Number recognition model training method, number recognition method and device

By generating and training sample data features for a number recognition model, and utilizing a convolutional neural network model, the problem of the inability to identify fraudulent numbers in advance in existing technologies is solved, achieving efficient identification of fraudulent numbers and reducing telephone fraud.

CN112396079BActive Publication Date: 2024-11-08CHINA MOBILE GRP GUANGDONG CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201910760686.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2019-08-16
Publication Date
2024-11-08
Estimated Expiration
2039-08-16

AI Technical Summary

Technical Problem

Existing number identification models can only identify numbers that have already been used for fraud, and cannot identify fraudulent numbers in advance.

Method used

By acquiring raw data and corresponding classification labels of multiple numbers within a predetermined time period, sample data features are generated. A number recognition model is then trained using a convolutional neural network model, ensuring that the loss function is less than a predetermined threshold, in order to identify fraudulent numbers.

Benefits of technology

It enables early identification of fraudulent numbers, reducing the occurrence of telephone fraud.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112396079B_ABST
    Figure CN112396079B_ABST
Patent Text Reader

Abstract

The present invention provides a number recognition model training method, a number recognition model training device, a number recognition method, a number recognition device and an electronic device, wherein the number recognition model training method comprises: obtaining the original data and corresponding classification labels of multiple numbers within a predetermined time period, wherein the original data comprises service data and signaling data; generating sample data features corresponding to each number based on the original data of each number, wherein the sample data features comprise service data features and signaling data features; training the number recognition model based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold. The technical solution of the embodiment of the present invention can train a number recognition model that can extract number feature data efficiently and quickly, and identify fraudulent numbers in advance through the model, thereby effectively reducing telephone fraud.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a number recognition model training method, a number recognition model training device, a number recognition method, a number recognition device, an electronic device, and a computer-readable storage medium. Background Art

[0002] Nowadays, more and more people use telephones, and various activities such as advertising and fraud are common through telephones. How to quickly and accurately identify telephone numbers used for fraud or advertising and promotion has become a focus of attention.

[0003] The number recognition model in the existing technical solution extracts feature data from the user's call signaling data analysis based on the fraudulent numbers that have occurred, and then classifies and recognizes them through common machine learning algorithms such as clustering and random forest. However, this technical solution can only identify numbers that have already committed fraudulent acts, and cannot identify fraudulent numbers in advance. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a number recognition model training method, a number recognition model training device, a number recognition method, a number recognition device, an electronic device and a computer-readable storage medium, which can obtain a number recognition model that can identify fraudulent numbers in advance. By using this model to identify fraudulent numbers in advance, telephone fraud can be effectively reduced.

[0005] To solve the above technical problems, the embodiments of the present invention are implemented as follows:

[0006] In the first aspect, an embodiment of the present invention provides a number recognition model training method, including: obtaining original data and corresponding classification labels of multiple numbers within a predetermined time period, the original data including service data and signaling data; generating sample data features corresponding to each number based on the original data of each number, the sample data features including service data features and signaling data features; training a number recognition model based on the sample data features and classification labels corresponding to each number so that the loss function corresponding to the number recognition model is less than a predetermined threshold.

[0007] In a second aspect, an embodiment of the present invention provides a number identification method, including: obtaining original data of a target number, the original data including service data and signaling data; generating production data features corresponding to the target number based on the original data of the target number, the production data features including service data features and signaling data features; inputting the production data features into a number identification model as described in the first aspect above to obtain output data; and determining a classification label corresponding to the target number based on the output data.

[0008] In the third aspect, an embodiment of the present invention provides a number recognition model training device, including: a data acquisition module, used to acquire original data and corresponding classification labels of multiple numbers within a predetermined time period, the original data including business data and signaling data; a sample data feature generation module, used to generate sample data features corresponding to each number based on the original data of each number, the sample data features including business data features and signaling data features; a model training module, used to train the number recognition model based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold.

[0009] In a fourth aspect, an embodiment of the present invention provides a number identification device, comprising: a data acquisition module, used to acquire original data of a target number, the original data including service data and signaling data; a production data feature generation module, used to generate production data features corresponding to the target number based on the original data of the target number, the production data features including service data features and signaling data features; a data input module, used to input the production data features into a number identification model as described in the first aspect above to obtain output data; and a classification label determination module, used to determine the classification label corresponding to the target number based on the output data.

[0010] In a fifth aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the number recognition model training method as described in the first aspect or the number recognition method as described in the second aspect is implemented.

[0011] In a sixth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the number recognition model training method as described in the first aspect or the number recognition method as described in the second aspect is implemented.

