Vehicle insurance fraud behavior monitoring method based on improved TCNexp algorithm

Abstract

The invention provides a vehicle insurance fraud behavior monitoring method based on an improved TCNexp algorithm, and belongs to the field of feature engineering. The vehicle insurance fraud behavior monitoring method has the advantages that the features of user behaviors are specifically screened from the aspect of data selection, and under the condition that the calculation burden is not changed, through an instructive feature extraction fusion strategy, the vehicle insurance fraud behavior is monitored; in combination with the TCNexp model, the weight of the model parameter is dynamically adjusted according to the sequence length, so that TCNexp is more practical when processing the current scene. Different coping strategies are made according to behavior characteristics of users under different base stations through base station risk coefficients in combination with space-time characteristics of base station data so as to pay more attention to data points related to base station signals. The weight coefficient and bias of the crowd recognition layer are optimized by using an optimization algorithm, so that the problem that the neural network is easy to fall into local optimum can be solved, and the optimization effect is improved. The three key modules are integrated, so that operator data analysis can be well understood, and valuable information can be obtained from the operator data analysis.

Description

Technical Field

[0001] The present invention relates to the fields of operators and feature engineering, and in particular to a method for monitoring auto insurance fraud using an improved TCN_exp algorithm. Background Art

[0002] Insurance fraud has always been a key component of insurance claims risk management, and auto insurance fraud is a high-incidence area, accounting for nearly 70% of all insurance fraud cases. Small claims account for over 70% of auto insurance claims, and the average payment cycle still takes around 11 days. For cost control and customer service reasons, insurance companies aim to quickly resolve and close these numerous small claims, reducing manpower investment. However, the high incidence of fraud in the auto insurance sector, coupled with a weak legal consciousness among some customers, who even believe that insurance companies are making money from them, leads to a high incidence of fraud cases even among legitimate claims. Balancing the time demands for expedited claims processing with insurance companies' anti-fraud risk management requirements requires new technological breakthroughs and innovations to provide effective solutions for insurance institutions.

[0003] In the existing technology, users who may potentially commit fraud are usually identified based on human experience or simple feature comparison, and these users are monitored or reminded in response. Summary of the Invention

[0004] To address the above technical issues, this paper provides a method for monitoring auto insurance fraud using an improved TCN_exp algorithm. Using user base station time series data, a high-risk auto insurance group identification model is established to identify user behavior and output risk level labels (low, medium, high). Identifying high-risk groups can help the insurance industry more accurately assess risk and take appropriate measures based on customer risk levels, preventing unfair pricing practices.

[0005] The technical solution of the present invention is:

[0006] A method for monitoring auto insurance fraud based on an improved TCN_exp algorithm.

[0007] First, the abnormal points in the pre-processed user movement trajectory behavior information are marked, and the corresponding data are intercepted according to the analysis of different types of user movement behavior characteristics to establish a behavior trajectory behavior dataset;

[0008] Then, a crowd identification and classification neural network model is established. At the same time, based on the relevant characteristics of the base stations, the base station risk coefficient is calculated and associated with each base station location, which is convenient for subsequent model parameter adjustment;

[0009] Through the overall model's special analysis of vehicle owner data, combined with the impact of mobile base station data on the vehicle appearance coefficient in the time and space dimensions, high-risk group labels are output for vehicle risk control.

[0010] Further,

[0011] The model is divided into two layers. The upper layer is the TCN_exp model, which performs data extraction on time series data, analyzes different behavioral characteristics, and optimizes the weight by combining the base station risk coefficient with exp to pay more attention to the learning of samples near the base station in time and space;

[0012] The lower layer uses the output of the upper layer as the output of the fully connected neural network model to perform crowd classification and recognition. The weight coefficient and bias of the fully connected neural network are optimized through the CS algorithm. Finally, the recognition rate of the overall model is brought to 98%, and several comparison models are set.

[0013] This paper mainly uses mobile base station data, combined with the multi-feature caliber time series of CDR / DPI data, uses the improved TCN_exp to extract relevant features, and adopts the improved fully connected neural network to classify and identify the input data. Combined with user location, app usage, communication and other dimensional data, the risk level of the auto insurance base station is associated. By learning the relationship between the input data and the corresponding label, the network can classify the new data and output the risk level label.

[0014] At the same time, TCN_exp is optimized by combining exp with the spatiotemporal distribution characteristics of base stations. Incorporating a specific base station risk factor, which is highly correlated with base station auto insurance, the accuracy of the overall model is improved through gradual iterative optimization of the loss function. Temporal correlation is captured from multiple perspectives, capturing the sequential and temporal dependence of user behavior as features for the next layer. This next layer utilizes an optimization algorithm combined with the specific base station risk factor to optimize the fully connected neural network for the task of identifying high-risk groups for auto insurance. This optimization algorithm accelerates the convergence of the neural network. By adaptively adjusting the learning rate and parameter update strategy, and combining data from base stations at different locations for comparison, the network can converge more quickly to a global or local optimal solution.

