CN113177733A

CN113177733A - Medium and small micro-enterprise data modeling method and system based on convolutional neural network

Info

Publication number: CN113177733A
Application number: CN202110554758.XA
Authority: CN
Inventors: 王鑫; 王莹; 陈进东; 张健; 曹丽娜
Original assignee: Beijing Information Science and Technology University
Current assignee: Beijing Information Science and Technology University
Priority date: 2021-05-20
Filing date: 2021-05-20
Publication date: 2021-07-27
Anticipated expiration: 2041-05-20
Also published as: CN113177733B

Abstract

The invention discloses a medium and small micro enterprise data modeling method and system based on a convolutional neural network, the invention is a specific application of the convolutional neural network in the field of medium and small micro enterprise information processing, the application range of a deep learning algorithm is further expanded, the method has higher pertinence and practicability, time sequence information (such as financial information and the like) and non-time sequence information (such as enterprise information and the like) are used for enterprise information, the innovation is higher in the specific application of the convolutional neural network, and the specific method and the flow for extracting and fusing time sequence information and non-time sequence information characteristics are explained through an application scene of credit risk identification.

Description

Medium and small micro-enterprise data modeling method and system based on convolutional neural network

Technical Field

The invention belongs to the technical field of enterprise data processing, and particularly relates to a medium and small enterprise data modeling method and system based on a convolutional neural network.

Background

The data of the enterprise comprises data of time series characteristics and data of non-time series characteristics, wherein the data of the time series characteristics is mainly financial data, the data of the non-time series characteristics is mainly non-financial data, the financial data is an important index for measuring the repayment capacity, the development capacity, the profit capacity and the operation capacity of the enterprise, the non-financial data comprises enterprise information such as registered capital, established years, education background, working years of enterprise owners and the like, and the enterprise information can reflect the strength of the enterprise to a certain extent.

Under the background of big data, the financial data of the internet is increasingly complicated, a comprehensive and objective online data acquisition system is constructed by using basic information and dynamic financial indexes of small and medium enterprises acquired on the internet, and the complex financial data with time sequence characteristics and the enterprise information with non-time sequence characteristics can be effectively processed by using a deep learning method. The convolutional neural network is used as a deep learning method, and can be used for deeply and accurately judging financial data with time series characteristics and complex financial data. In recent years, convolutional neural networks have been introduced into the internet, financial field, to assess personal credit, customer credit and enterprise credit risk. The convolutional neural network can not only automatically extract features from data, but also has strong learning capability in a network structure. Through the adjustment to the network structure, the convolutional neural network not only can carry out deep learning to the data that have the time series characteristic, carries out accurate predictive analysis, can also carry out the analysis to non-financial data such as enterprise information.

Aiming at the problems that a data processing model for small and medium-sized micro enterprises in the prior art is inaccurate, high in cost, complex and low in timeliness, a convolutional neural network is introduced, and a data modeling method for the small and medium-sized micro enterprises based on the convolutional neural network is provided by combining the convolutional neural network with a processing method for time series data and non-time series data.

Disclosure of Invention

The invention aims to provide a medium and small micro enterprise data modeling method and system based on a convolutional neural network, so that the problems of high cost, high complexity and low timeliness of enterprise credit risk evaluation are solved, the processed data is more accurate, and the method and system have more important guiding significance for measuring main indexes of enterprises.

In order to achieve the purpose, the invention adopts the following technical scheme:

a convolutional neural network-based medium and small micro enterprise data modeling method comprises the following steps:

(1) collecting related data of small and medium-sized micro-enterprises;

(2) preprocessing the acquired data;

(3) constructing a convolutional neural network model;

(4) and training and testing the constructed convolutional neural network model, and optimizing and determining parameters of the convolutional neural network model.

Preferably, the preprocessing in step (2) includes missing value processing and data normalization processing. The missing value processing is to fill missing values in the downloaded data set, namely, the current missing value is filled by using the global average value of the index sequence to obtain a complete data set; the data normalization process employs a z-score normalization process.

The method for training and testing the constructed convolutional neural network model in the step (4) comprises the following steps: randomly dividing the acquired enterprise data into a training set and a testing set, performing iterative training on the convolutional neural network model by taking the training set data as input, testing the model by using the testing set, and adjusting parameters according to the testing result.

The convolutional neural network model comprises two parallel sub-convolutional neural networks which are respectively used for receiving data of time series characteristics and data of non-time series characteristics.