[0012] According to the technical solution of the embodiment of the present invention, on the one hand, the original data and corresponding classification labels of multiple numbers within a predetermined time period are obtained, and the original data includes business data and signaling data, so that the business data and various signaling data for training the model can be obtained; on the other hand, sample data features corresponding to each number are generated based on the original data of each number, and the sample data features include business data features and signaling data features, so that various sample feature data of the number can be efficiently generated without manual analysis and extraction; on the other hand, the number recognition model is trained based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold, and a number recognition model for extracting and identifying fraudulent numbers can be trained. Fraudulent numbers can be identified in advance through this model, which can effectively reduce telephone fraud. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0014] Figure 1 A schematic diagram showing a flow chart of a number recognition model training method provided according to some embodiments of the present invention;

[0015] Figure 2 A schematic diagram showing a flow chart of a number identification method provided according to some embodiments of the present invention;

[0016] Figure 3 A schematic diagram showing a flow chart of a number identification method provided according to some other embodiments of the present invention;

[0017] Figure 4 A schematic diagram showing a number recognition model training method provided according to some other embodiments of the present invention;

[0018] Figure 5 A schematic diagram showing multi-layer matrix data provided according to some embodiments of the present invention is shown;

[0019] Figure 6 A schematic diagram of a number recognition model provided according to some embodiments of the present invention is shown;

[0020] Figure 7 A schematic diagram showing a number identification method according to some further embodiments of the present invention;

[0021] Figure 8A schematic block diagram of a number recognition model training device according to some embodiments of the present invention is shown;

[0022] Fig. 9 A schematic block diagram of a number identification device provided according to some embodiments of the present invention is shown; and

[0023] Fig.10 A schematic structural diagram of an electronic device provided according to some embodiments of the present invention is shown. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0025] Figure 1 FIG. 1 is a flow chart showing a method for training a number recognition model according to some embodiments of the present invention. Figure 1 As shown, the number recognition model training method includes steps S110, S120, and S130, and can be applied to the server side. Figure 1 The number recognition model training method in the example embodiment is described in detail.

[0026] Reference Figure 1 As shown, in step S110, original data of multiple numbers within a predetermined time period and corresponding classification labels are obtained, and the original data includes service data and signaling data.

[0027] In an example embodiment, a data acquisition cycle is defined, that is, a predetermined time period, such as 30 days, specifically, the original data of the number from T-1 to T-30 and the corresponding classification labels are obtained on date T. For example, on August 2, 2019, the original data of the number from August 1, 2019 to July 3, 2019 are obtained. The original data of the same number includes business data and signaling data. Business data includes but is not limited to: account opening information, business processing, and network access channels; signaling data includes but is not limited to: number of call digits, Internet traffic, and location information. In addition, the original data of different numbers are divided into normal number data and fraud number data. The classification labels corresponding to the normal number data are normal classification 1, normal classification 2, normal classification 3, etc.; the classification labels corresponding to the fraud number data are fraud classification 1, fraud classification 2, fraud classification 3, etc.

[0028] In step S120, sample data features corresponding to each number are generated based on the original data of each number, and the sample data features include service data features and signaling data features.

[0029] In an example embodiment, the original data of each number is preprocessed, and the service data and signaling data in the original data are converted into customized multi-layer matrix data to obtain sample data features. Specifically, based on the time sequence of the predetermined time period, multi-layer matrix data corresponding to the original data of each number is generated, and the multi-layer matrix data includes but is not limited to: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix, and a location data layer matrix; each layer of matrix data of the multi-layer matrix data is converted into a floating point number and normalized to obtain sample data features corresponding to the original data of the number, and the sample data features include service data features and signaling data features.

[0030] Further, in an exemplary embodiment, based on the time sequence of a predetermined time period, the step of generating a multi-layer matrix data corresponding to the original data of each number includes: generating data rows of the matrix data of the corresponding data layer based on the data of the corresponding data layer per unit time of the predetermined time period, the corresponding data layer being one of a service data layer, a signaling call data layer, an Internet traffic data layer, and a location data layer; generating the matrix data of the corresponding data layer based on the data rows of the corresponding data layer per unit time according to the time sequence of the predetermined time period.

[0031] For example, a multi-layer matrix data is divided into 4 layers, namely, the business data layer matrix, the signaling call data layer matrix, the Internet traffic data layer matrix and the location data layer matrix. The business data layer matrix is ​​the first layer, which is the data in a single-layer matrix format obtained after the business-related data is integrated, and the same applies to other matrices. The columns of the business data matrix layer are preset fixed values, such as 15, and the rows are the data acquisition cycles mentioned in step S110, such as 30. On August 2, 2019, the original data of the number for 30 days between August 1, 2019 and July 3, 2019 is obtained. The cycle starts from July 3, 2019, and the data of the day is obtained and assembled into 15 columns of data and added to the matrix table until August 1, 2019, that is, the 15x30 time series data in the business data matrix layer is completed. Similarly, the signaling call data layer matrix, the Internet traffic data layer matrix and the location data layer matrix also assemble data according to this method. If there is no data for a number on a certain day, it is set to 0 by default.

