Intermediate number call risk identification method and device, electronic equipment and storage medium

Abstract

The present application relates to a middle number call risk identification method and device, electronic equipment and storage medium, the method comprises: obtaining the communication data of the middle number in the call process, and preprocessing the communication data to extract the mel spectrum feature in the communication data. Extract the number attribute feature and the call entity feature from the preprocessed communication data, and construct a risk level model based on the random forest algorithm to output the risk level corresponding to the middle number of the risk level model. The mel spectrum feature is used as the input of the CNN model to capture the local frequency band feature in the mel spectrum feature, and the BiLSTM model is called to identify the risk probability of the call process based on the local frequency band feature. Based on the risk identification rule, the risk level and the risk probability are evaluated to obtain the comprehensive risk score corresponding to the communication data. When the comprehensive risk score exceeds the preset range, it is determined that the call process is an abnormal call, which improves the accuracy of call risk identification.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method, device, electronic device and storage medium for determining the risk of an intermediate number call. Background Art

[0002] In the communications sector, intermediary number calls, while a common method of communication, offer convenience but also conceal risky behavior. Some individuals exploit the flexibility of intermediary number calls to engage in various activities that compromise user safety and legal rights, posing significant challenges to communications industry regulation. Due to the diverse scenarios, complex content, and dynamic nature of intermediary number calls, traditional call risk identification methods struggle to accurately identify them. Therefore, integrating number attribute information, voice content characteristics, and service scenario information to construct an efficient risk identification model has become a pressing technical challenge.

[0003] Currently, existing methods for identifying risk in middle-level numbers mostly perform simple keyword detection on the text content of voice messages. Due to the diversity and flexibility of risk terminology, many risky content does not contain obvious keywords. Furthermore, keyword detection after voice-to-text conversion can lead to inaccurate text translation due to factors such as voice quality, the wide variety of dialects, and the presence of noise, resulting in low detection accuracy. Secondly, existing call risk assessment models often use fixed feature weights or rules, making it difficult to adapt to the changing call characteristics of different industry scenarios. For example, normal calls in the financial industry may contain a large number of words related to monetary amounts. If existing models uniformly identify such words as high-risk features, they will result in a large number of false positives. Furthermore, existing technologies lack an effective mechanism to comprehensively consider multiple factors, including number information, voice content features, and call scenario, making it difficult to comprehensively and accurately determine whether middle-level number calls are risky. Finally, existing technologies lack a systematic set of comprehensive identification rules, making it difficult to effectively integrate user risk level, voice risk probability, and scenario matching, resulting in poor reliability of the final risk assessment results. For example, when the user risk level is low but the voice risk probability is high, the existing risk identification method is difficult to determine a more reasonable assessment result.

[0004] In summary, the traditional method of identifying call risk with an intermediate number has low flexibility in identifying call risk due to the lack of comprehensive identification rules and the limitations of fixed features and simple keywords. At the same time, the accuracy and comprehensiveness of call risk identification need to be further improved. Summary of the Invention

[0005] The present invention aims to provide a method, device, electronic device and storage medium for determining the risk of intermediate number calls to solve the deficiencies in the prior art. The technical problem to be solved by the present invention is achieved through the following technical solutions.

[0006] The present invention provides a method for determining the risk of an intermediate number call, the method comprising:

[0007] Acquire communication data of the intermediate number during the call, and preprocess the communication data to extract Mel spectrum features in the communication data;

[0008] Extracting number attribute features and call entity features from the preprocessed communication data, and building a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model;

[0009] The mel spectrum features are used as input to the CNN model to capture local frequency band features in the mel spectrum features, and the BiLSTM model is called to identify the risk probability of the call process based on the local frequency band features;

[0010] Evaluate the risk level and risk probability based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and determine that the call process of the intermediate number is an abnormal call when the comprehensive risk score exceeds a preset range;

[0011] The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

[0012] In one embodiment, obtaining communication data of the intermediate number during a call and preprocessing the communication data to extract mel spectrum features in the communication data includes:

[0013] Cleaning and standardizing the call number information to remove invalid and abnormal data in the call number information;

[0014] Extracting entity data based on the reporting speech template from the call number information, and constructing an entity data sequence based on the reporting speech template based on the extracted different entity data;

[0015] Converting the call voice data into a digital signal, and performing framing and windowing processing on the digital signal to obtain a voice stream sequence;

[0016] The call number information includes user information, number location information and call time, and the standardization processing includes encoding and converting the user information and location information, and converting the call time of the intermediate number into time series features.

[0017] In one embodiment, obtaining communication data of the intermediate number during a call and preprocessing the communication data to extract mel spectrum features in the communication data further includes:

[0018] Decomposing the noisy call voice data into noise coefficients of different frequency bands through wavelet transform, and performing soft threshold processing on the high-frequency noise coefficients whose noise frequency exceeds a first threshold to obtain a voice signal reconstructed by inverse wavelet transform;

[0019] Extracting the Mel spectrum feature from the speech signal through a Mel filter bank;

[0020] Call the BERT word segmenter to segment the customer's report text to obtain a word segmentation sequence, and input the word segmentation sequence into the BERT model for encoding to obtain the hidden state vector of the context corresponding to each word segmentation;

[0021] The softmax function is called to calculate the probability distribution of different entity labels corresponding to each segmentation, and the predicted entity labels of each segmentation are merged through the decoder to obtain the structured call entity features.