Sub-convolution neural networks for time series feature data: extracting relevant features among different indexes and time sequence features of the same index by using 1 × 3 and 3 × 1 convolution kernels in the first convolution layer, further extracting the features by using a 2 × 2 maximum pooling layer, then, extracting the features by using a convolution layer only comprising a 3 × 3 single convolution kernel, and finally, performing downsampling by using a 2 × 2 maximum pooling layer;

for a sub-convolution neural network of non-time series feature data: only one layer of 1 × 3 convolution kernel is used, and then a 2 × 2 pooling layer is used for feature extraction;

and finally, tiling the output matrixes of the two sub-convolution neural networks by using a flat, unifying the multidimensional input, combining the multidimensional input and the input through a full connection layer, and selecting a softmax function as an output classifier by using the output layer behind the full connection layer.

The convolutional neural network model is optimized by specifically adopting the following steps: the target loss function is 'bank-cross', the optimizer is Adam, the standard for measuring the model quality is precision accuracycacy, then in the process of training the model, the value-split is 0.2, the sequence of input samples is randomly disturbed before each epoch, the data size of each batch of training is 10, and the training is carried out 50 times.

The invention also discloses a medium and small micro enterprise data processing system based on the convolutional neural network, which comprises:

at least one memory cell;

at least one processing unit;

the storage unit stores at least one instruction; the instructions are loaded by at least one processing unit and perform the steps of:

collecting related data of small and medium-sized micro-enterprises;

preprocessing the acquired data;

constructing a convolutional neural network model;

and training and testing the constructed convolutional neural network model, and optimizing and determining parameters of the convolutional neural network model.

The processing unit includes:

a data acquisition subunit: the system is used for collecting related data of small and medium-sized micro enterprises;

a data preprocessing subunit: preprocessing the acquired data;

a model building subunit: the method is used for constructing a convolutional neural network model;

a training subunit: the method is used for training and testing the constructed convolutional neural network model.

The invention has the advantages that:

(1) the invention is a specific application of the convolutional neural network in the field of information processing of small and medium-sized micro-enterprises, further expands the application range of a deep learning algorithm, has higher pertinence and practicability, uses time sequence information (such as financial information and the like) and non-time sequence information (such as enterprise information and the like) for enterprise information, has higher innovation in the specific application of the convolutional neural network, and can more accurately reflect the basic situation of the small and medium-sized micro-enterprises through quantitative and qualitative multi-dimensional data analysis.

(2) In the data processing process, different processing methods are respectively adopted for data with time series characteristics and data with non-time series characteristics, and in consideration of more input characteristics of the data with the time series characteristics, in order to improve the classification accuracy, two convolution layers and two pooling layers are adopted for a sub-convolution network connected with the data with the time series characteristics. Considering that the input features of the non-time-series feature data are less, in order to avoid the problems of gradient extinction and explosion, which may be faced by too many layers of convolution layers and pooling layers, an overfitting phenomenon is generated, so that the sub-convolution network connected with the data without the time-series feature data only adopts one convolution layer and one pooling layer. In order to verify the effectiveness of the method, the data of the small and medium-sized micro-enterprises are evaluated, and the result shows that the accuracy of the training set is 0.95, the accuracy of the verification set is 0.99, the classification accuracy is high, and the overfitting phenomenon is not generated. Therefore, different processing methods are adopted for the data with the time series characteristics and the data with the non-time series characteristics, the processing accuracy of the enterprise data is improved, and the basic conditions of small and medium-sized micro enterprises can be more accurately reflected.

Drawings

FIG. 1 is a graph of training set accuracy using different numbers of layers of convolution and pooling, respectively;

FIG. 2 is a graph of validation set accuracy using different numbers of layers of convolution and pooling layers, respectively;

FIG. 3 is a graph of training set accuracy using both 2-layer convolution and 2-layer pooling;

FIG. 4 is a graph of validation set accuracy using both 2-layer convolution and 2-layer pooling;

FIG. 5 is a graph of training set accuracy using both 1-layer convolution and 1-layer pooling;

FIG. 6 is a validation set accuracy using both 1-layer convolution and 1-layer pooling;

fig. 7 is a process for processing data of small and medium-sized micro-enterprises by using a convolutional neural network.

Detailed Description

The invention will be further described with reference to specific embodiments, and the advantages and features of the invention will become apparent as the description proceeds.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. In addition, any methods and materials similar or equivalent to those described herein can be used in the practice of the present invention. The preferred embodiments and materials described herein are intended to be exemplary only.