[0032] In the example embodiment, the sample data features obtained after preprocessing the normal number data are called normal samples, and the sample data features obtained after preprocessing the fraud number data are called fraud samples. In the original data obtained in step S110, the normal samples are much more than the fraud samples. The normal samples and the fraud samples are assembled in a predetermined ratio to obtain a sample group. For example, the normal samples and the fraud samples assembled in a ratio of 3:1 are one sample group.

[0033] In step S130, the number recognition model is trained based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold.

[0034] In an example embodiment, the sample group consisting of the sample data features obtained in step S120 is divided into training data samples and test data samples according to a predetermined ratio, for example, the predetermined ratio is 7:3, that is, in the sample data features, 70% of the sample group is used as training data samples for training the model, and 30% of the sample group is used as test data samples for testing the model. Based on the training data samples and the corresponding classification labels, the number recognition model is trained by a gradient optimizer. The number recognition model is a convolutional neural network model. The number recognition model is tested using the test data samples and the corresponding classification labels, and its parameters are adjusted. When the loss function corresponding to the number recognition model is less than a predetermined threshold, it can be regarded that the model can realize the function of identifying fraudulent numbers and has achieved the purpose of training. If the number recognition model passes the test, the number recognition model can be used to identify fraudulent numbers.

[0035] In an example embodiment, training data samples and corresponding classification labels are respectively stored in the storage space corresponding to the one-to-one field train_x and the storage space corresponding to the field train_y, and test data samples and corresponding classification labels are respectively stored in the storage space corresponding to the one-to-one field test_x and the storage space corresponding to the field test_y.

[0036] In an example embodiment, the process of training the number recognition model is specifically as follows: preprocessing the original data of the number as described in step S120; starting the convolutional neural network adapter, reading the preset network configuration parameters, loading the preset network structure, defining the loss function and the gradient optimizer; setting the model training result save path, judging whether there is a previous training result, then loading it; if not, creating it; according to the preset number of single training data inputs, that is, parameter batch_size, obtaining the corresponding number of training data samples from the field train_x and the field train_y to input the model; looping the training data according to the preset training rounds; testing the model based on the test data sample, and saving the model if it passes the test.

[0037] In an exemplary embodiment, the loss function loss is defined, and the cross entropy can be used as the loss function of classification, and the reference formula is: The selection of this loss function is only an example, and this application may also select other functions as the loss function.

[0038] In the exemplary embodiment, a gradient optimizer is defined, and an "Adam" optimizer may be used, with a learning rate of 0.001. The selection of the above optimizer and the learning rate are only examples, and the present application does not impose any special restrictions on this.

[0039] according to Figure 1 The number recognition model training method in the example embodiment, on the one hand, obtains the original data and corresponding classification labels of multiple numbers within a predetermined time period, the original data includes business data and signaling data, and can obtain business data and various signaling data for training the model; on the other hand, based on the original data of each number, generates sample data features corresponding to each number, the sample data features include business data features and signaling data features, and can efficiently generate various sample feature data of the number without manual analysis and extraction; on the other hand, based on the sample data features and classification labels corresponding to each number, the number recognition model is trained so that the loss function corresponding to the number recognition model is less than a predetermined threshold, and a number recognition model for extracting and identifying fraudulent numbers can be trained. Fraudulent numbers can be identified in advance through the model, which can effectively reduce telephone fraud.

[0040] Figure 2 FIG. 1 is a flow chart showing a method for number identification according to some embodiments of the present invention. Figure 2 As shown, the number identification method includes steps S210, S220, S230, and S240, and can be applied to terminal devices, and the mobile terminal includes but is not limited to smart terminals such as mobile phones, tablet computers, computers, and wearable devices. Figure 2 The number identification method in the exemplary embodiment is described in detail.

[0041] Reference Figure 2 As shown, in step S210, original data of the target number is obtained, and the original data includes service data and signaling data.

[0042] In an example embodiment, the original data of the target number to be identified is obtained based on a predetermined time period. The predetermined time period is the data acquisition cycle, for example, 30 days, specifically obtaining the original data of the target number from T-1 to T-30 on date T, such as obtaining the original data of the number for 30 days between August 1, 2019 and July 3, 2019 on August 2, 2019. The original data of the target number comes from the actual production process, so it is also called production data. The original data of the same number includes business data and signaling data. Business data includes but is not limited to: account opening information, business handling, and network access channels; signaling data includes but is not limited to: number of call digits, Internet traffic, and location information.

[0043] It should be noted that, in the exemplary embodiment, the original data of the target number obtained in step S210 is the same as Figure 1 The original data of the target number does not have a classification label, and before being identified by the number recognition model, it is impossible to know whether it is a fraudulent number.