[0022] In one embodiment, the extracting number attribute features and call entity features from the preprocessed communication data, and constructing a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model includes:

[0023] Extracting number attribute features and call entity features from the preprocessed communication data, and combining the number attribute features and call entity features into a feature vector according to feature dimensions;

[0024] Selecting multiple feature elements from the feature vector and determining the optimal split feature based on the Gini index to construct a decision tree, wherein the random forest algorithm is composed of multiple decision trees;

[0025] The decision tree is called to output a risk level prediction result based on the feature vector, and a probability distribution of the risk level is obtained according to multiple risk level prediction results output by multiple decision trees through a voting mechanism to determine the risk level of the intermediate number.

[0026] In one of the embodiments, the Mel spectrum features are taken as inputs of the CNN model to capture local frequency band features in the Mel spectrum features, and a BiLSTM model is called to identify the risk probability of the call process based on the local frequency band features, including:

[0027] The Mel spectrum features are input into the CNN model to capture local frequency band features in the Mel spectrum features through the convolution layer of the CNN model, and the local frequency band features are processed by the dimension reduction through the pooling layer to obtain a feature sequence;

[0028] The feature sequence is input into the BiLSTM model to call the BiLSTM model to capture forward and reverse timing features in the speech stream sequence, and output bidirectional fusion features through vector splicing;

[0029] Based on the bidirectional fusion features, a full connection layer is called to output the risk probability of the current call of the intermediate number through the output layer weight and the sigmoid activation function.

[0030] In one of the embodiments, the Mel spectrum features are taken as inputs of the CNN model to capture local frequency band features in the Mel spectrum features, and a BiLSTM model is called to identify the risk probability of the call process based on the local frequency band features, and then further including:

[0031] Abnormal call voice data are obtained, and the CNN model and the BiLSTM model are fused and trained based on the abnormal call voice data;

[0032] The model parameters are adjusted through the back propagation method, and the model parameters are optimized through a binary cross-entropy loss function to obtain a trained CNN-BiLSTM composite model.

[0033] In one of the embodiments, the risk level and the risk probability are evaluated based on the preset risk discrimination rule to obtain a comprehensive risk score corresponding to the communication data, and when the comprehensive risk score exceeds a preset range, it is determined that the call process of the intermediate number is an abnormal call, including:

[0034] The current communication data of the intermediate number are input into the trained CNN-BiLSTM composite model to output a risk level and a risk probability corresponding to the current communication data;

[0035] A risk score fusion formula is called to calculate the comprehensive risk score corresponding to the current communication data based on the weight coefficients corresponding to the risk level and the risk probability, respectively, to determine whether the comprehensive risk score exceeds a preset range;

[0036] When the comprehensive risk score exceeds a preset range, the risk level of the current call process is divided according to the preset risk identification rules, and the risk level division result is fed back to the user end.

[0037] The present invention further provides a device for determining the risk of an intermediate number call, for implementing any of the above methods for determining the risk of an intermediate number call, the device comprising:

[0038] A data preprocessing module is used to obtain communication data of the intermediate number during the call and preprocess the communication data to extract Mel spectrum features in the communication data;

[0039] a risk level output module, configured to extract number attribute features and call entity features from the preprocessed communication data, and construct a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model;

[0040] a risk probability output module, configured to use the mel spectrum features as input to a CNN model to capture local frequency band features in the mel spectrum features, and to call a BiLSTM model to identify the risk probability of the call process based on the local frequency band features;

[0041] a comprehensive identification module, configured to evaluate the risk level and risk probability based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and to determine that the call process of the intermediate number is an abnormal call when the comprehensive risk score exceeds a preset range;

[0042] The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

[0043] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements any of the above-mentioned methods for determining the risk of intermediate number calls.

[0044] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for determining the risk of an intermediate number call.

[0045] The aforementioned method, device, electronic device, and storage medium for identifying the risk of calls involving middle-number numbers obtain communication data from the middle-number during a call and preprocess the data to extract mel-spectrogram features from the data. Number attribute features and call entity features are then extracted from the preprocessed data. A risk rating model is constructed based on a random forest algorithm, which then uses the risk rating output by the risk rating model to determine the risk level corresponding to the middle-number number. The mel-spectrogram features are then used as input to a CNN model to capture the local frequency band characteristics within the mel-spectrogram features. A BiLSTM model is then used to identify the risk probability of the call based on these local frequency band characteristics. Finally, the risk level and risk probability are evaluated based on pre-set risk identification rules to obtain a comprehensive risk score corresponding to the communication data. If the comprehensive risk score exceeds a pre-set range, the call involving the middle-number number is deemed abnormal. The present invention constructs a multi-dimensional data collection system for number information and call voice, combines the number risk level classification based on the random forest algorithm and the risk identification of call voice data by the CNN+Bilstm fusion model, and finally realizes the fusion of multi-model collaboration and risk judgment rules. It not only improves the flexibility and accuracy of risk judgment in the call process of the intermediate number, but also improves the comprehensiveness of the risk judgment results of the intermediate number based on the comprehensive scoring of the number risk level and the call risk probability. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 This is a flow chart of the method for determining the risk of calls with intermediate numbers provided by the present invention;

[0048] Figure 2 Schematic diagram of the architecture of the middle number call risk identification method for identifying the middle number call risk in a specific embodiment of the present invention;

[0049] Figure 3 This is a second flow chart of the method for determining the risk of calls with intermediate numbers provided by the present invention;

[0050] Figure 4 This is a third flow chart of the method for determining the risk of calls with intermediate numbers provided by the present invention;

[0051] Figure 5 This is a fourth flow chart of the method for determining the risk of a call with an intermediate number provided by the present invention;

[0052] Figure 6 This is a fifth flow chart of the method for determining the risk of a call with an intermediate number provided by the present invention;

[0053] Figure 7 This is a sixth flow chart of the method for determining the risk of an intermediate number call provided by the present invention;

[0054] Figure 8 This is the seventh flow chart of the method for determining the risk of calls with intermediate numbers provided by the present invention;

[0055] Figure 9 This is a schematic diagram of the structure of the device for determining the risk of calls with intermediate numbers provided by the present invention;

[0056] Figure 10 This is a diagram of the internal structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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 making creative efforts shall fall within the scope of protection of the present invention.

