Vehicle dynamic pricing system and platform based on reinforcement learning and risk rule fusion

Abstract

This invention discloses a vehicle mortgage dynamic pricing system and platform based on the fusion of reinforcement learning and risk rules. The system includes: a data acquisition module that acquires static attributes and dynamic time-series features electronic data and loads them into a first data buffer; a residual value decay prediction module that generates a tensor data of the predicted disposal residual value probability distribution at the preset mortgage maturity time based on a time-series feature encoder; a risk constraint construction module that extracts the stress residual value and writes it into a second data buffer; and a reinforcement learning pricing module that constructs an asymmetric action exploration constraint function based on the ratio of the stress residual value to the vehicle purchase price field, respectively limiting the lower limit of the interest rate and the upper limit of the loan-to-value ratio, and outputting pricing data. This invention solves the technical problems of data read / write conflicts and asynchronous risk constraints in mortgage pricing through a dual physical address space isolation architecture and an asymmetric constraint mechanism.

Description

Technical Field

[0001] This invention relates to the field of computer financial risk control technology, and in particular to a vehicle dynamic pricing system and platform based on the fusion of reinforcement learning and risk rules. Background Technology

[0002] As the auto finance market continues to expand, vehicle mortgage loans have become an important segment of financial institutions' lending business. In vehicle mortgage lending, lending institutions need to assess the vehicle's value upon accepting the mortgage application and determine the mortgage interest rate and loan-to-value ratio accordingly. Traditional pricing methods rely on manual experience rules or static valuation models, which are ill-suited to the dynamic characteristic of vehicle residual value rapidly declining over time. In recent years, reinforcement learning technology has been increasingly applied to the field of dynamic pricing of financial products. By constructing state and action spaces, the pricing model autonomously learns the optimal pricing strategy through interaction with the environment. Simultaneously, time-series prediction models based on Long Short-Term Memory (LSTM) networks have also been initially applied in used car residual value assessment, capable of predicting future value trends using time-series data such as vehicle history maintenance records and mileage.

[0003] However, existing technical solutions suffer from fundamental technical flaws: the exploration process of the reinforcement learning pricing module is separated from the risk constraint stage, causing the model to frequently cross risk control red lines in pursuit of profit maximization, resulting in invalid pricing results that do not meet risk requirements; the output of the residual value prediction module is usually a single fixed value, which cannot quantify the uncertainty of the prediction results, resulting in a lack of quantitative basis for risk exposure assessment; in addition, the data processing modules share the same cache area in memory, and residual value prediction data and pricing calculation data frequently experience read-write conflicts, resulting in excessive power consumption of the processor when performing a large number of invalid pricing calculations, and the pricing decision and risk constraint cannot be synchronized in time. These flaws collectively lead to low computational efficiency, insufficient risk control accuracy, and serious waste of processor resources in existing systems when processing vehicle mortgage pricing tasks.

[0004] Therefore, this invention proposes a vehicle dynamic pricing system and platform based on the fusion of reinforcement learning and risk rules. Summary of the Invention

[0005] This invention provides a vehicle dynamic pricing system and platform based on the fusion of reinforcement learning and risk rules. It avoids data read / write conflicts between residual value prediction and pricing calculation through a dual physical address space isolation architecture to reduce processor power consumption. Furthermore, it completely eliminates invalid pricing outputs and model convergence oscillations caused by the separation of reinforcement learning pricing exploration and risk constraints.

[0006] This invention provides a vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules, running on a computer device including a processor and memory, comprising: The data acquisition module is configured to acquire static attribute electronic data and dynamic time-series characteristic electronic data of the mortgaged vehicle through the processor control data interface, and load the static attribute electronic data and dynamic time-series characteristic electronic data into the first data cache area in the memory. The first data cache area is the first physical address space in the memory that is configured with exclusive read and write permissions for the data acquisition module and the residual value decay prediction module. The residual value decay prediction module is configured to have the processor call the weight file of the first neural network model with built-in time-series feature encoder, read the dynamic time-series feature electronic data in the first data buffer, and calculate and generate tensor data representing the probability distribution of the predicted disposal residual value of the mortgaged vehicle at the time of the expiration of the preset mortgage period based on the duration of the preset mortgage period. The risk constraint construction module is configured to have the processor execute a quantile extraction algorithm on the tensor data to extract the pressure residual value that meets the preset risk confidence threshold condition, and write the pressure residual value into the second data cache in the memory. The second data cache is a second physical address space in the memory that is configured with exclusive read and write permissions for the risk constraint construction module and the reinforcement learning pricing module. The second physical address space is isolated from the first physical address space. The reinforcement learning pricing module is configured to have the processor call the weight file of the second neural network model, read the pressure residual value from the second data cache, and construct an asymmetric action exploration constraint function based on the ratio of the pressure residual value to the vehicle purchase price field in the static attribute electronic data. As the ratio decreases, the asymmetric action exploration constraint function synchronously narrows the upper limit boundary of the mortgage ratio dimension and raises the lower limit boundary of the mortgage interest rate dimension. The asymmetric action exploration constraint function limits the lower limit boundary of the mortgage interest rate dimension and the upper limit boundary of the mortgage ratio dimension in the action space of the second neural network model, respectively, and outputs the floating-point values ​​of the mortgage interest rate and the mortgage ratio. The pricing output interface is configured to encapsulate the floating-point values ​​of the mortgage rate and the mortgage percentage into electronic messages and send them to the front-end interactive device.