Embodiment 1 medium and small micro enterprise data modeling method based on convolutional neural network

Firstly, collecting enterprise related data:

firstly, 1158 medium and small enterprises in the manufacturing industry of the medium and small enterprise boards and the entrepreneur enterprise boards are selected from a national tai-an database, ST and non-ST companies refer to a document of credit evaluation research of the small and small enterprises based on an improved dynamic combination evaluation method published in the management academic newspaper of Zhang invention 2019, financial factors are divided into 4 dimensions of profitability, debt paying capacity, operation capacity and growth capacity, and non-financial factors are divided into 2 dimensions of enterprise quality and enterprise owner quality. The financial indexes are data of 31 days in 12 months in 2017, 31 days in 12 months in 2018 and 31 days in 12 months in 2019, and the non-financial indexes are data of 31 days in 12 months in 2019. The specific indexes are selected as shown in table 1:

TABLE 1 basic information of small and micro enterprises

Secondly, quantifying non-financial data under the basic condition of an enterprise, wherein a specific quantification method can refer to a document of 'small micro enterprise credit evaluation research based on an improved dynamic combination evaluation method' published in 'management bulletin' in 2019, and directly quantifies non-financial index data such as enterprise registered capital, established years, education background, enterprise owner working years and the like into values between [0 and 1] according to threshold values, for example, the registered capital is divided into three intervals according to two threshold values of 1 hundred million yuan and 10 hundred million yuan, and the three intervals are respectively quantified into 0.6, 0.8 and 1; the established age is greater than 9 years, the quantization is 1, the quantization is 0.8 in (7, 9), the quantization is 0.6 in (5, 7), the quantization is 0.4 in (3, 5), the quantization is 0.2, not more than 1 year, the quantization is 0 in (1, 3), the educational background is the researchers and above, the quantization is 1, the national quantization is 0.8, the major expertise quantization is 0.7, the high-middle-expertise quantization is 0.5, the first-middle-school quantization is 0.3, the primary school quantization is 0.1, the academical history is 0, the business owner's working age is greater than 9 years, the quantization is 1, the quantization is 0.8 in (7, 9), the quantization is 0.6 in (5, 5), the quantization is 0.4 in (1, 3), the quantization is 0.2, not more than 1 year, and the quantization is 0.

Secondly, determining the credit risk judgment standard of the small and medium-sized micro enterprises:

whether the small and medium enterprises in the market are ST or not is used as a criterion for judging whether credit risk exists, the Shanghai Shen stock exchange is announced at 22/4/1998, and the stock exchange of the listed company with abnormal financial conditions or other conditions is specially processed according to the stock marketing rules implemented in 1998, namely (ST), the judgment that the enterprise is ST means that the listed company has at least one of the following six conditions:

1) the loss continues in the last two years.

2) The last annual audit of accounting showed that its shareholder equity was lower than registered capital, i.e., net assets per share was lower than the equity of the stock.

3) The registered accountant presents an audit report which can not represent opinion or deny opinion to the financial statement of the last accounting year.

4) The capital of registered accountants and related departments is lower than the capital of registered accounts when the capital of stockholders is audited in the last accounting year and the authority of registered accountants and related departments is not confirmed.

5) The last annual income was adjusted at the last audited financial report, resulting in two consecutive accounting annual losses.

6) The financial condition is considered abnormal by the exchange or the national certificate.

After the company on the market becomes ST, the fact that the company has a certain degree of risk of repayment capacity is meant, the repayment capacity of the company is possibly weakened, the company has higher credit risk, the repayment liability capacity is reduced, the repayment liability capacity is used as a definition standard of the credit risk, ST is an enterprise with high credit risk, and non-sT is an enterprise with low credit risk;

thirdly, preprocessing the data with the time series characteristics and the data with the non-time series characteristics:

the situation that the acquired original data is lost is possible, and meanwhile, different indexes have the problems of non-uniform dimension and the like, which are mainly reflected in that the sizes of the index values are not uniform, some index values appear in a specific gravity mode, and some indexes appear in a ratio mode. Firstly, for the problem of data loss, because each enterprise needs to acquire a lot of index data, if the enterprise uses simple deletion, the sample data volume is greatly reduced, which is not an optimal mode; for the problem of non-uniform dimension, if the original index value is directly analyzed, the influence of some indexes with overlarge values on the model is amplified, and thus the effectiveness of the model is influenced. The data therefore requires some pre-processing before the model is trained using the sample data. The preprocessing of the data comprises two steps of missing value processing and data standardization processing.