[0044] In step S220, production data features corresponding to the target number are generated based on the original data of the target number, and the production data features include service data features and signaling data features.

[0045] In an example embodiment, the original data of the target number is preprocessed, and the service data and signaling data in the original data are converted into customized multi-layer matrix data to obtain production data features. Specifically, based on the time sequence of the predetermined time period, multi-layer matrix data corresponding to the original data of the target number is generated, and the multi-layer matrix data includes but is not limited to: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix, and a location data layer matrix; each layer of matrix data of the multi-layer matrix data is converted into a floating point number and normalized to obtain production data features corresponding to the original data of the target number, and the production data features include service data features and signaling data features.

[0046] Further, in an exemplary embodiment, based on the time sequence of a predetermined time period, a method for generating multi-layer matrix data corresponding to the original data of a target number is as follows: based on the data of the corresponding data layer per unit time of the predetermined time period, data rows of matrix data of the corresponding data layer are generated, the corresponding data layer being one of a service data layer, a signaling call data layer, an Internet traffic data layer, and a location data layer; and according to the time sequence of the predetermined time period, matrix data of the corresponding data layer is generated based on the data rows of the corresponding data layer per unit time.

[0047] For example, a multi-layer matrix data is divided into 4 layers, namely, the business data layer matrix, the signaling call data layer matrix, the Internet traffic data layer matrix and the location data layer matrix. The business data layer matrix is ​​the first layer, which is the data in a single-layer matrix format obtained after the business-related data is integrated, and the same applies to other matrices. The columns of the business data matrix layer are preset fixed values, such as 15, and the rows are the data acquisition cycles mentioned in step S110, such as 30. On August 2, 2019, the original data of the number for 30 days between August 1, 2019 and July 3, 2019 is obtained. The cycle starts from July 3, 2019, and the data of the day is obtained and assembled into 15 columns of data and added to the matrix table until August 1, 2019, that is, the 15x30 time series data in the business data matrix layer is completed. Similarly, the signaling call data layer matrix, the Internet traffic data layer matrix and the location data layer matrix also assemble data according to this method. If there is no data for a number on a certain day, it is set to 0 by default.

[0048] In step S230, the production data features are input into the number recognition model to obtain output data.

[0049] In the example embodiment, the configuration parameters of the number recognition model are read and the network model structure is loaded; the address of the trained model is read and the model data is loaded; the production data features are input into the model in batches to obtain output data, which can be named out_data.

[0050] Based on the output data, determine the classification label corresponding to the target number.

[0051] In an example embodiment, based on the output data out_data, the probability of classification of the target number is determined, which may be named logits_data; based on the probability of classification logits_data, the classification label identification code corresponding to the target number is determined, which may be named pre_label_idx; the classification label identification code pre_label_idx is matched with the classification label name in the classification label library, which stores identification codes of multiple classification labels, such as label_idx, and names, such as label_name, to obtain the classification label name corresponding to the target number, which may be named pre_label_name.

[0052] In the example embodiment, the probability of the classification of the target number is determined by outputting the data, which can be calculated using the softmax() function or by other algorithms. Based on the probability of the classification, the classification label identification code corresponding to the target number is determined, which can be calculated using the argmax() function or by other algorithms.

[0053] In an example embodiment, the target number may be named user_num, and the target number may be matched with an account opening channel name in an account opening channel table, such as num_chanel_name, to obtain the account opening channel name corresponding to the target number.

[0054] In an exemplary embodiment, the target number user_num and corresponding parameters, such as the target number classification probability logits_data, classification label identification code pre_label_idx, classification label name pre_label_name and account opening channel name num_chanel_name, are assembled into final output data.

[0055] It should be noted that, in the exemplary embodiment, the number of output data is the same as the number of input data, that is, the number of input target numbers is the same as the number of final output data including the target numbers and corresponding parameters.

[0056] In the example embodiment, the suspected fraudulent numbers and channel information are outputted, specifically, the final output data is pushed to the corresponding storage table of the database according to the system settings; and then pushed to other platforms according to the settings.

[0057] according to Figure 2 The number identification method in the example embodiment, on the one hand, obtains the original data of the target number, the original data includes business data and signaling data, and can obtain the original data of the target number; generates production data features corresponding to the target number based on the original data of the target number, the production data features include business data features and signaling data features, and can efficiently generate multiple production data features of the target number without manual analysis and extraction; inputs the production data features into the number identification model to obtain output data, and can obtain the output data of the number identification model; based on the output data, determines the classification label corresponding to the target number, can obtain the classification label corresponding to the target number, and determines whether the target number is a fraud number, thereby reducing telephone fraud.

[0058] Figure 3 A schematic flow chart of a number identification method provided according to some other embodiments of the present invention is shown.

[0059] In step S310, data is input.