[0015] Further,

[0016] The base station risk coefficient calculation formula is as follows:

[0017] w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ

[0018] Among them, w_d represents the base station risk coefficient, α represents the distance factor, the greater the base station radiation distance, the lower the value, β represents the district and county factor, the more remote the district and county, the lower the value, γ represents the density factor, the lower the number of daily base station visitors, the lower the value, and other factors; δ represents the special scenario factor, which is a fixed character, and q represents the corresponding weight of each factor.

[0019] When TCN_exp was built, the algorithm optimization adopted: dilated causal convolution, void convolution / convolutional filter, residual connection / 1x1 conv. The structure consists of multiple repeated modules, each module contains two causal void convolution layers + a residual connection. The dilation factor and number of filters of each module are different; a softmax layer is added after the last module to output the result.

[0020] After building the TCN_exp model parameters, the steps for parameter adjustment are as follows:

[0021] 1) Based on the base station risk coefficients with different scores, the weight of the i-th action is calculated to obtain the degree to which each action is affected by different base station risk coefficients;

[0022] The parameter w_d is introduced into the weight calculation part; the base station risk coefficients of different scores are normalized so that the normalized value of the base station risk coefficient of each score is the weight of the i-th action, that is, Wi, and the specific formula is as follows:

[0023]

[0024] Pi represents the probability of the i-th action, w_d represents the base station risk coefficient, Wi represents the weight of the i-th action, ∑(Wj) represents the sum of all action weights, K represents the total length of the base station risk coefficient selection, and the base station risk coefficient is w_d = (α*q1+β*q2+γ*q3…+ω*qn)+δ;

[0025] w_d is the base station risk coefficient and the calculation method is: w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ

[0026] 2) Based on the calculation results in Step 1, adjust the weight of each action;

[0027] Through the above method, the calculated probability of each action is affected by different base station risk coefficients, and the parameter adjustment commands and rewards corresponding to different actions are set. After executing the action and obtaining the corresponding reward, the weight of each action can be adjusted according to different rewards and probabilities Pi. The formula is as follows:

[0028] Wi'=Wi*exp((reward*(1-w_d)) / (K*Pi))

[0029] Wi: represents the current weight value, reward: represents the reward value obtained after executing the action in the current state; after each iteration, the base station risk coefficient is updated according to the base station position to control the relationship between different base stations, and then the selected reward is updated. In this way, the action selection strategy is gradually adjusted according to different rewards and probability updates to gradually find a better action;

[0030] 3) Further adjust the frequency of the action adjustment under the model

[0031] Use the iterative update method to limit the adjustment of the action frequency. The detailed process is in the step of updating the weight in the parameter dynamic adjustment algorithm. For the selected action, the Euclidean distance between the recognition degree and the action is calculated, and the update rate is adjusted according to the distance. That is, the farther away from the identification parameter, the faster the update rate, and the closer to the identification parameter, the slower the update rate. This method is used to further control the behavior of the result, and then the selected action weight is updated according to the adjusted update rate. The specific formula is as follows:

[0032]

[0033] Where γ is the update iteration coefficient, which is set to 0.23 here, v min and v max is the upper and lower limits of the rate update, k represents the number of iterations, w_d a The median of the base station risk coefficient is used here instead. The base station risk coefficient is obtained by calculating the update rate at the kth iteration, combined with the upper and lower limits of the maximum update rate and the influence of the previous update rate.

[0034] Crowd recognition via fully connected neural networks

[0035] The output item extracted from the upper layer data is used as the input X of the fully connected neural network through the classification model. After classification through the network hidden layer and output layer, the user's risk score is divided into three groups: low, medium, and high, and the prediction result is obtained. Combining the loss function softmax with the cross entropy, the scores of each category are first transformed by the e-exponential function, and then normalized, and finally the population classification of each category is output.