[0058] The following combination Figures 1 to 10 The present invention describes the method, device, electronic device and storage medium for determining the risk of an intermediate number call.

[0059] like Figure 1 As shown, in one embodiment, a method for determining the risk of an intermediate number call includes the following steps:

[0060] Step S110 , obtaining communication data of the intermediate number during the call, and preprocessing the communication data to extract Mel spectrum features in the communication data.

[0061] Among them, communication data includes call number information and call voice data. Preprocessing includes standardization of call number information, extraction of call entity features of the customer reporting script template, conversion of call voice data and Mel spectrum feature extraction.

[0062] Specifically, the call number information and call voice information of the middle number to be identified during the call are obtained, and the call number information is standardized. At the same time, the call entity features of the customer's reporting speech template are extracted, and the call voice information is converted and Mel spectrum features are extracted.

[0063] Combine Figure 2As shown, in a specific embodiment, the method for determining the risk of a call with an intermediate number provided by the present invention is as follows: Figure 2 The illustrated implementation of the risk assessment architecture for call processing with an intermediary number includes a data collection layer, a feature processing layer, a model building layer, and a comprehensive assessment layer. The data collection layer collects relevant data on intermediary number calls, including number information and voice data.

[0064] During the data collection process, we collect data related to calls with intermediate numbers, including number information and voice data. Number information includes information about the company using the number, location, time of day, plan type, call patterns, daily call frequency, and customer behavior scores. It also includes usage scenarios and language used as reported by customers. Voice data is the audio content of the call.

[0065] The feature processing layer is used to preprocess and extract features from the data collected by the data collection layer. This includes entity extraction and feature vector construction. The data preprocessing process is divided into number information preprocessing and voice data preprocessing:

[0066] (1) Number information preprocessing: Standardize and encode enterprise information, such as mapping industry types to 20 first-level industry codes; geocode location information; and convert call time periods into time features. Collected number information is cleaned and standardized to remove invalid data and outliers, and then data in different formats is converted to a standard format. At the same time, call entity features are extracted from the entity features of the customer reporting script template to facilitate input and processing of subsequent models.

[0067] The customer report script is segmented by BERT's tokenizer, and special [CLS] and [SEP] tags are automatically added to mark the beginning and end of the sentence, mapped to ID ( ):

[0068]

[0069] Where X is the input text.

[0070] The segmented sequence is input into the Transformer layer of the BERT model for encoding, and the context-related hidden state vector H corresponding to each token is obtained. The expression is:

[0071]

[0072] The probability distribution of each token belonging to different entity labels is calculated through the softmax function. Finally, the predicted labels of each token are optimized and merged using the decoding algorithm to obtain the structured entity feature results, which are expressed as follows:

[0073]

[0074]

[0075]

[0076] Where h is the latent vector of each token, is the label probability distribution for each token, is the predicted label for each token, is the predicted label of the i-th token, W is the classification layer weight, b is the bias, For the merged structured entity list, B-type and I-type are two key tags in the BIO annotation system, which are used to identify the starting and internal positions of the entity. represents the score of transferring from the i-th label to the j-th label.

[0077] (2) Speech data preprocessing: Speech data is subjected to noise reduction and sampling rate conversion to improve speech quality. Then, the speech data is converted into a digital signal and framed and windowed. The frame length is set to 25ms and the frame shift is 10ms to obtain a speech stream sequence. The audio data is then subjected to noise reduction using the wavelet threshold denoising algorithm, which is expressed as:

[0078]

[0079] Where, is a noisy speech signal, For pure speech signals, is Gaussian white noise, is the noise standard deviation, For time point.

[0080] Afterwards, the noisy speech signal is decomposed into coefficients of different frequency bands through wavelet transform, and the high-frequency coefficients are processed by soft thresholding. The expression is:

[0081]

[0082] Where, is the wavelet coefficient, is the threshold value ( , is the signal length), is the processed wavelet coefficient, Represents others. After the above processing, the speech signal is reconstructed by inverse wavelet transform.

[0083] Finally, the speech signal reconstructed by inverse wavelet transform is framed and the Mel spectrum is obtained by Mel filter bank to extract Mel spectrum features as the input of subsequent models. The expression is:

[0084]

[0085] Where, is the audio frequency (Hz).

[0086] Step S120, extracting number attribute features and call entity features from the preprocessed communication data, and building a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model.

[0087] Specifically, the number attribute features and call entity features are extracted from the communication data preprocessed in step S110, and a random algorithm forest is used to construct a risk level model based on the extracted number attribute features and call entity features. The constructed risk level model is used to output the risk level corresponding to the intermediate number.

[0088] It should be noted that the number attribute features are the attribute information of the call number itself, such as the identity information of the number registered user and the number's location, and the call entity features are the relevant nouns or keywords mentioned in the voice data during the call.