Missing value processing: because the obtained original data set has the conditions of missing values and the like, missing value filling is performed on the downloaded data set first, that is, the current missing value is filled by using the global average value of the index sequence, so that a complete data set is obtained.

And (3) data standardization treatment: because the magnitude of data of the data set is different, for example, the magnitude of surplus cash guarantee multiple, sales profit rate, capital profit rate and the like are greatly different, in order to eliminate the influence of different magnitudes of data, the data of different magnitudes are uniformly converted into the same magnitude, so the data are subjected to z-score standardization processing, firstly, the original data are assumed to be subjected to normal distribution, the standardization process is to process the data through the mean value and the standard deviation of a data sequence, the processed original data conform to the standard normal distribution, the median (mu) of the observation value group is subtracted from the observation value, and then the median (sigma) is divided to obtain the data, and the training speed and the prediction accuracy of the model are improved. The formula is as follows:

x′＝(x-μ)/σ

fourthly, constructing a convolutional neural network model:

and constructing a convolutional neural network model, and forming a stable convolutional neural network model by designing structures in the convolutional neural network, a convolutional layer, a pooling layer, an output layer Softmax and a full connection layer.

Fifthly, training and testing the constructed convolutional neural network model, and learning and optimizing parameters of the convolutional neural network model:

data in an enterprise data information base are randomly divided into a training set and a testing set, iterative training is carried out on the convolutional neural network model by taking the training set data as input, then the model is tested by the testing set, and parameter adjustment is carried out according to a testing result, so that the model tends to be stable.

The model includes two parallel sub-convolutional neural networks, a first sub-network and a second sub-network, for inputting financial data with time-series characteristics and non-financial data without time-series characteristics, respectively.

In the data processing process, different processing methods are respectively adopted for the data with the time series characteristics and the data with the non-time series characteristics, and in consideration of more input characteristics of the data with the time series characteristics, in order to improve the classification accuracy, a sub-convolution network connected with the time series characteristic data adopts two convolution layers and two pooling layers. Considering that the input features of the non-time-series feature data are fewer, in order to avoid the problems of gradient extinction and explosion caused by excessive layers of the convolutional layer and the pooling layer, an overfitting phenomenon is generated, so that only one convolutional layer and one pooling layer are adopted in the sub-convolutional network connected with the non-time-series feature data, as can be seen from fig. 1 and 2, the accuracy of the training set is 0.95, the accuracy of the verification set is 0.99, the classification accuracy is higher, and the overfitting phenomenon is not generated.

If the number of convolutional and pooling layers is too large, problems of gradient disappearance and explosion may be encountered, and even an overfitting phenomenon may occur. In an experiment of the invention, the two kinds of data are processed by adopting the first sub-network, under the condition of ensuring that other experiment parameter settings are not changed, two layers of convolution and two layers of pooling layers are adopted for the data of the time series characteristic and the data of the non-time series characteristic, and as a result, as shown in fig. 3 and 4, the accuracy of a training set is 0.85, the accuracy of a verification set is 0.87, and the classification accuracy is reduced.

If the number of the convolutional layers and the pooling layers is too small, the recognition capability of the model on the data characteristics is weak, so that the classification accuracy is reduced. In another experiment of the present invention, both the two kinds of data are processed by using the second sub-network, and under the condition that other experiment parameter settings are not changed, a layer of convolution and a layer of pooling layer are used for both the data of the time series characteristic and the data of the non-time series characteristic, and as a result, as shown in fig. 5 and fig. 6, the accuracy of the training set is 0.79, the accuracy of the verification set is 0.76, and the classification accuracy is reduced.

Therefore, specifically, the following processing is adopted:

for a sub-convolution neural network that receives time series feature data (this embodiment refers to financial data): the first layer of convolutional layer uses two convolutional kernels of 1 × 3 and 3 × 1 to extract related features among different indexes and time sequence features of the same index, the depth of the convolutional kernels is 128, padding is same, BN (boron nitride) batch normalization operation is performed, and an activation function is relu; then padding is filled to same and Dropout is 0.1 through the maximum pooling layer of 2 multiplied by 2; then, a convolution layer which only comprises 3 multiplied by 3 single convolution kernels is connected to further extract the features, the fill is same, the depth of the convolution kernels is 128, BN normalization is carried out, and the activation function is relu; and finally, performing down-sampling through a 2 × 2 maximum pooling layer, wherein padding is filled to same, and Dropout is 0.1.