[0036] Optimize model parameters by combining base station risk coefficient w_d

[0037] The specific process is as follows:

[0038] 1) Combined with the FCNN structural parameters, randomly generate α sets X, each set represents a different combination of weights and biases;

[0039] 2) For each set, calculate the fitness A based on the network output error j , calculation formula:

[0040]

[0041] in, Is a constant, in order to limit the maximum value of fitness, Y min is the result when the metric value is minimum, and ε is a very small number of machines;

[0042] 3) For the globally optimal set, new individuals are generated through random mutation, thereby increasing the diversity of parameter results and avoiding premature convergence to the local optimal solution. The mutation formula is as follows:

[0043]

[0044] Among them, x jk is the position of the jth set in the kth dimension, is the lower bound of the jth set in the kth dimension, is the upper limit, and the value exceeding the parameter boundary is adjusted to the set range by random mapping. sl ,X xl are the upper and lower bounds of the optimization solution, respectively, and U(0,1) is a uniformly randomly distributed random number in the interval [0, 1], that is:

[0045] x kl =X xl +U(0,1)*(X sl -X xl )

[0046] 4) Calculate the Euclidean distance between each set x and the optimal solution, combine the median selection probability of the base station risk coefficient w_d, and select the set N in the next iteration process from the total set C of all parameters according to this probability. Dis(x j ) is the set x j The sum of distances to other sets, x k is another set in C, and the probability formula is as follows:

[0047]

[0048] Then repeat steps 2)-4) until the fitness reaches the preset accuracy requirement. The set x obtained at this time is the optimal parameter configuration of the fully connected neural network. Calling these configurations can realize crowd recognition at this layer.

[0049] After the model is built, the training set is input for model training and the test set is input for verification. In order to verify the optimization ability of this model, the same data is used to train and verify TCN_exp, and the recognition rate and loss function are used as parameters for model evaluation.

[0050] The beneficial effects of the present invention are

[0051] The present invention integrates the characteristic results of base station data into the overall model to achieve richer feature construction without complicated manual intervention. First, feature engineering is performed by combining base station dimensional data with other operator data, and then processed through the data extraction layer using a time series algorithm to obtain user base station data with obvious spatiotemporal characteristics. At the same time, the optimization algorithm has obvious advantages in optimization effect compared to a simple fully connected neural network. Combined with the unique spatial characteristics of base station data, the optimization effect of the neural network is improved through exploration and utilization strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the working structure of the present invention;

[0053] Figure 2 It is a schematic diagram of the TCN structure;

[0054] Figure 3 It is a schematic diagram of the fully connected neural network training process;

[0055] Figure 4 It is the curve of TCN_exp recognition rate and loss function change. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] Regarding the monitoring and reminder methods of the existing technology, we believe that the existing technology cannot achieve AI recognition, let alone accurate large-scale recognition capabilities. Therefore, in response to the problems of the existing technology, the present invention proposes a monitoring method for auto insurance fraud based on base station data and an improved TCN_exp algorithm. It mainly uses mobile base station data, combined with the multi-feature time series of CDR / DPI data, uses the improved TCN_exp to extract relevant features, and uses an improved fully connected neural network to classify and identify the input data. Combined with user location, app usage, communication and other dimensional data, the risk level of auto insurance base stations is associated. The network learns the relationship between input data and corresponding labels, so that it can classify new data and output risk level labels.

[0058] At the same time, TCN_exp is optimized by combining exp with the spatiotemporal distribution characteristics of base stations. Incorporating a specific base station risk factor, which is highly correlated with base station auto insurance, the accuracy of the overall model is improved through gradual iterative optimization of the loss function. Temporal correlation is captured from multiple perspectives, capturing the sequential and temporal dependence of user behavior as features for the next layer. This next layer utilizes an optimization algorithm combined with the specific base station risk factor to optimize the fully connected neural network for the task of identifying high-risk groups for auto insurance. This optimization algorithm accelerates the convergence of the neural network. By adaptively adjusting the learning rate and parameter update strategy, and combining data from base stations at different locations for comparison, the network can converge more quickly to a global or local optimal solution.

[0059] The present invention integrates the characteristic results of base station data into the overall model to achieve richer feature construction without complicated manual intervention. First, feature engineering is performed by combining base station dimensional data with other operator data, and then processed through the data extraction layer using a time series algorithm to obtain user base station data with obvious spatiotemporal characteristics. At the same time, the optimization algorithm has obvious advantages in optimization effect compared to a simple fully connected neural network. Combined with the unique spatial characteristics of base station data, the optimization effect of the neural network is improved through exploration and utilization strategies.

[0060] Based on the problems existing in the above methods, the present invention uses user base station time series data to establish a high-risk group identification model for auto insurance to identify user behavior and output risk level labels (low, medium, high). Identifying high-risk groups can help the insurance industry assess risks more accurately and take corresponding measures to more accurately assess risks according to customer risk levels and avoid unfair pricing and other behaviors.