[0089] Combine Figure 2 As shown in the following example, in a specific embodiment of the method for identifying call risk using an intermediate number, the model building layer is used to construct a customer risk level model and a voice risk identification model, respectively outputting a risk level and risk probability. During the model building process, the customer risk level model is constructed using the random forest algorithm based on preprocessed number information. Random forest is an ensemble learning algorithm that effectively improves model accuracy and stability by constructing multiple decision trees and integrating their prediction results.

[0090] First, perform feature selection to convert the number information and entity list into model input features, including:

[0091] (1) Number attribute characteristics: industry to which the enterprise belongs (one-hot code), location (latitude and longitude code), call time characteristics (whether it is peak time, weekday / weekend, etc.), package type (category code), calling frequency (normalized), average call duration (normalized), customer historical behavior score (normalized value).

[0092] (2) Entity characteristics: the existence of entity types (e.g., whether it contains ID number, amount, etc., which forms a binary feature), the number of sensitive entities (e.g., the number of occurrences of amount entities), and the matching degree between entities and industries (e.g., whether the frequency of amount entities in the financial industry is within a reasonable range).

[0093] The above number attribute features and entity features are combined to form a corresponding feature vector, such as X=[x1,x2,...,xk], where k is the feature dimension.

[0094] After that, the model is constructed. Since the random forest is composed of multiple decision trees, the construction of each decision tree is explained below. First, multiple samples are randomly selected from the training set constructed by the above-mentioned collected data through bootstrap sampling to form a sub-training set. For each node in the decision tree, s features are randomly selected from the k features in the feature vector X=[x1,x2,...,xk] ( ), and select the optimal split feature based on the Gini index.

[0095] Specifically, for a certain value v of feature x, the Gini index after splitting is Expressed as:

[0096]

[0097] Where, , , (The proportion of category c in D).

[0098] In this embodiment, the decision tree grows to the maximum depth without pruning. For the feature vector X, each decision tree outputs a risk level prediction result, and the final risk level probability distribution is obtained through the voting mechanism. :

[0099]

[0100] Where, is the prediction result of the mth decision tree, is the indicator function, .

[0101] According to the above probability distribution, the user risk level can be determined: Risk Level .

[0102] In this example, we assume that the number of decision trees is 100, the maximum depth is 10, and the minimum number of sample splits is 2. The model is then trained using the training set, and model performance is optimized by adjusting model parameters such as the number of decision trees and the maximum depth. Furthermore, the trained model is evaluated using the test set, and metrics such as accuracy, precision, and recall are calculated. The results show that the constructed model achieves an accuracy of over 90%, with both precision and recall exceeding 85%, demonstrating the good performance of the constructed model.

[0103] In step S130 , the mel spectrum features are used as input to the CNN model to capture the local frequency band features in the mel spectrum features, and the BiLSTM model is called to identify the risk probability of the call process based on the local frequency band features.

[0104] Specifically, the mel-spectrogram features (represented as mel-spectrograms) extracted in step S120 are used as input to the CNN model. The CNN model is called to capture the more significant local frequency band features in the mel-spectrogram features. The captured local frequency band features are then input into the BiLSTM model, which is called to output the risk probability corresponding to the current call process based on the input local frequency band features.

[0105] Combine Figure 2 As shown, in a specific embodiment, the method for distinguishing the risk of middle-number calls provided by the present invention adopts the CNN+BiLSTM model to perform risk probability identification on the preprocessed Mel spectrum during the process of identifying the risk probability of call voice.

[0106] First, the preprocessed Mel-spectrogram features are fed into the CNN model. Convolutional and pooling layers are used to extract features and reduce the dimensionality of the speech stream. During this process, the CNN automatically learns local features within the speech stream and reduces the feature dimensionality through pooling, improving model efficiency. Convolution captures local, significant features in the spectrogram (e.g., energy in unusual frequency bands). The first convolution layer outputs:

[0107]

[0108] Where, is a 3×3 convolution kernel, is the bias term, In this example, after three layers of convolution (the convolution kernel sizes are 3×3, 3×3, and 2×2, respectively) and maximum pooling (step size 2), the output feature sequence is (T is the time step, D is the feature dimension).

[0109] Afterwards, the features extracted by CNN are input into the BiLSTM model. BiLSTM can simultaneously capture the forward and reverse time sequence features of the speech stream sequence, obtain the time sequence dependency (such as the logical coherence of the speech), and effectively utilize the sequence relationship of the speech to improve the accuracy of risk probability identification. Among them, BiLSTM bidirectional fusion feature The calculation formula is as follows:

[0110]

[0111] Where, represents the forward LSTM hidden state, represents the reverse LSTM hidden state, represents vector concatenation, (d is the dimension of the unidirectional LSTM hidden layer).

[0112] Finally, the risk probability of the call is output through the fully connected layer:

[0113]

[0114] Where, is the bidirectional LSTM output of the last time step ( ), is the output layer weight, is the bias, is the sigmoid activation function, (A larger value indicates a higher likelihood of risk).

[0115] In this embodiment, the CNN+BiLSTM model is trained using labeled risky speech data, and the model parameters are adjusted through the back propagation algorithm to optimize the model performance. The binary cross entropy loss function is used to optimize the model parameters:

[0116]

[0117] Where, is the number of training samples, is the sample label (1 for abnormality, 0 for normal), is the risk probability prediction value of the nth sample.

[0118] In this example, a batch size of 32, a learning rate of 0.001, and 50 training rounds are assumed. During training, a cross-entropy loss function and the Adam optimizer are used, along with dropout technology at a dropout rate of 0.2, to effectively prevent model overfitting. The training results show that the constructed model achieves an accuracy of over 88%, with both precision and recall rates exceeding 85%, demonstrating good performance.