[0060] In an example embodiment, the original data of the acquired number is input, and the data includes training data and production data, both of which include business data and signaling data. Business data includes account opening, handling, and channel data; signaling data includes call, location, and Internet access data.

[0061] In step S320, data preprocessing is performed.

[0062] In an exemplary embodiment, the service data and signaling data of the numbers in the data obtained in step S310 are assembled into multi-layer matrix data according to the time sequence of number generation, and then normalized to obtain sample data features.

[0063] In step S330, the deep learning model is trained.

[0064] In the example embodiment, the convolutional neural network model is initialized, and parameters such as the number of convolution layers, the size and number of convolution kernels, and training parameters are defined; the model is trained with preprocessed data to achieve the optimal effect; the model is then tested, and if it passes the test, the model parameters and configuration are saved.

[0065] In step S340, suspected fraud numbers are extracted.

[0066] In an example embodiment, production data is input into a trained and tested model to obtain output results, and classification labels corresponding to each number in the production data are determined based on the output results. Normal numbers and fraudulent numbers can be extracted based on the classification labels.

[0067] In step S350, the result is output.

[0068] The suspected fraud numbers and channel information will be output to the corresponding storage table in the database and pushed to other platforms according to the settings.

[0069] Figure 4 A schematic diagram of a number recognition model training method provided according to some other embodiments of the present invention is shown.

[0070] Reference Figure 4 As shown, in step S410, training data is input.

[0071] In an example embodiment, the input training data is as follows Figure 1 In step S110, original data of multiple numbers within a predetermined time period and corresponding classification labels are obtained.

[0072] In step S420, a data preprocessing operation is performed on the training data.

[0073] In an exemplary embodiment, based on the time sequence of a predetermined time period, multi-layer matrix data corresponding to the original data of each number is generated, and the multi-layer matrix data includes but is not limited to: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix, and a location data layer matrix; each layer of matrix data of the multi-layer matrix data is converted into a floating point number and normalized.

[0074] In step S430, feature data X is obtained.

[0075] In the exemplary embodiment, the feature data X, ie, the sample data feature, is obtained by converting the training data into corresponding multi-layer matrix data.

[0076] In step S440, a classification label Y is obtained.

[0077] In an exemplary embodiment, the classification label Y is obtained from the original data of the number and the corresponding classification label obtained in step S410.

[0078] In step S450, convolutional neural network 1 is input.

[0079] In the exemplary embodiment, the convolutional neural network 1 is an example of a number recognition model, and the feature data X is input into the convolutional neural network 1 to obtain output data.

[0080] In step S460, the output data obtained in step S450 and the classification label Y obtained in step S440 are input into the loss function.

[0081] In the exemplary embodiment, the output data is the theoretical value of the classification label, and the classification label Y is the actual value of the classification label. The two are operated through the loss function to adjust the parameters of the convolutional neural network 1.

[0082] In step S470, the gradient optimizer optimizes the convolutional neural network 1.

[0083] In an exemplary embodiment, a gradient optimizer acts on the loss function to optimize the parameters of the convolutional neural network 1.

[0084] In step S480, the trained model is stored.

[0085] In the exemplary embodiment, after training the convolutional neural network 1, it is tested, and if it passes the test, the model is stored in a specified location.

[0086] Figure 5 A schematic diagram of multi-layer matrix data provided according to some embodiments of the present invention is shown.

[0087] Reference Figure 5 As shown, the multi-layer data matrix is ​​divided into C layers, for example, 3 layers, namely the business data layer matrix, the signaling call data layer matrix, and the Internet traffic data layer matrix; H is the number of rows of each layer of matrix data, that is, the data acquisition cycle, and W is the number of columns of each layer of matrix data, which is a preset fixed value.

[0088] It should be noted that in the example embodiment, W / H / C are all adjustable parameters. For example, when a new data feature "location information data" that is different from business data, signaling call data, and Internet traffic data needs to be added during data preprocessing, the number of layers C increases by 1, from 3 to 4.

[0089] Figure 6 A schematic diagram of a number recognition model provided according to some embodiments of the present invention is shown.

[0090] Reference Figure 6 The number recognition model is a convolutional neural network model. In step S610, the model receives input data, namely, sample data features or production data features in a multi-layer matrix data format, and its parameters are matrix data columns W, matrix data rows H, matrix layers C, and data size WxHxC.

[0091] In step S620, the parameters of the first convolutional layer conv1 are set as follows: the convolution kernel size is 3x3, the output is 256, the stride is 1, and the expansion edge is the same, that is, pad=SAME.

[0092] In step S630, the parameters of the second convolutional layer conv2 are set as follows: the convolution kernel size is 5x5, the output is 512, the stride is 1, and the expansion edge is the same, that is, pad=SAME.

[0093] Then in step S640, a maximum pooling layer maxpool is connected, and its parameters are as follows: the convolution kernel size is 2x2, and the data size becomes matrix data column / 2x matrix data row / 2x512.