[0061] The model is divided into two layers. The upper layer is the improved TCN_exp algorithm after base station risk coefficient combined with exp processing. According to the advantage of this algorithm that is easier to process time series, and combined with the unique spatial characteristics of base station data, it can more effectively capture user characteristics. The characteristics are then used as input items of the fully connected neural network and enter the lower layer of the model. The base station risk coefficient is combined with the optimization algorithm to optimize the fully connected neural network and enhance the overall advantage of the model. The architecture is as follows Figure 1 shown.

[0062] Risk control high-risk population characteristics analysis, associated base station risk coefficient calculation

[0063] Carrier data is rich in dimensions and diverse in types. To ensure uniform distribution characteristics and facilitate subsequent model optimization, we preprocess the original dataset using methods such as normalization and regularization to ensure that different data have similar distribution characteristics and the same dimensional range.

[0064] Crowd characteristics analysis:

[0065] Unlike other methods of analyzing and identifying populations, this application combines operator data and base station data to identify and analyze high-risk groups for auto insurance risk control. It focuses on analyzing user base station data, deriving several common characteristics of high-risk groups in auto insurance risk control from network statistics, and then, based on rules, obtains multiple dimensional data related to these characteristics from user base station data and operator data, including identity characteristics, communication characteristics, consumption capacity, navigation / travel app usage, base station location, etc. The characteristics are briefly listed in the following table:

[0066] Scope of work Traffic conditions Working hours Income level … Risk Score districts and counties Average speed Usually during the day Stablize … Low urban area Sometimes fast, sometimes slow Usually at night Unstable … middle Between provinces and cities Keep on the highway Not fixed Unstable … high

[0067] Table 1: Characteristics of high-risk groups in auto insurance risk control Based on the above characteristics, operator-related indicators are extracted and combined with auto insurance risk control scenarios to derive certain indicators:

[0068]

[0069]

[0070] Table 2

[0071] Base station risk factor calculation:

[0072] The evaluation is conducted through dimensions such as base station location, navigation / travel app usage, etc., and the base station risk coefficient associated with each base station is set as an important coefficient for subsequent model parameter adjustment. The unique spatial attributes of mobile base stations, including longitude and latitude, the county to which they belong, and other special influencing factors are utilized. Each indicator has a certain weight in the base station risk coefficient, and the final parameter is composed of the weighted sum of the scores of each indicator.

[0073] The base station risk coefficient calculation formula is as follows:

[0074] w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ

[0075] Among them, w_d represents the base station risk coefficient, α represents the distance factor, the greater the base station radiation distance, the lower the value, β represents the district and county factor, the more remote the district and county, the lower the value, γ represents the density factor, the lower the number of daily base station visitors, the lower the value, and other factors; δ represents the special scenario factor (such as highway), which is a fixed character, q represents the corresponding weight of each factor, combined with the auto insurance risk control scenario, where q1 is 0.18, q2 is 0.25, q3 is 0.27, and δ is 1.

[0076] Based on the analysis of the characteristics of user behavior data, the time points that meet the characteristics are found, and 30 data are intercepted separately using a sliding window. Each data contains a matrix of 18 features such as speed, passed base stations, longitude and latitude, etc. as a driving behavior data set. However, the accuracy of the training model with a small amount of data will be relatively low. Therefore, 3,000 data segments are intercepted in the present invention, of which 700 segments belong to the behavioral characteristics of low-risk groups belonging to deliverymen, 700 segments belong to medium- and high-risk groups, 700 segments belong to high-risk groups, and 900 segments have other characteristics. 80% of the data segments are selected as training sets, and the remaining 20% are used as test sets, and it is ensured that the proportion of each group behavior in the training set and the test set is the same.

[0077] Use the optimized TCN_exp model for data feature extraction

[0078] To better preserve temporal continuity, TCN_exp serves as a data extraction layer, linking historical data with current data. Furthermore, TCN_exp leverages TCN's parallel computing capabilities and its ability to capture long-range dependencies, effectively processing long sequences of data. TCN_exp's forgetting layer selectively forgets historical data that is not crucial to the results, further improving the model's efficiency and performance. To prevent abnormal data from interfering with the model, we adjust model parameter updates by assigning different weights to data at different times, optimizing model performance at critical moments. This combination of TCN_exp and weight adjustment better adapts to the characteristics of time series data, enhancing the model's robustness and predictive power.

[0079] 1)TCN_exp construction

[0080] In the present invention, TCN_exp is used to utilize the time series information in the original data to learn the long-term dependencies in the sequence, so as to output more refined features for future identification of people involved in risk control of auto insurance. The TCN model usually contains multiple convolutional layers and residual connections to extract the features of the data. Each convolutional layer contains multiple convolution kernels, and different convolution kernels can capture different features. Residual connections can help alleviate the gradient vanishing problem and improve the training effect of the model. The TCN model consists of multiple convolutional layers and nonlinear activation functions, which can effectively capture the features and patterns in time series data.

