	1755.38	2258.36	2615.9	2542.17	2590.65
Total R&D personnel		person	4005	7297	7960	9302	11217
Grouped by source of funds	Government funds	Ten thousand yuan	2059.39	2005.86	2439.15	2958.29	2593.68
	Corporate funds	Ten thousand yuan	93345.21	147204.5	183464.5	236200.6	284049.37
	Overseas funds	Ten thousand yuan	332.29	–	282.8	–	–
	Other funds	Ten thousand yuan	1341.28	1151.4	2056.36	1932.13	2467.43
Number of institutions		Pieces	187.86	651.45	723.16	1045.35	1067.57
Total staff of the organization		per-son	4478	10828	11717	15004	14419
Ph.D.		per-son	222	297	369	367	348
Master’s degree		per-son	444	1173	1214	1610	1911

The patent achievements of high-tech enterprises often have distinctive characteristics of the post-industrial era, which is a powerful proof of the effective transformation of scientific and technological achievements. These patents have the characteristics of high technical content, low resource consumption and low environmental carrying capacity, and their contribution rate to the transformation of scientific and technological achievements is relatively high. Table 3 shows the patent application statistics of high-tech enterprises. The classification of patents in each region is shown in Table 4.

Table 3

Statistics on patent applications of high-tech enterprises

County (district)	Patent filing amount	Effective patent filing amount	Average patent filing amount	Average effective patent filing amount
1	2977	1578	61	32
2	1228	657	40	21
3	644	476	24	18
4	780	461	22	13
5	496	291	23	13
6	251	202	17	13
7	35	27	35	27
8	8	8	8	8
9	12	1l	6	6
Total	643I	3711	35	20

Table 4

Table of Patent Classifications in each Region

County (district)	Patent of invention	Patent of utility model	Design patent
1	232	650	2095
2	216	455	557
3	138	460	46
4	218	366	196
5	168	219	109
6	29	212	2095
7	14	21	O
8	O	8	O
9	1	11	0
Total	1001	2362	3013

Based on the above data analysis, this paper combines the research model to analyze, get related problems, and get the corresponding strategy.

7 Analysis of data mining results

Through data mining analysis, it summarizes the advantages, disadvantages, opportunities and threats of the city’s enterprises in the transformation of results. Advantages include the deepening of industry-university-research cooperation, the strengthening of talent pools, the growth of research and development institutions, and the increase in financial support. Disadvantages include small scale of enterprises, weak sense of innovation, backward management model, and inadequate system. Opportunities include the enhancement of government service awareness, the gradual improvement of incentive measures, the gradual establishment and improvement of intermediaries, and the initial results of spillover benefits. Threats include inefficient use of funds, increased complexity of research and development results, unpredictable market conditions, and low levels of sustainability. In the process of implementing the transformation of scientific and technological achievements, if a good environment for promoting the transformation of results emerges, such as a sound science and technology policy system, increased investment in science and technology, and enhanced cooperation and exchanges, the advantages that enterprises have can be magnified, which is conducive to the transformation of results.

As a industry with certain industrial base and certain comparative advantages, high-tech industry is the pillar industry that promotes the city’s economic development and leads the transformation of scientific and technological achievements of enterprises in the future. Functional materials are the basic industries for economic development. In terms of enterprise development, the establishment of a “double new and one special” industrial system is of great significance, and it should adhere to the development requirements of “industrial strong city and industrial industrialization”. The intellectual property awareness and ability of high-tech enterprises is the new normal of economic development, and it is a direct reflection of accelerating the transformation of development mode, accelerating industrial transformation and upgrading, and improving the efficiency of scientific and technological achievements transformation, and directly affect the implementation and effectiveness of the city’s innovation-driven transformation development strategy.

The special fund for the transformation of scientific and technological achievements is to solve the problem of transformation of achievements and is an important basis for mobilizing the enthusiasm of independent innovation, enhancing innovation capability and cultivating strategic emerging industries. In order to improve the efficiency of the use of special funds for the transformation of scientific and technological achievements of enterprises in the city, it is necessary to improve the effectiveness of project organization, create a good project implementation scope, and improve the organization level of the project as a breakthrough.

Small and medium-sized enterprises are the largest industrial enterprises in the city, and they are the foundation of industry. They have the characteristics of large amount, scattered distribution, small scale, high enthusiasm for independent research and development, and urgent need to implement scientific and technological achievements. It is an important participant in the transformation of scientific and technological achievements of enterprises in the city. Increasing support for small and medium-sized enterprises, helping small and medium-sized enterprises to grow rapidly and increasing their contribution rate to the transformation of scientific and technological achievements is an important way to strengthen scientific and technological work and promote the transformation of results under the new situation.