[0119] Step S140 , evaluate the risk level and risk probability based on the preset risk identification rules to obtain a comprehensive risk score corresponding to the communication data, and when the comprehensive risk score exceeds a preset range, determine that the call process of the intermediate number is an abnormal call.

[0120] Specifically, based on the pre-set risk identification rules, the risk level of the intermediate number obtained in the above steps and the risk probability of the call process are evaluated to obtain the corresponding comprehensive risk score of this call process. When the comprehensive risk score exceeds the set safety range, the call process of the intermediate number is determined to be an abnormal call.

[0121] Combine Figure 2 As shown, in a specific embodiment, the present invention provides a method for identifying call risk using an intermediate number. The comprehensive evaluation layer is used to output a comprehensive assessment result based on preset rules, combining the enterprise risk level, voice risk probability, and scenario matching. During the comprehensive evaluation process, the output of the customer risk level model and the risk probability identification results of the CNN+BiLSTM model are combined with the scenario information of the current call, and a comprehensive evaluation is performed according to preset evaluation rules to determine whether the current call is risky. The evaluation rules are shown in Table 1:

[0122] Table 1

[0123]

[0124] In this embodiment, the comprehensive risk score is calculated by combining the number risk level R∈[1,5] (1 is very low risk, 5 is very high risk) and the voice risk probability P∈[0,1], and the expression is:

[0125]

[0126] Where, is the weight coefficient, which can be optimized through cross-validation (usually 0.4 to 0.6 in this case).

[0127] Finally, based on the range of the comprehensive risk score (Score), the call is determined to be risk 1, risk 2, or normal. Experimental results show that this example achieved 92% risk identification accuracy, 90% precision, 93% recall, and an F1 score of 91.5%. This demonstrates a clear advantage in risk identification accuracy. Therefore, by integrating multi-dimensional data, optimizing the model architecture, and establishing a comprehensive evaluation mechanism, this example achieves efficient and accurate risk assessment of calls with intermediate numbers, demonstrating significant practical application value and potential for widespread adoption.

[0128] The aforementioned method for identifying the risk of calls involving middle-number numbers obtains communication data from the middle-number call and preprocesses it to extract mel-spectrogram features. It then extracts number attribute features and call entity features from the preprocessed data. A risk rating model is constructed based on a random forest algorithm, which uses the risk rating output by the risk rating model to determine the risk level for each middle-number call. The mel-spectrogram features are then used as input to a CNN model to capture the local frequency band characteristics within the mel-spectrogram features. A BiLSTM model is then used to identify the risk probability of the call based on these local frequency band characteristics. Finally, the risk level and risk probability are evaluated based on pre-set risk identification rules to obtain a comprehensive risk score corresponding to the communication data. If the comprehensive risk score exceeds a preset range, the call involving the middle-number number is deemed abnormal. This method constructs a multi-dimensional data collection system for number information and call voice, combines the number risk level classification based on the random forest algorithm and the risk identification of call voice data using the CNN+Bilstm fusion model, and ultimately achieves the integration of multi-model collaboration and risk judgment rules. It not only improves the flexibility and accuracy of risk judgment in the call process of the intermediate number, but also improves the comprehensiveness of the risk judgment results of the intermediate number based on the comprehensive scoring of the number risk level and the call risk probability.

[0129] like Figure 3 As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention, step S110 specifically includes the following steps:

[0130] Step S111 , cleaning and standardizing the call number information to remove invalid data and abnormal data in the call number information.

[0131] Step S112: extracting entity data based on the reporting speech template from the call number information, and constructing an entity data sequence based on the reporting speech template based on the extracted different entity data.

[0132] Step S113: convert the call voice data into a digital signal, and perform frame division and windowing processing on the digital signal to obtain a voice stream sequence.

[0133] Among them, the call number information includes user information, number location information and call time. The standardization processing includes encoding and converting the user information and location information, and converting the call time of the intermediate number into time series features.

[0134] like Figure 4 As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention, step S110 specifically further includes the following steps:

[0135] Step S114 , decomposing the noisy call voice data into noise coefficients of different frequency bands by wavelet transform, and performing soft threshold processing on the high-frequency noise coefficients whose noise frequencies exceed a first threshold, to obtain a voice signal reconstructed by inverse wavelet transform.

[0136] Step S115 : extracting mel spectrum features from the speech signal through a mel filter bank.

[0137] Step S116: Call the BERT word segmenter to segment the customer report language to obtain a word segmentation sequence, and input the word segmentation sequence into the BERT model for encoding to obtain a hidden state vector of the context corresponding to each word segmentation.

[0138] In step S117, the softmax function is called to calculate the probability distribution of different entity labels corresponding to each word segment, and the predicted entity labels of each word segment are merged through the decoder to obtain structured call entity features.

[0139] Specifically, the customer report script is segmented through BERT's tokenizer, and special [CLS] and [SEP] tags are automatically added to mark the beginning and end of the sentence. The segmented sequence is input into the BERT model for encoding, and the context-related hidden state vector corresponding to each token is obtained. The probability distribution of each token belonging to different entity labels is calculated through the softmax function. Finally, the predicted label of each token is optimized and merged using a decoding algorithm to obtain a structured entity feature result.

[0140] like Figure 5 As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention, step S120 specifically includes the following steps:

[0141] Step S121 : extracting number attribute features and call entity features from the preprocessed communication data, and combining the number attribute features and call entity features into a feature vector according to feature dimensions.