[0081] TCN(x)=σ(Conv1(x)οConv2(x)ο...οConv n (x))

[0082] Where Conv1(x) represents the output of the i-th convolutional layer, ο represents the convolution operation, and σ represents the activation function.

[0083] TCN structure Figure 2 As shown on the right, this can be considered a specialized convolution operation. It involves sliding a convolution kernel over a large data input and performing a weighted summation of the data within each kernel to produce a new data output. This extracts local features from the input while reducing the number of parameters.

[0084] The algorithm optimization adopts: dilated causal convolution, void convolution / convolutional filter (by inserting blank positions / dilation factors in the convolution kernel, the convolution kernel covers a longer range of input without increasing the number of parameters, thereby increasing the receptive field), residual connection / 1x1 conv (which can build deeper networks and improve performance),

[0085] The structure consists of multiple repeated modules, each module contains two causal hole convolution layers + a residual connection, the expansion factor and the number of filters of each module are different; a softmax layer is added after the last module to output the result.

[0086] 2) Construct a dynamic parameter adjustment method for the constructed TCN_exp model parameters:

[0087] The advantage of the dynamic parameter adjustment algorithm designed in this invention is that it can effectively balance exploration and utilization in an environment with random changes, while also having good performance and convergence properties. Based on its properties, it can better actively associate the fluctuations of the base station risk coefficient with the relevant parameters of TCN_exp. The main parameter adjustment steps are as follows:

[0088] Step 1: Based on the base station risk coefficients with different scores, calculate the weight of the i-th action and obtain the degree to which each action is affected by different base station risk coefficients;

[0089] The parameter w_d is introduced into the weight calculation part. The base station risk coefficients of different scores are normalized so that the normalized value of the base station risk coefficient of each score is the weight of the i-th action, that is, Wi. The specific formula is as follows:

[0090]

[0091] In this invention, Pi represents the probability of the i-th action, w_d represents the base station risk coefficient, which is defined as 0.43 in this invention, Wi represents the weight of the i-th action (derived from the base station risk coefficient), ∑(Wj) represents the sum of all action weights, and K represents the total length of the base station risk coefficient selection. The base station risk coefficient is w_d = (α*q1+β*q2+γ*q3…+ω*qn)+δ. Here, when the corresponding base station risk coefficient is higher, the probability of the corresponding action is higher, and the parameter update step size is shorter, which means the update rate will be faster. A smaller step size means a smaller update amplitude, which helps the algorithm converge stably.

[0092] w_d is the base station risk coefficient and the calculation method is: w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ

[0093] Step 2: Based on the calculation results in Step 1, adjust the weight of each action;

[0094] Through the above method, the calculated probability of each action is affected by different base station risk coefficients, and the parameter adjustment commands and rewards corresponding to different actions are set. After executing the action and obtaining the corresponding reward, the weight of each action can be adjusted according to different rewards and probabilities Pi. The formula is as follows:

[0095] Wi'=Wi*exp((reward*(1-w_d)) / (K*Pi))

[0096] Wi: represents the current weight value, and reward: represents the reward value obtained after executing the action in the current state. After each iteration, the base station risk coefficient is updated based on the base station location to control the relationship between different base stations. The selected reward is then updated. This gradually adjusts the action selection strategy based on the different rewards and probability updates to gradually find the optimal action. Based on these actions, the parameters of TCN_exp, such as n_estimators, learning_rate, max_bin, num_leaves, and bagging_fraction, are adjusted. Step 3: Further adjust the frequency of action adjustments under the model.

[0097] The present invention uses an iterative update method to limit the adjustment of the action frequency. The detailed process is that when updating the weight in the parameter dynamic adjustment algorithm, for the selected action, the Euclidean distance between the recognition degree and the action is calculated, and the update rate is adjusted according to the distance. That is, the farther away from the identification parameter, the faster the update rate, and the closer to the identification parameter, the slower the update rate. This method can be used to further control the behavior of the result, and then the selected action weight is updated according to the adjusted update rate. The specific formula is as follows:

[0098]

[0099] Where γ is the update iteration coefficient, which is set to 0.23 here, v min and v max The upper and lower limits of the rate update are set here to (0,1), k represents the number of iterations, w_d a The median of the base station risk coefficient is used here instead. The base station risk coefficient is obtained by calculating the update rate at the kth iteration, combined with the upper and lower limits of the maximum update rate and the influence of the previous update rate.