[0094] In step S650, the third convolutional layer conv3 is connected, and its parameters are set as follows: the convolution kernel size is matrix data columns / 2xmatrix data rows / 2, the output is 1024, the step size stride is 1, and the expansion edge is invalid, that is, pad=VALID. At this time, the data output size is 1x1x1024.

[0095] In step S660, the parameters of the four-layer convolution layer conv4 are set as follows: the convolution kernel size is 1x1, the sum of the normal number classification and the fraud number classification cls_num is output, referred to as the classification sum, the step length stride is 1, and the expansion edge is the same, that is, pad = SAME.

[0096] In step S670, the model outputs data, and the size of the output data is 1x1x classification sum.

[0097] Figure 7 A schematic diagram of a number identification method according to some further embodiments of the present invention is shown.

[0098] Reference Figure 7 As shown, in step S710, production data is obtained.

[0099] In an exemplary embodiment, original data, ie, production data, of a target number is acquired based on a predetermined period of time.

[0100] In step S720, data is preprocessed.

[0101] In an example embodiment, data preprocessing operations are performed on production data, specifically, based on the time sequence of a predetermined time period, multi-layer matrix data corresponding to the original data of each number is generated, the multi-layer matrix data including but not limited to: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix, and a location data layer matrix; each layer of matrix data of the multi-layer matrix data is converted into floating point numbers and normalized.

[0102] In step S730, feature data X is acquired.

[0103] In an exemplary embodiment, the feature data X, ie, the production data feature, is obtained by converting the production data into corresponding multi-layer matrix data.

[0104] In step S740, the feature data X is input into the convolutional neural network model 1.

[0105] In the exemplary embodiment, the convolutional neural network 1 is an example of a number recognition model, and the feature data X is input into the convolutional neural network 1 to obtain output data.

[0106] In step S750, based on the output results of the convolutional neural network model 1, suspected fraud numbers are extracted.

[0107] In an example embodiment, based on the output data, the probability of the classification of the target number is determined; based on the probability of the classification, the classification label identification code corresponding to the target number is determined; the classification label identification code is matched with the classification label name in the classification label library, where the classification label library stores the identification codes and names of multiple classification labels, to obtain the classification label name corresponding to the target number. According to the classification label name corresponding to the target number, the suspected fraud number can be extracted.

[0108] In step S760, the fraudulent number and channel information are output.

[0109] In the example embodiment, the fraudulent number and channel information are stored and pushed to other platforms. Specifically, according to the system settings, the final output data is pushed to the corresponding storage table of the database; then pushed to other platforms according to the settings

[0110] Figure 8FIG. 1 is a schematic block diagram of a number recognition model training device provided according to some embodiments of the present invention. Figure 8 As shown, the number recognition model training device 800 includes: a data acquisition module 810, a sample data feature generation module 820 and a model training module 830. The data acquisition module is used to acquire the original data and corresponding classification labels of multiple numbers within a predetermined time period, and the original data includes service data and signaling data; the sample data feature generation module is used to generate sample data features corresponding to each number based on the original data of each number, and the sample data features include service data features and signaling data features; the model training module is used to train the number recognition model based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold.

[0111] In some embodiments of the present invention, based on the above scheme, the sample data feature generation module includes: a multi-layer matrix data generation unit, which is used to generate multi-layer matrix data corresponding to the original data of the number based on the timing of a predetermined time period, and the multi-layer matrix data includes: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix and a location data layer matrix; a data normalization unit, which is used to convert each layer of matrix data of the multi-layer matrix data into a floating point number, and perform normalization processing to obtain sample data features corresponding to the original data of the number.

[0112] In some embodiments of the present invention, based on the above scheme, the multi-layer matrix data generation unit includes: a data row generation unit, which is used to generate data rows of matrix data of the corresponding data layer based on the data of the corresponding data layer per unit time of a predetermined time period, and the corresponding data layer is one of a business data layer, a signaling call data layer, an Internet traffic data layer, and a location data layer; a matrix data generation unit, which is used to generate matrix data of the corresponding data layer based on the data rows of the corresponding data layer per unit time according to the timing of the predetermined time period.

[0113] In some embodiments of the present invention, based on the above scheme, the model training module includes: a data sample classification unit, which is used to divide the sample data features into training data samples and test data samples according to a predetermined ratio; a model training unit, which is used to train the number recognition model through a gradient optimizer based on the training data samples and the corresponding classification labels; a parameter adjustment unit, which is used to adjust the parameters of the number recognition model based on the test data samples and the corresponding classification labels.