[0007] Furthermore, the dynamic time-series characteristic electronic data acquired by the data acquisition module includes the time-series encoding of the vehicle's historical maintenance and repair records, the monthly cumulative mileage increment numerical sequence, and the slope parameter of the used car price index curve for the same model. The residual value decay prediction module has a built-in time-series feature encoder, which is a long short-term memory network computing unit. The long short-term memory network computing unit is used to convert dynamic time-series feature electronic data into a fixed-dimensional vehicle life cycle feature vector. The tensor data generated by the residual value decay prediction module, which characterizes the probability distribution of the predicted disposal residual value, includes mean and variance parameters. The risk constraint construction module executes a quantile extraction algorithm that calculates the risk value function value corresponding to the preset tail probability threshold based on the mean parameter and variance parameter, and determines the risk value function value as the stress residual value.

[0008] Furthermore, the input layer state vector dimension of the reinforcement learning pricing module includes the encoding dimension of static attribute electronic data, the feature dimension of vehicle life cycle feature vector, and the scalar dimension of pressure residual value. The action output layer of the second neural network model includes output nodes for the mortgage rate adjustment range and the mortgage ratio adjustment range.

[0009] Furthermore, the reinforcement learning pricing module has a built-in reward signal calculation submodule; In offline training mode, the reward signal calculation submodule is used to calculate the instantaneous reward value; When the ratio of the collateral ratio floating-point value output by the second neural network model to the electronic data of the actual residual value of the mortgaged vehicle obtained by the data acquisition module at the expiration of the mortgage period is greater than the preset bad debt threshold constant, the reward signal calculation submodule outputs a negative gradient back transmission signal, triggering the update process of the weight parameters of the second neural network model to suppress the power consumption of the processor performing invalid pricing calculations. When the ratio of the collateral ratio floating-point value output by the second neural network model to the actual residual value electronic data at the end of the collateral term is less than or equal to the preset bad debt threshold constant, and the collateral interest rate floating-point value meets the preset income threshold condition, the reward signal calculation submodule outputs a positive gradient feedback signal to solidify the weight parameters of the second neural network model.

[0010] Furthermore, it also includes a risk rule fusion module, which runs in the processor; The risk rule fusion module loads a set of expert risk hard boundary rules pre-stored in the memory. The set of expert risk hard boundary rules includes the maximum allowable vehicle age threshold variable, the minimum allowable mortgage residual value multiple threshold variable, and the upper and lower limit threshold variables of the mortgage interest rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. The risk rule fusion module calls the comparator logic circuit to perform the following judgments: compare the vehicle age value of the mortgaged vehicle with the maximum allowable vehicle age threshold variable, and compare the ratio of the residual value of the pressure to the vehicle purchase price field with the minimum allowable residual value multiple threshold variable. When the vehicle age value exceeds the maximum allowable vehicle age threshold, an interest rate adjustment is triggered. When the ratio of the residual value of the pressure to the vehicle purchase price field is lower than the minimum allowable residual value multiple threshold variable, the upper limit boundary of the mortgage ratio dimension will be forcibly set to zero. After the reinforcement learning pricing module outputs the original floating-point value of the mortgage rate, the risk rule fusion module determines whether the original floating-point value of the mortgage rate conforms to the upper and lower limit threshold variables of the mortgage rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. If it does not conform, the original floating-point value of the mortgage rate is truncated and corrected to generate a corrected floating-point value of the mortgage rate that meets the hard boundary conditions.