[0100] Through this method, the dynamic parameter adjustment algorithm can adopt different update strategies for different actions during the iteration process. Actions farther from the identification parameters will have their weights updated more frequently to more quickly approach the actual probability distribution; actions closer to the identification parameters will be updated more slowly to avoid over-adjustment. This way, the algorithm can more accurately estimate the true probability distribution of each action while maintaining a certain degree of exploration, making the optimal choice.

[0101] 3) Obtain the extracted spatiotemporal data features.

[0102] Finally, the following time series features are generated: the risk score y^j obtained by user p^j at date^j takes the value of (1: low, 2: medium, 3: high), as well as the highly correlated dimensional features. jz_n represents some features of the base station, such as the number of base stations passed, the average speed of base station passing, etc. app_n represents some features of app behavior, such as the number of times the app is used and the duration, as shown in Table 3.

[0103]

[0104]

[0105] Table 3

[0106] Crowd recognition is performed based on the spatiotemporal data features extracted by the improved TCN_exp and combined with other dimensional data.

[0107] By combining the data extracted by the improved TCN_exp with the user's data on communication, app usage, consumption and other dimensions, this layer uses the nonlinear fitting capability of the fully connected neural network to classify and identify high-risk groups for risk control.

[0108] 1) Crowd recognition through fully connected neural networks

[0109] Here, the output item extracted from the upper layer data is used as the input X of the fully connected neural network through the classification model. After classification through the network hidden layer and output layer, the user's risk score is divided into three groups: low, medium, and high, and the prediction result is obtained. Combining the loss function softmax with cross entropy, the scores of each category are first transformed by the e-exponential function, and then normalized (each category score is divided by the total score of all categories), and finally the population classification of each category is output. The optimization of model parameters focuses on weights and learning rates, etc., which are achieved through multiple iterations of base station risk coefficients combined with optimization methods, such as Learning Rate, Regularization, Batch Normalization, Dropout ratio, etc. The specific process is as follows: Figure 3 Shown:

[0110] FCNN, as the main body of the classification model, is a multi-layer perceptron that seeks the most reasonable and robust hyperplane between categories. During model training, batch normalization can be used to achieve preliminary optimization of the learning rate. At the same time, in order to avoid overfitting of the model, the information allowed to be stored in the model is adjusted to constrain it. At the same time, Dropout is used on a certain layer, which means randomly discarding some outputs of the layer (setting the output value to 0). These discarded neurons are like being deleted by the network. By adjusting the random inactivation ratio (Dropout ratio), the neurons in the hidden layer are prevented from being activated with a certain probability.

[0111] 2) Optimize model parameters based on base station risk coefficient w_d

[0112] This invention uses an improved adaptive optimization algorithm at this level, combined with the base station's risk factor, to optimize the fully connected model. The algorithm uses the population position and fitness information generated during the optimization iteration to estimate the shape of the objective function's search space. Adaptive adjustments are then made based on the search results. The risk factor is used to control the probability of selecting new parameter sets during optimization iterations, enabling more efficient exploration of the search space and a better balance between global and local search during the search process. The specific process is as follows:

[0113] Step 1: Combined with the FCNN structural parameters, randomly generate α sets X, each set represents a different combination of weights and biases.

[0114] Step 2: For each set, calculate the fitness A based on the network output error j This step determines the degree of influence of the set on other sets during the optimization process. The greater the influence, the more obvious the impact on the parameters. The calculation formula is:

[0115]

[0116] in, Is a constant, in order to limit the maximum value of fitness, Y min is the result when the metric value is minimum, and ε is a very small machine number.

[0117] Step 3: For the global optimal set, new individuals are generated through random mutation to increase the diversity of parameter results and avoid premature convergence to the local optimal solution. The mutation formula is as follows:

[0118]

[0119] Among them, x jk is the position of the jth set in the kth dimension, is the lower bound of the jth set in the kth dimension, is the upper limit, and the value exceeding the parameter boundary is adjusted to a reasonable range by random mapping. sl ,X xl are the upper and lower bounds of the optimization solution, respectively, and U(0,1) is a uniformly randomly distributed random number in the interval [0, 1]. That is:

[0120] x kl =X xl +U(0,1)*(X sl -X xl )

[0121] Step 4: Calculate the Euclidean distance between each set x and the optimal solution, combine the median selection probability of the base station risk coefficient w_d, and select the set N in the next iteration process from the total set C of all parameters according to this probability. Dis(x j ) is the set x j The sum of distances to other sets, x k is another set in C, and the probability formula is as follows:

[0122]

[0123] Then repeat steps 2-4 until the fitness reaches the preset accuracy requirement. The set x obtained at this time is the optimal parameter configuration of the fully connected neural network. Calling these configurations can realize crowd recognition at this layer.