[0114] Fig. 9 FIG. 1 is a schematic block diagram of a number recognition device provided according to some embodiments of the present invention. Fig. 9As shown, the group recommendation device 900 includes: a data acquisition module 910, a production data feature generation module 920, a data input module 930 and a classification label determination module 940. The data acquisition module is used to acquire the original data of the target number, and the original data includes service data and signaling data; the production data feature generation module is used to generate production data features corresponding to the target number based on the original data of the target number, and the production data features include service data features and signaling data features; the data input module is used to input the production data features into the number recognition model to obtain output data; the classification label determination module is used to determine the classification label corresponding to the target number based on the output data.

[0115] In some embodiments of the present invention, based on the above scheme, the classification label determination module includes: a classification probability determination unit, which is used to determine the probability of classification of the target number based on the output data; an identification code determination unit, which is used to determine the classification label identification code corresponding to the target number based on the classification probability; a label name matching unit, which is used to match the classification label identification code with the classification label name in the classification label library, which stores the identification codes and names of multiple classification labels.

[0116] Furthermore, an embodiment of the present application also provides an electronic device, Fig.10 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown in FIG. Fig.10 As shown, the device includes: a memory 1001, a processor 1002, a bus 1003 and a communication interface 1004. The memory 1001, the processor 1002 and the communication interface 1004 communicate via the bus 1003, and the communication interface 1004 may include an input and output interface, which includes but is not limited to a keyboard, a mouse, a display, a microphone, a loudspeaker, etc.

[0117] Fig.10 In the embodiment, the memory 1001 stores computer executable instructions that can be run on the processor 1002, and the computer executable instructions, when executed by the processor 1002, implement the following process: obtain original data and corresponding classification labels of multiple numbers within a predetermined time period, the original data including business data and signaling data; generate sample data features corresponding to each number based on the original data of each number, the sample data features including business data features and signaling data features; train the number recognition model based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold.

[0118] Optionally, when the computer executable instructions are executed by processor 1002, sample data features corresponding to each number are generated based on the original data of each number, including: based on the timing of a predetermined time period, generating multi-layer matrix data corresponding to the original data of the number, the multi-layer matrix data including: a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix, and a location data layer matrix; converting each layer of matrix data of the multi-layer matrix data into floating point numbers, and normalizing them to obtain sample data features corresponding to the original data of the number.

[0119] Optionally, when the computer executable instructions are executed by the processor 1002, multi-layer matrix data corresponding to the original data of the number is generated based on the timing of a predetermined time period, including: generating data rows of matrix data of the corresponding data layer based on the data of the corresponding data layer per unit time of the predetermined time period, the corresponding data layer being one of a service data layer, a signaling call data layer, an Internet traffic data layer, and a location data layer; generating matrix data of the corresponding data layer based on the data rows of the corresponding data layer per unit time according to the timing of the predetermined time period.

[0120] Optionally, when the computer executable instructions are executed by processor 1002, the number recognition model is a convolutional neural network model, and the number recognition model is trained based on sample data features and classification labels corresponding to each number, including: dividing the sample data features into training data samples and test data samples according to a predetermined ratio; training the number recognition model through a gradient optimizer based on the training data samples and the corresponding classification labels; and adjusting the parameters of the number recognition model based on the test data samples and the corresponding classification labels.

[0121] Optionally, when the computer executable instructions are executed by processor 1002, the following process is implemented: obtaining original data of the target number, the original data including service data and signaling data; generating production data features corresponding to the target number based on the original data of the target number, the production data features including service data features and signaling data features; inputting the production data features into a number recognition model to obtain output data; and determining a classification label corresponding to the target number based on the output data.

[0122] Optionally, when the computer executable instructions are executed by the processor 1002, the classification label corresponding to the target number is determined based on the output data, including: determining the probability of classification of the target number based on the output data; determining the classification label identification code corresponding to the target number based on the probability of classification; matching the classification label identification code with the classification label name in a classification label library, which stores identification codes and names of multiple classification labels.

[0123] According to the technical solution of the embodiment of the present invention, on the one hand, the original data and corresponding classification labels of multiple numbers within a predetermined time period are obtained, and the original data includes business data and signaling data, so that the business data and various signaling data for training the model can be obtained; on the other hand, sample data features corresponding to each number are generated based on the original data of each number, and the sample data features include business data features and signaling data features, so that various sample feature data of the number can be efficiently generated without manual analysis and extraction; on the other hand, the number recognition model is trained based on the sample data features and classification labels corresponding to each number, so that the loss function corresponding to the number recognition model is less than a predetermined threshold, and a number recognition model for extracting and identifying fraudulent numbers can be trained. Fraudulent numbers can be identified in advance through this model, which can effectively reduce telephone fraud.

[0124] The electronic device in the embodiment of the present application can implement each process of the aforementioned number recognition model training method or number recognition method, and achieve the same effects and functions, which will not be repeated here.

[0125] Furthermore, an embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned number recognition model training method or number recognition method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0126] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0127] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the method described in the above-mentioned embodiment of the present invention.