[0142] Step S122: Select multiple feature elements from the feature vector and determine the optimal split feature based on the Gini index to construct a decision tree. The random forest algorithm is composed of multiple decision trees.

[0143] Step S123, calling the decision tree to output the risk level prediction result based on the feature vector, and obtaining the probability distribution of the risk level according to the multiple risk level prediction results output by multiple decision trees through a voting mechanism to determine the risk level of the middle number.

[0144] like Figure 6As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention, step S130 specifically includes the following steps:

[0145] In step S131 , the Mel spectrum features are input into the CNN model to capture the local frequency band features in the Mel spectrum features through the convolution layer of the CNN model, and the local frequency band features are subjected to dimensionality reduction processing through the pooling layer to obtain a feature sequence.

[0146] Step S132: Input the feature sequence into the BiLSTM model to call the BiLSTM model to capture the forward and reverse time sequence features in the speech stream sequence, and output the bidirectional fusion features through vector splicing.

[0147] Step S133: Based on the bidirectional fusion features, the fully connected layer is called to combine the output layer weights and the sigmoid activation function to output the risk probability of the current call of the middle number.

[0148] like Figure 7 As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention further includes the following steps after step S130:

[0149] Step S710: Acquire abnormal call voice data, and perform fusion training on the CNN model and the BiLSTM model based on the abnormal call voice data.

[0150] Step S720: Adjust the model parameters through the back-propagation method, and call the binary cross entropy loss function to optimize the model parameters to obtain the trained CNN-BiLSTM composite model.

[0151] like Figure 8 As shown, in one embodiment, the method for determining the risk of an intermediate number call provided by the present invention, step S140 specifically includes the following steps:

[0152] Step S141: Input the current communication data of the intermediate number into the trained CNN-BiLSTM composite model to output the risk level and risk probability corresponding to the current communication data.

[0153] Step S142: The risk score fusion formula is called to calculate the comprehensive risk score corresponding to the current communication data based on the weight coefficients corresponding to the risk level and the risk probability, so as to determine whether the comprehensive risk score exceeds a preset range.

[0154] Step S143: When the comprehensive risk score exceeds a preset range, the current call process is classified into risk levels according to preset risk identification rules, and the risk level classification result is fed back to the user end.

[0155] The intermediate number call risk identification device provided by the present application is described below. The intermediate number call risk identification device described below can be referred to in correspondence with the intermediate number call risk identification method described above.

[0156] As shown in FIG. 10, in one embodiment, an intermediate number call risk identification device includes a data preprocessing module 910, a risk level output module 920, a risk probability output module 930, and a comprehensive identification module 940. Figure 9

[0157] The data preprocessing module 910 is configured to obtain communication data of an intermediate number in a call process, and preprocess the communication data to extract a mel spectrum feature in the communication data.

[0158] The risk level output module 920 is configured to extract a number attribute feature and a call entity feature from the preprocessed communication data, and construct a risk level model based on a random forest algorithm to output a risk level corresponding to the intermediate number.

[0159] The risk probability output module 930 is configured to input the mel spectrum feature into a CNN model to capture a local frequency band feature in the mel spectrum feature, and call a BiLSTM model to identify a risk probability of the call process based on the local frequency band feature.

[0160] The comprehensive identification module 940 is configured to evaluate the risk level and the risk probability based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and determine that the call process of the intermediate number is an abnormal call when the comprehensive risk score exceeds a preset range.

[0161] The communication data includes call number information and call voice data, and the preprocessing includes standardization of the call number information, extraction of the call entity feature, and conversion of the call voice data and extraction of the mel spectrum feature.

[0162] In this embodiment, the data preprocessing module 910 of the intermediate number call risk identification device provided by the present application is specifically configured to:

[0163] clean and standardize the call number information to remove invalid data and abnormal data in the call number information.

[0164] extract entity data based on a report speech template from the call number information, and construct an entity data sequence based on the report speech template based on the extracted different entity data.

[0165] convert the call voice data into a digital signal, and frame and window the digital signal to obtain a speech stream sequence.

[0166] ​The call number information includes user information, number home location information and call time, and the standardization processing includes encoding conversion of the user information and the home location information, and conversion of the call time of the intermediate number into a time sequence feature.

[0167] In the embodiment, the intermediate number call risk identification device provided by the application is provided, and the data preprocessing module 910 is specifically further used for:

[0168] The wavelet transform is used to decompose the call voice data carrying noise into noise coefficients of different frequency bands, and the soft threshold value processing is performed on the high-frequency noise coefficients whose noise frequency exceeds a first threshold value, so as to obtain a voice signal reconstructed by inverse wavelet transform.

[0169] The mel filter bank is used to extract mel spectrum features from the voice signal.

[0170] The BERT tokenizer is called to perform tokenization processing on the customer report speech, to obtain a tokenization sequence, and the tokenization sequence is input into the BERT model for encoding, to obtain a hidden state vector corresponding to the context of each token.

[0171] The softmax function is called to calculate the probability distribution of different entity labels corresponding to each token, and the predicted entity labels of each token are merged by the decoder, to obtain structured call entity features.

[0172] In the embodiment, the intermediate number call risk identification device provided by the application is provided, and the risk level output module 920 is specifically used for:

[0173] The number attribute features and the call entity features are extracted from the preprocessed communication data, and the number attribute features and the call entity features are combined into a feature vector according to the feature dimension.

[0174] A plurality of feature elements are selected from the feature vector, and an optimal split feature is determined based on a Gini index, to construct a decision tree, and the random forest algorithm is composed of a plurality of decision trees.