[0124] Model Evaluation

[0125] After the model is built, the training set is input for model training and the test set is input for verification. In order to verify the optimization ability of this model, the same data is used to train and verify TCN_exp (for easy distinction, hereinafter referred to as TCN_exp), and the recognition rate and loss function are used as parameters for model evaluation. Figure 3 The figure shows the recognition rate and loss function of TCN_exp in the present invention, where the blue is the training set result and the red is the test set result. It can be seen that after 500 iterations, both parameters tend to be stable and the effect is ideal, as shown in Figure 2. Figure 4 As shown, Figure 4 (a) is the traditional TCN model, Figure 4 (b) is the TCN_exp model, where the blue line segment is the result of the training set and the red line segment is the result of the test set. It can be seen that although the results of the TCN model gradually converge with the increase in the number of training times, the recognition effect is general and the recognition rate is not as high as TCN_exp. The model was experimented with 5 times and the average recognition rate was taken for comparison. The comparison results are shown in Table 4. The average recognition rate of TCN is 95.2%, and the recognition rate of TCN_exp reaches 98.4%. It can be seen that the present invention has certain advantages.

[0126]

[0127] Table 4

[0128] The above description is only a preferred embodiment of the present invention and is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A method for monitoring auto insurance fraud based on an improved TCN_exp algorithm, characterized in that: First, the abnormal points in the pre-processed user movement trajectory behavior information are marked, and the corresponding data are intercepted according to the analysis of different types of user movement behavior characteristics to establish a behavior trajectory behavior dataset; Then, a crowd identification and classification neural network model is established. At the same time, based on the relevant characteristics of the base stations, the base station risk coefficient is calculated and associated with each base station location, which is convenient for subsequent model parameter adjustment; Through the overall model's special analysis of vehicle owner data, combined with the impact of mobile base station data on the vehicle appearance coefficient in the time and space dimensions, high-risk group labels are output for vehicle risk control.

2. The method according to claim 1, characterized in that The model is divided into two layers. The upper layer is the TCN_exp model, which performs data extraction on time series data, analyzes different behavioral characteristics, and optimizes the weight by combining the base station risk coefficient with exp to pay more attention to the learning of samples near the base station in time and space; The lower layer uses the output of the upper layer as the output of the fully connected neural network model to perform crowd classification and recognition. The weight coefficient and bias of the fully connected neural network are optimized through the CS algorithm. Finally, the recognition rate of the overall model is brought to 98%, and several comparison models are set.

3. The method according to claim 2, characterized in that Using mobile base station data as the primary input, combined with multi-feature time series data from CDR / DPI, the improved TCN_exp extracts relevant features. An improved fully connected neural network is then used to classify and identify the input data. This data is then combined with user location, app usage, and communication data to correlate the risk level of vehicle insurance base stations. The network learns the relationship between input data and corresponding labels, classifying new data and outputting risk level labels. At the same time, TCN_exp is optimized by combining exp with the spatiotemporal distribution characteristics of base stations, and the special base station risk coefficient related to base station auto insurance is combined. The accuracy of the overall model is improved by gradually iteratively optimizing the loss function, and time correlation is obtained from several aspects. The sequential and time-dependent nature of user behavior is captured as the characteristics of the next layer. The next layer uses the optimization algorithm combined with the special base station risk coefficient to optimize the fully connected neural network to adapt to the task of identifying high-risk groups for auto insurance. The convergence process of the neural network is accelerated by the optimization algorithm, and the learning rate and parameter update strategy are adaptively adjusted. At the same time, the base station data in different locations are combined for comparison, so that the network can converge to the global optimal solution or the local optimal solution faster.

4. The method according to claim 3, characterized in that The base station risk coefficient calculation formula is as follows: w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ Among them, w_d represents the base station risk coefficient, α represents the distance factor, the greater the base station radiation distance, the lower the value, β represents the district and county factor, the more remote the district and county, the lower the value, γ represents the density factor, the lower the number of daily base station visitors, the lower the value, and other factors; δ represents the special scenario factor, which is a fixed character, and q represents the corresponding weight of each factor.

5. The method according to claim 4, characterized in that When TCN_exp was built, the algorithm optimization adopted: dilated causal convolution, void convolution / convolutional filter, residual connection / 1x1 conv. The structure consists of several repeated modules, each of which contains two causal void convolution layers + a residual connection. The dilation factor and number of filters of each module are different; a softmax layer is added after the last module to output the result.