8 Conclusion

The investment in science and technology has become more diversified, and capital is entering the field related to the development of industrial technology, which has brought about major changes in the operation and management mode of scientific and technological activities. Promoting the combination of production, education and research, and realizing the transfer and transformation of scientific and technological achievements has become the only way for countries around the world to develop science and technology and economy. This paper expounds the importance of the transformation of scientific and technological achievements of enterprises to economic and social development, enumerates and introduces the relevant literature, and builds a model of scientific and technological achievements transformation based on deep learning and data mining. Moreover, this paper discusses the factors affecting the transformation of scientific and technological achievements of enterprises. Through the above introduction and comments, it analyzes the status quo and problems of the transformation of scientific and technological achievements of case city enterprises, and deeply analyzes the causes of these problems. In addition, this paper proposes countermeasures and suggestions for solving the problem of the transformation of scientific and technological achievements of enterprises, so that in the future, when formulating policies and measures for the transformation of scientific and technological achievements, the problem orientation can be more prominent, and the transformation of scientific and technological achievements can better serve the economic and social development of the city.

Footnotes

Acknowledgment

Sichuan Science and Technology Achievement Conversion Platform Project, Platform for Transforming Scientific and Technological Achievements and Promoting New Technologies in Petroleum and Petrochemical Industry (NO:2019ZHCG0010).

References

Lee

E.B.

, Kim

and Lee

S.G.

, Predicting customer churn in mobile industry using data mining technology[J], Industrial Management & Data Systems117(1) (2017), 90–109.

Jin

, Wang

and Zeng

, Application of Data Mining Technology in Financial Risk Analysis[J], Wireless Personal Communications (2018).

Liu

, Wang

and Hong

, Wearing comfort analysis from aspect of numerical garment pressure using 3D virtual-reality and data mining technology[J], International Journal of Clothing Science and Technology29(2) (2017), 166–179.

Wei

, Application of Data Mining Technology Based on Wireless Sensor Networks in Oceanographic Forecasting[J], International Journal of Online Engineering (iJOE)14(04) (2018), 137.

Bai

, Data cleansing method of talent management data in wireless sensor network based on data mining technology[J], EURASIP Journal on Wireless Communications and Networking2019(1) (2019).

Yin

, Computer Network Virus Defense Technology Based on Data Mining Technology[J], IOP Conference Series Materials Science and Engineering (2018), 382.

Lou

, Spatial-temporal dynamics of rainstorm in East China based on spatial-temporal datamining technology[J], Chinese Journal of Applied Ecology (2017).

Liu

, The Application of Data Mining Technology in the Teaching Evaluation in Colleges and Universities[J], Journal of Computational and Theoretical Nanoscience (2017), 14.

Chen

and He

, IEEE 2017 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) - Kansas City, MO (2017.11.13-2017.11.16) 2017, IEEE International Conference on Bioinformatics and Biomedicine (BIBM) - Application of data mining technology in TCM diagnosis and treatment[C]// IEEE International Conference on Bioinformatics & Biomedicine. IEEE Computer Society2017 (2017), 1350–1353.

10.

Liu

, Li

, Wei

, et al., Seismic vulnerability assessment at urban scale using data mining and GI Science technology: application to Urumqi (China)[J], Geomatics Natural Hazards and Risk10(1) (2019), 958–985.

11.

Angeli

, Howard

, Ma

, et al., Data mining in educational technology classroom research: Can it make a contribution?[J], Computers & Education2017, S0360131517301264.

12.

Triulzi

, Alstott

and Magee

C.L.

, Estimating Technology Performance Improvement Rates by Mining Patent Data[J], SSRN Electronic Journal (2018).

13.

Hou

, Guo

and Nevin

, Research and Prospect of Multimedia Information DataMining[J], Recent Patents on Computer Science (2017).

14.

Feng

, Li

, Zhang

, et al., Proactive Radio Resource Optimization with Margin Prediction: A Data Mining Approach[J], IEEE Transactions on Vehicular Technology2017, 1–1.

15.

Sorour

, Goda

and Mine

, Comment Data Mining to Estimate Student Performance Considering Consecutive Lessons[J], Educational Technology & Society20(1) (2017), 73–86.

16.

Yang

, Application of Data Mining Technology in the Subject Tactical Teaching of Badminton[J], International Journal of Emerging Technologies in Learning (iJET)13(07) (2018).

17.

Liu

, Garment Wearing Comfort Analysis Using Data Mining Technology[M]// Artificial Intelligence for Fashion Industry in the Big Data Era. (2018).

18.

and Liu

, CET-4 score analysis based on data mining technology[J], Cluster Computing2018(5), 1–11.

19.

Feng

, Lian-Zhong

, Yi-Fan

, et al., A method of evaluating diesel engine performance by using data mining technology[J], Journal of Dalian Maritime University43(3) (2017), 83–88.

20.

Sprigle

and Johnson Taylor

, Data-mining analysis of the provision of mobility devices in the United States with emphasis on complex rehab technology[J], Assistive Technology2017 (2017), 10400435.2017.1402391.

21.

Kimmons

and Veletsianos

, Public Internet Data Mining Methods in Instructional Design, Educational Technology, and Online Learning Research[J], Tech Trends62(5) (2018), 492–500.