[0175] The decision tree is called to output a risk level prediction result based on the feature vector, and a voting mechanism is used to obtain a probability distribution of the risk level according to a plurality of risk level prediction results output by the plurality of decision trees, to determine the risk level of the intermediate number.

[0176] In the embodiment, the intermediate number call risk identification device provided by the application is provided, and the risk probability output module 930 is specifically used for:

[0177] The mel spectrum features are input into the CNN model, to capture local frequency band features in the mel spectrum features through a convolutional layer of the CNN model, and to perform dimension reduction processing on the local frequency band features through a pooling layer, to obtain a feature sequence.

[0178] The feature sequence is input into the BiLSTM model to call the BiLSTM model to capture the forward and reverse time sequence features in the speech stream sequence, and output the bidirectional fusion features through vector splicing.

[0179] Based on the bidirectional fusion features, the fully connected layer is called to combine the output layer weights and the sigmoid activation function to output the risk probability of the current call of the middle number.

[0180] In this embodiment, the device for determining the risk of an intermediate number call provided by the present invention further includes a model fusion module for:

[0181] Acquire abnormal call voice data and perform fusion training on the CNN model and BiLSTM model based on the abnormal call voice data.

[0182] The model parameters are adjusted through the back-propagation method, and the binary cross entropy loss function is called to optimize the model parameters to obtain the trained CNN-BiLSTM composite model.

[0183] In this embodiment, the device for determining the risk of an intermediate number call provided by the present invention, the comprehensive determination module 940 is specifically configured to:

[0184] The current communication data of the intermediate number is input into the trained CNN-BiLSTM composite model to output the risk level and risk probability corresponding to the current communication data.

[0185] The risk score fusion formula is called to calculate the comprehensive risk score corresponding to the current communication data based on the weight coefficients corresponding to the risk level and risk probability, so as to determine whether the comprehensive risk score exceeds the preset range.

[0186] When the comprehensive risk score exceeds the preset range, the current call process is divided into risk levels according to the preset risk judgment rules, and the risk level division results are fed back to the user end.

[0187] Figure 10 The following is a schematic diagram of the physical structure of an electronic device. The electronic device may be a smart terminal, and its internal structure diagram may be as follows: Figure 10 As shown. The electronic device includes a processor, an internal memory, and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it implements a method for determining the risk of intermediate number calls, which includes:

[0188] Acquire the communication data of the middle number during the call and pre-process the communication data to extract the Mel spectrum features in the communication data;

[0189] Extract number attribute features and call entity features from the preprocessed communication data, and build a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model;

[0190] The CNN model uses the Mel-spectrogram features as input to capture the local frequency band characteristics in the Mel-spectrogram features. The BiLSTM model is then used to identify the risk probability of the call process based on the local frequency band characteristics.

[0191] The risk level and risk probability are evaluated based on preset risk identification rules to obtain a comprehensive risk score corresponding to the communication data. If the comprehensive risk score exceeds the preset range, the call process of the middle number is determined to be an abnormal call;

[0192] The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

[0193] Those skilled in the art will understand that Figure 10 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the electronic device to which the solution of the present invention is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0194] In another aspect, the present invention further provides a computer storage medium storing a computer program. When the computer program is executed by a processor, a method for determining the risk of an intermediate number call is implemented. The method includes:

[0195] Acquire the communication data of the middle number during the call and pre-process the communication data to extract the Mel spectrum features in the communication data;

[0196] Extract number attribute features and call entity features from the preprocessed communication data, and build a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model;

[0197] The CNN model uses the Mel-spectrogram features as input to capture the local frequency band characteristics in the Mel-spectrogram features. The BiLSTM model is then used to identify the risk probability of the call process based on the local frequency band characteristics.

[0198] The risk level and risk probability are evaluated based on preset risk identification rules to obtain a comprehensive risk score corresponding to the communication data. If the comprehensive risk score exceeds the preset range, the call process of the middle number is determined to be an abnormal call;

[0199] The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

[0200] In yet another aspect, a computer program product or computer program is provided. The computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium. When the processor executes the computer instructions, a method for determining the risk of an intermediate number call is implemented. The method includes:

[0201] Acquire the communication data of the middle number during the call and pre-process the communication data to extract the Mel spectrum features in the communication data;

[0202] Extract number attribute features and call entity features from the preprocessed communication data, and build a risk level model based on the random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model;

[0203] The CNN model uses the Mel-spectrogram features as input to capture the local frequency band characteristics in the Mel-spectrogram features. The BiLSTM model is then used to identify the risk probability of the call process based on the local frequency band characteristics.

[0204] The risk level and risk probability are evaluated based on preset risk identification rules to obtain a comprehensive risk score corresponding to the communication data. If the comprehensive risk score exceeds the preset range, the call process of the middle number is determined to be an abnormal call;

[0205] The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

[0206] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0207] By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0208] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0209] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for determining the risk of an intermediate number call, characterized in that: The method comprises: Acquire communication data of the intermediate number during the call, and preprocess the communication data to extract Mel spectrum features in the communication data; Extracting number attribute features and call entity features from the preprocessed communication data, and building a risk level model based on a random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model; The mel spectrum features are used as input to the CNN model to capture local frequency band features in the mel spectrum features, and the BiLSTM model is called to identify the risk probability of the call process based on the local frequency band features; Evaluate the risk level and risk probability based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and determine that the call process of the intermediate number is an abnormal call when the comprehensive risk score exceeds a preset range; The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

2. The method for determining the risk of an intermediate number call according to claim 1, characterized in that: The acquiring of communication data of the intermediate number during the call and preprocessing the communication data to extract Mel spectrum features in the communication data includes: Cleaning and standardizing the call number information to remove invalid and abnormal data in the call number information; Extracting entity data based on the reporting speech template from the call number information, and constructing an entity data sequence based on the reporting speech template based on the extracted different entity data; Converting the call voice data into a digital signal, and performing framing and windowing processing on the digital signal to obtain a voice stream sequence; The call number information includes user information, number location information and call time, and the standardization processing includes encoding and converting the user information and location information, and converting the call time of the intermediate number into time series features.