For a sub-convolutional neural network that receives non-time-series feature data (this embodiment refers to non-financial data): because the non-financial data does not have time sequence characteristics, input information is less, the partial sub-network only uses a layer of 1 × 3 convolution kernels, the fill is same, the depth of the convolution kernels is 128, BN normalization is performed, and an activation function is relu; feature extraction was then performed using a 2 × 2 pooling layer with a dropout operation of 0.1.

And finally, tiling the output matrixes of the two sub-convolution neural networks by using a flat, unifying the multidimensional input, combining the multidimensional input and the input, passing through a full connection layer, and finally selecting a softmax function as an output classifier by using the output layer.

Optimizing the model, wherein the target loss function is 'binary-cross', the optimizer is Adam (the learning rate is 0.004), and the standard for measuring the quality of the model is precision accuracy; then, in the process of training the model, the evaluation-split value is 0.2, the sequence of input samples is randomly disturbed before each epoch, the data size of each batch of training is 10, and the training is carried out for 50 times.

In the training process, each sample in the training set is processed, the loss of the output result and the label classification result is solved for training, the test set is used for calculating the accuracy of the output result and the label classification result, and the test set does not participate in the training; the following method is adopted for training each sample data (namely, data of each enterprise) so that the data tends to be stable:

1) inputting enterprise financial data input _1, wherein the data structure is (None, 3, 43, 1); and inputting enterprise non-financial data input _2, wherein the data structure is (None, 1, 4, 1).

2) For the first sub-network, the step 1) of enterprise financial data input _1 is subjected to a first layer of convolution operation: the first convolution layer comprises two convolutions, two convolution kernels of 1 × 3 and 3 × 1 are respectively used, wherein the convolution kernel of 1 × 3 aims to extract correlation characteristics between different indexes, and the convolution kernel of 3 × 1 aims to extract time sequence characteristics of the same index. According to the convolution kernel decomposition, n multiplied by n convolution kernels are decomposed into two one-dimensional convolutions (1 multiplied by n, n multiplied by 1), so that the calculation can be accelerated, the parameter scale can be reduced, and the network depth and the nonlinearity can be increased because the number of the convolution kernels is doubled; the filling is same, so that the size of the input vector is unchanged after the convolution; the depth of the convolution kernel is 128, and padding is same as same; BN batch normalization operation; the activation function is ReLU, and compared with a Sigmoid function, the ReLU eliminates the gradient saturation effect at the x part which is greater than or equal to 0, and the calculation of the ReLU is simpler. However, the ReLU itself has drawbacks: if the input becomes negative, the network training will not be successfully completed with a gradient equal to 0. Even so, ReLU is still one of the most commonly used activation functions in the current deep learning field. The activation functions used by the model are all relus.

Both convolution output shapes are (None, 3, 43, 128).

3) For the first sub-network, the connection operation: and (3) splicing the results (None, 3, 43, 128) output by the two convolutions of the first layer convolution operation in the step 2) in depth, wherein the output result is (None, 3, 43, 256).

4) For the first subnetwork, max pooling: and (3) outputting the output result (None, 3, 43, 256) spliced in the step (3), wherein padding is filled to same through the maximum pooling layer of 2 multiplied by 2, Dropout is 0.1, and the output result (None, 2, 22, 256) is output.

5) For the first subnetwork, a second layer of convolution operations is performed: outputting the result (None, 2, 22, 256) of the step 4), and using a convolution kernel of 3 multiplied by 3; the filling is same, so that the size of the input vector is unchanged after the convolution; the depth of the convolution kernel is 128; BN normalization; the activation function is relu and the output result is (None, 2, 22, 128).

6) For the first subnetwork, the second max pooling: performing down-sampling on the output result (None, 2, 22, 128) in the step 5) through a 2 multiplied by 2 maximum pooling layer; padding is filled to same, dropout operation is 0.1, and the output result is (None, 1, 11, 128).

7) For the first subnetwork, tiling layer: the Flatten layer is used to "Flatten" the input, i.e., to dimension the multidimensional input, often used in the transition from convolutional layers to fully-connected layers, to tile the output result (None, 1, 11, 128) of step 6), and output the result (None, 1408).

8) For the second subnetwork, step 7) enterprise non-financial data input _2 convolution operation: a layer of 1 × 3 convolution kernels is used; the filling is same, so that the size of the input vector is unchanged after the convolution; the depth of the convolution kernel is 128; BN normalization; the activation function is relu and the output result is (None, 1, 4, 128).

9) For the second subnetwork, max pooling: using a 2 multiplied by 2 pooling layer for the output (None, 1, 4, 128) of step 8); dropout operation is 0.1 and the output result is (None, 1, 2, 128).