6. The method according to claim 5, characterized in that After constructing the TCN_exp model parameters, the steps for parameter adjustment are as follows: 1) Based on the base station risk coefficients with different scores, the weight of the i-th action is calculated to obtain the degree to which each action is affected by different base station risk coefficients; The parameter w_d is introduced into the weight calculation part; the base station risk coefficients of different scores are normalized so that the normalized value of the base station risk coefficient of each score is the weight of the i-th action, that is, Wi, and the specific formula is as follows: Pi represents the probability of the i-th action, w_d represents the base station risk coefficient, Wi represents the weight of the i-th action, ∑(Wj) represents the sum of all action weights, K represents the total length of the base station risk coefficient selection, and the base station risk coefficient is w_d = (α*q1+β*q2+γ*q3…+ω*qn)+δ; w_d is the base station risk coefficient and the calculation method is: w_d=(α*q1+β*q2+γ*q3…+ω*qn)+δ 2) Based on the calculation results in Step 1, adjust the weight of each action; Through the above method, the calculated probability of each action is affected by different base station risk coefficients, and the parameter adjustment commands and rewards corresponding to different actions are set. After the action is executed and the corresponding reward is obtained, the weight of each action can be adjusted according to different rewards and probabilities Pi. The formula is as follows: Wi'=Wi*exp((reward*(1-w_d)) / (K*Pi)) Wi: represents the current weight value, reward: represents the reward value obtained after performing the action in the current state; After each iteration, the base station risk coefficient is updated according to the base station position to control the relationship between different base stations, and then the selected reward is updated. In this way, the action selection strategy is gradually adjusted according to different rewards and probability updates to gradually find a better action; 3) Further adjust the frequency of the action adjustment under the model Use iterative update method to limit the adjustment of action frequency. The detailed process is in the step of updating weights in the parameter dynamic adjustment algorithm. For the selected action, calculate the Euclidean distance between the recognition degree and the action, and adjust the update rate according to the distance. That is, the farther away from the recognition parameter, the faster the update rate, and the closer to the recognition parameter, the slower the update rate. This method is used to further control the behavior of the result. Then, the selected action weight is updated according to the adjusted update rate. The specific formula is as follows: Where γ is the update iteration coefficient, which is set to 0.23 here, v min and v max is the upper and lower limits of the rate update, k represents the number of iterations, w_d a The median of the base station risk coefficient is used here instead. The base station risk coefficient is obtained by calculating the update rate at the kth iteration, combined with the upper and lower limits of the maximum update rate and the influence of the previous update rate.

7. The method according to claim 6, characterized in that Crowd recognition via fully connected neural networks The output item extracted from the upper layer data is used as the input X of the fully connected neural network through the classification model. After classification through the network hidden layer and output layer, the user's risk score is divided into three groups: low, medium, and high, and the prediction result is obtained. Combining the loss function softmax with the cross entropy, the scores of each category are first transformed by the e-exponential function, and then normalized, and finally the population classification of each category is output.

8. The method according to claim 7, characterized in that Optimize model parameters by combining base station risk coefficient w_d The specific process is as follows: 1) Combined with the FCNN structural parameters, randomly generate α sets X, each set represents a different combination of weights and biases; 2) For each set, calculate the fitness A based on the network output error j , calculation formula: in, Is a constant, in order to limit the maximum value of fitness, Y min is the result when the metric value is minimum, and ε is a very small number of machines; 3) For the globally optimal set, new individuals are generated through random mutation, thereby increasing the diversity of parameter results and avoiding premature convergence to the local optimal solution. The mutation formula is as follows: Among them, x jk is the position of the jth set in the kth dimension, is the lower bound of the jth set in the kth dimension, is the upper limit, and the value exceeding the parameter boundary is adjusted to the set range by random mapping. sl ,X xl are the upper and lower bounds of the optimization solution, respectively, and U(0,1) is a uniformly randomly distributed random number in the interval [0, 1], that is: x kl =X xl +U(0,1)*(X sl -X xl ) 4) Calculate the Euclidean distance between each set x and the optimal solution, combine the median selection probability of the base station risk coefficient w_d, and select the set N in the next iteration process from the total set C of all parameters according to this probability. Dis(x j ) is the set x j The sum of distances to other sets, x k is another set in C, and the probability formula is as follows: Then repeat steps 2)-4) until the fitness reaches the preset accuracy requirement. The set x obtained at this time is the optimal parameter configuration of the fully connected neural network. Calling these configurations can realize crowd recognition at this layer.

9. The method according to claim 8, characterized in that After the model is built, the training set is input for model training and the test set is input for verification. In order to verify the optimization ability of this model, the same data is used to train and verify TCN_exp, and the recognition rate and loss function are used as parameters for model evaluation.

Images