[0128] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A number identification method, characterized in that: include: Acquire original data of the target number, the original data including service data and signaling data, the service data including account opening information and network access channel; generating production data features corresponding to the target number based on the original data of the target number, the production data features including service data features and signaling data features; Inputting the production data features into a number recognition model for processing to obtain output data, wherein the number recognition model is a convolutional neural network model, the convolutional neural network model includes a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer, the convolution kernel size of the first convolutional layer is 3x3, the convolution kernel size of the second convolutional layer is 5x5, the convolution kernel size of the third convolutional layer is matrix data column / 2xmatrix data row / 2, and the convolution kernel size of the fourth convolutional layer is 1x1; Determine the probability of the classification of the target number based on the output data, determine the classification label identification code corresponding to the target number based on the classification probability, and match the classification label identification code with the classification label name in the classification label library to determine the classification label corresponding to the target number; The training method of the number recognition model includes: Obtaining raw data of multiple numbers within a predetermined time period and corresponding classification labels; Based on the data of the corresponding data layer per unit time of the predetermined time period, data rows of matrix data of the corresponding data layer are generated, and the corresponding data layer includes a service data layer, a signaling call data layer, an Internet traffic data layer and a location data layer; according to the time sequence of the predetermined time period, based on the data rows of the corresponding data layer per unit time, multi-layer matrix data corresponding to the original data of each number is generated, and each layer of matrix data of the multi-layer matrix data is converted into a floating point number, and normalized to obtain sample data features corresponding to the original data of each number, wherein the multi-layer matrix data includes a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix and a location data layer matrix, the service data layer matrix is ​​the first layer, the columns of the service data matrix layer are preset fixed values, and the sample data features include service data features and signaling data features; The sample data features corresponding to each of the numbers are divided into training data samples and test data samples according to a predetermined ratio, the number recognition model is trained through a gradient optimizer based on the training data samples and the corresponding classification labels, and the parameters of the number recognition model are adjusted based on the test data samples and the corresponding classification labels so that the loss function corresponding to the number recognition model is less than a predetermined threshold, wherein the predetermined ratio is 7:

3.

2. The method according to claim 1, characterized in that The classification label library stores identification codes and names of various classification labels.

3. A number recognition device, characterized in that: include: A data acquisition module, used to acquire the original data of the target number, wherein the original data includes service data and signaling data, wherein the service data includes account opening information and network access channels; A production data feature generation module, configured to generate production data features corresponding to the target number based on the original data of the target number, wherein the production data features include service data features and signaling data features; A data input module, used for inputting the production data features into a number recognition model for processing to obtain output data, wherein the number recognition model is a convolutional neural network model, and the convolutional neural network model includes a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer, the convolution kernel size of the first convolutional layer is 3x3, the convolution kernel size of the second convolutional layer is 5x5, the convolution kernel size of the third convolutional layer is matrix data column / 2xmatrix data row / 2, and the convolution kernel size of the fourth convolutional layer is 1x1; A classification label determination module, used to determine the probability of the classification of the target number based on the output data, determine the classification label identification code corresponding to the target number based on the classification probability, and match the classification label identification code with the classification label name in the classification label library to determine the classification label corresponding to the target number; The training device of the number recognition model comprises: A data acquisition module, used to acquire the original data of multiple numbers within a predetermined time period and corresponding classification labels, wherein the original data includes service data and signaling data, and the service data includes account opening information and network access channels; A sample data feature generation module is used to generate data rows of matrix data of corresponding data layers based on data of corresponding data layers per unit time of the predetermined time period, wherein the corresponding data layers include a service data layer, a signaling call data layer, an Internet traffic data layer and a location data layer; according to the time sequence of the predetermined time period, based on the data rows of corresponding data layers per unit time, generate multi-layer matrix data corresponding to the original data of each number, convert each layer of matrix data of the multi-layer matrix data into floating point numbers, and perform normalization processing to obtain sample data features corresponding to the original data of each number, wherein the multi-layer matrix data includes a service data layer matrix, a signaling call data layer matrix, an Internet traffic data layer matrix and a location data layer matrix, the service data layer matrix is ​​the first layer, the columns of the service data matrix layer are preset fixed values, and the sample data features include service data features and signaling data features; A model training module is used to divide the sample data features corresponding to each of the numbers into training data samples and test data samples according to a predetermined ratio, train the number recognition model through a gradient optimizer based on the training data samples and the corresponding classification labels, and adjust the parameters of the number recognition model based on the test data samples and the corresponding classification labels so that the loss function corresponding to the number recognition model is less than a predetermined threshold, wherein the predetermined ratio is 7:

3.

4. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the number recognition method according to claim 1 or 2 when executed by the processor.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the number recognition method according to claim 1 or 2 is implemented.