10) For the second subnetwork, tiling layer: the Flatten layer is used to "Flatten" the input, i.e., to dimension the multidimensional input, often used in the transition from convolutional layers to fully-connected layers, and to tile the output result (None, 1, 2, 128) of step 9), and the output result is (None, 256).

11) And (3) splicing the first sub-network and the second sub-network, and splicing the output result (None, 1408) in the step 7) with the output result (None, 256) in the step 10), wherein the output result is (None, 1664).

12) Full connection operation: and (4) performing full connection operation on the output result (None, 1664) in the step 11), wherein units are 128, the activation function is relu, and the output result is (None, 128).

18) An output layer: the softmax function is selected as the output classifier.

19) Compiling the model, optimizing the model: the target loss function is "binary-cross loss", the optimizer is Adam (the learning rate is 0.004), and the criterion for measuring the model is accuracy accuracuracy.

20) Training a model: and (3) evaluation-split (taking a value of 0.2), namely, one fifth of data in a specified training set is taken as a verification set in each training, and relevant indexes of the model, such as accuracy, a loss function and the like, are tested after each training is finished. And the shuffle value is set as true, which indicates that the sequence of the input samples is randomly disturbed before each epoch in the training process, thereby ensuring the randomness of the verification set and increasing the robustness of the model. The size of each batch of data batch-size (value 10) indicates that the data size of each batch of training is 10 in the training process. The number of times epochs (value 50) represents 50 times of training.

Example 2

A convolutional neural network-based medium and small micro enterprise data modeling system, the system comprising:

at least one memory cell;

at least one processing unit;

collecting related data of small and medium-sized micro-enterprises;

preprocessing the acquired data;

constructing a convolutional neural network model;

and training and testing the constructed convolutional neural network model, and learning and optimizing parameters of the convolutional neural network model.

The processing unit includes:

a data preprocessing subunit: preprocessing the acquired data;

The specific data processing process of the system adopts the method of embodiment 1.

The above-mentioned embodiments are merely preferred embodiments of the present invention, which are merely illustrative and not restrictive, and it should be understood that other embodiments may be easily made by those skilled in the art by replacing or changing the technical contents disclosed in the specification, and therefore, all changes and modifications that are made on the principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. The medium and small micro enterprise data modeling method based on the convolutional neural network is characterized by comprising the following steps:

(1) collecting related data of small and medium-sized micro-enterprises;

(2) preprocessing the acquired data;

(3) constructing a convolutional neural network model;

2. The method according to claim 1, wherein the preprocessing in step (2) includes missing value processing and data normalization processing.

3. The method according to claim 1, wherein the missing value processing is missing value filling of the downloaded data set, i.e. filling the current missing value with the global average value of the index sequence to obtain a complete data set; the data normalization process employs a z-score normalization process.

4. The method of claim 1, wherein the training and testing of the constructed convolutional neural network model in step (4) is performed by: randomly dividing the acquired enterprise data into a training set and a testing set, performing iterative training on the convolutional neural network model by taking the training set data as input, testing the model by using the testing set, and adjusting parameters according to the testing result.

5. The method of claim 1, wherein the convolutional neural network model comprises two parallel sub-convolutional neural networks for receiving time-series characteristic data and non-time-series characteristic data, respectively.

6. The method of claim 5, wherein for a sub-convolutional neural network receiving time series feature data: extracting relevant features among different indexes and time sequence features of the same index by using 1 × 3 and 3 × 1 convolution kernels in the first convolution layer, further extracting the features by using a 2 × 2 maximum pooling layer, then, extracting the features by using a convolution layer only comprising a 3 × 3 single convolution kernel, and finally, performing downsampling by using a 2 × 2 maximum pooling layer;

7. The method according to claim 1, wherein the optimization of the convolutional neural network model specifically employs: the target loss function is 'bank-cross', the optimizer is Adam, the standard for measuring the model quality is precision accuracycacy, then in the process of training the model, the value-split is 0.2, the sequence of input samples is randomly disturbed before each epoch, the data size of each batch of training is 10, and the training is carried out 50 times.

8. A convolutional neural network-based small and medium enterprise data processing system, comprising:

at least one memory cell;

at least one processing unit;

collecting related data of small and medium-sized micro-enterprises;

preprocessing the acquired data;

constructing a convolutional neural network model;

9. The system of claim 8, wherein the processing unit comprises:

a data preprocessing subunit: preprocessing the acquired data;