3. The method for determining the risk of a call with an intermediate number according to claim 2, wherein: The acquiring of communication data of the intermediate number during the call and preprocessing the communication data to extract Mel spectrum features in the communication data further includes: Decomposing the noisy call voice data into noise coefficients of different frequency bands through wavelet transform, and performing soft threshold processing on the high-frequency noise coefficients whose noise frequency exceeds a first threshold to obtain a voice signal reconstructed by inverse wavelet transform; Extracting the Mel spectrum feature from the speech signal through a Mel filter bank; Call the BERT word segmenter to segment the customer's report text to obtain a word segmentation sequence, and input the word segmentation sequence into the BERT model for encoding to obtain the hidden state vector of the context corresponding to each word segmentation; The softmax function is called to calculate the probability distribution of different entity labels corresponding to each segmentation, and the predicted entity labels of each segmentation are merged through the decoder to obtain the structured call entity features.

4. The method for determining the risk of a call with an intermediate number according to claim 2, wherein: The extracting number attribute features and call entity features from the preprocessed communication data, and constructing a risk level model based on a random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model, includes: Extracting number attribute features and call entity features from the preprocessed communication data, and combining the number attribute features and call entity features into a feature vector according to feature dimensions; Selecting multiple feature elements from the feature vector and determining the optimal split feature based on the Gini index to construct a decision tree, wherein the random forest algorithm is composed of multiple decision trees; The decision tree is called to output a risk level prediction result based on the feature vector, and a probability distribution of the risk level is obtained according to multiple risk level prediction results output by multiple decision trees through a voting mechanism to determine the risk level of the intermediate number.

5. The method for determining the risk of a call with an intermediate number according to claim 2, wherein: The method of using the mel spectrum feature as the input of the CNN model to capture the local frequency band features in the mel spectrum feature, and calling the BiLSTM model to identify the risk probability of the call process based on the local frequency band features includes: Inputting the Mel spectrum features into a CNN model to capture local frequency band features in the Mel spectrum features through the convolution layer of the CNN model, and performing dimensionality reduction processing on the local frequency band features through a pooling layer to obtain a feature sequence; Inputting the feature sequence into the BiLSTM model to call the BiLSTM model to capture the forward and reverse time sequence features in the speech stream sequence, and outputting bidirectional fusion features through vector splicing; Based on the bidirectional fusion features, the fully connected layer is called in combination with the output layer weights and the sigmoid activation function to output the risk probability of the current call of the middle number.

6. The method for determining the risk of a call with an intermediate number according to claim 5, characterized in that: The method further includes using the mel spectrum feature as an input to the CNN model to capture the local frequency band features in the mel spectrum feature, and calling the BiLSTM model to identify the risk probability of the call process based on the local frequency band features. Acquire abnormal call voice data, and perform fusion training on the CNN model and the BiLSTM model based on the abnormal call voice data; The model parameters are adjusted by the back-propagation method, and the binary cross entropy loss function is called to optimize the model parameters to obtain the trained CNN-BiLSTM composite model.

7. The method for determining the risk of a call with an intermediate number according to claim 6, wherein: The risk level and risk probability are evaluated based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and when the comprehensive risk score exceeds a preset range, the call process of the intermediate number is determined to be an abnormal call, including: Inputting the current communication data of the intermediate number into the trained CNN-BiLSTM composite model to output the risk level and risk probability corresponding to the current communication data; Invoking a risk score fusion formula to calculate the comprehensive risk score corresponding to the current communication data based on the weight coefficients corresponding to the risk level and the risk probability, so as to determine whether the comprehensive risk score exceeds a preset range; When the comprehensive risk score exceeds a preset range, the risk level of the current call process is divided according to the preset risk identification rules, and the risk level division result is fed back to the user end.

8. A device for identifying the risk of calls with intermediate numbers, characterized in that: The device for implementing the method for determining the risk of an intermediate number call according to any one of claims 1 to 7 comprises: A data preprocessing module is used to obtain communication data of the intermediate number during the call and preprocess the communication data to extract Mel spectrum features in the communication data; A risk level output module is used to extract number attribute features and call entity features from the preprocessed communication data, and to build a risk level model based on a random forest algorithm to call the risk level corresponding to the intermediate number output by the risk level model; a risk probability output module, configured to use the mel spectrum features as input to a CNN model to capture local frequency band features in the mel spectrum features, and to call a BiLSTM model to identify the risk probability of the call process based on the local frequency band features; a comprehensive identification module, configured to evaluate the risk level and risk probability based on a preset risk identification rule to obtain a comprehensive risk score corresponding to the communication data, and to determine that the call process of the intermediate number is an abnormal call when the comprehensive risk score exceeds a preset range; The communication data includes call number information and call voice data, and the preprocessing includes standardization of call number information, extraction of call entity features, conversion of call voice data, and extraction of Mel spectrum features.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for determining the risk of an intermediate number call according to any one of claims 1 to 7 are implemented.

10. A computer storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for determining the risk of an intermediate number call according to any one of claims 1 to 7 are implemented.

Images