[0011] Furthermore, the residual value decay prediction module and the reinforcement learning pricing module are configured with an asynchronous clock signal triggering mechanism; The weight file of the first neural network model in the residual decay prediction module is updated with parameters at the end of the first clock cycle. The input layer of the second neural network model in the reinforcement learning pricing module receives the new state vector dimension boundary generated by the updated weight file of the first neural network model after the first clock cycle ends, and performs policy iterative update calculation in the second clock cycle.

[0012] Furthermore, the data acquisition module is also equipped with an on-board diagnostic system signal receiving port, which is used to acquire in real time the on-board diagnostic system fault code data and the cumulative number of traffic violations of the mortgaged vehicle during the mortgage period; The risk constraint construction module obtains the first attenuation coefficient by querying a preset fault code attenuation coefficient mapping table based on the fault code data of the vehicle diagnostic system, and obtains the second attenuation coefficient by querying a preset violation number attenuation coefficient mapping table based on the cumulative number of traffic violations. The product of the first attenuation coefficient and the second attenuation coefficient is used as the dynamic attenuation coefficient. The pressure residual value stored in the second data cache is dynamically attenuated and corrected by multiplication. The corrected pressure residual value is then updated in real time to the address space of the second data cache of the reinforcement learning pricing module through atomic synchronization operation. This triggers the reinforcement learning pricing module to recalculate the floating-point values ​​of the mortgage rate and the mortgage ratio based on the corrected pressure residual value.

[0013] Furthermore, it also includes a module for generating a visual interactive interface for mortgage risk assessment; The module for generating a visual interactive interface for mortgage risk assessment is used to drive the display device to render a graphical user interface that includes a pressure residual value decay curve layer and a pricing confidence interval dashboard layer; The pressure residual value decay curve layer uses the time axis as the horizontal axis and the mean of the pressure residual value value and the probability distribution of the predicted disposal residual value as the vertical axis. It displays the safety boundary of vehicle value during the mortgage period in real time and responds to the user's drag operation of the vertex control on the time axis by sending a mortgage period change signal to the risk constraint construction module to trigger the recalculation of the pressure residual value value. The pricing confidence interval dashboard layer is used to respond to the floating-point value of the collateral rate output by the reinforcement learning pricing module. It marks the risk-return balance range of the current floating-point value of the collateral rate on the dashboard, and responds to the user's drag operation on the target interest rate pointer on the dashboard. It sends an interest rate adjustment signal to the reinforcement learning pricing module to trigger the recalculation of the floating-point value of the collateral ratio.

[0014] Furthermore, the augmented learning pricing module includes an online inference working mode and an offline training working mode; When the reinforcement learning pricing module is in online inference mode, the processor only calls the weight parameter file of the second neural network model stored in memory to perform forward propagation calculation to generate the floating-point value of the mortgage rate. It does not perform backpropagation gradient update calculation. It obtains the asymmetric action exploration constraint function parameters corresponding to the current pressure residual value by reading the mapping table data pre-calculated based on the preset risk confidence threshold. Based on the asymmetric action exploration constraint function parameters, it determines the lower boundary of the mortgage rate dimension and the upper boundary of the mortgage ratio dimension.

[0015] This invention provides a vehicle dynamic pricing platform based on the fusion of reinforcement learning and risk rules, comprising: The mortgage application processing unit is used to parse the received mortgage request data packet containing the vehicle identification number field and the preset mortgage term field; The risk assessment interface unit is used to initiate structured query language requests to external vehicle credit databases to obtain electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records. The pricing engine unit's processor is loaded with computer-executable instructions stored in memory, such as those described above, for a vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules, used to generate floating-point values ​​for the mortgage rate and the mortgage ratio. The result feedback unit is used to encapsulate the floating-point values ​​of the collateral interest rate and the collateral ratio generated by the pricing engine unit into a Hypertext Transfer Protocol response message and send it to the front-end interactive device.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a first data cache and a second data cache with mutually isolated dual physical address spaces, and assigning dedicated read and write permissions to the data acquisition module, residual value decay prediction module, risk constraint construction module, and reinforcement learning pricing module respectively, this invention fundamentally solves the technical problem of frequent read and write conflicts between residual value prediction data and pricing calculation data in memory in existing technologies, significantly reducing the power consumption overhead of the processor executing invalid pricing calculations. Simultaneously, by having the residual value decay prediction module output tensor data of the predicted disposal residual value probability distribution and the risk constraint construction module extract the pressure residual value value, this invention, for the first time, quantifies the uncertainty of vehicle residual value prediction into a numerical indicator that can be used for pricing constraints, overcoming the limitations of existing technologies. Existing technologies that only output a single fixed residual value result in a lack of quantitative basis for risk exposure assessment. To address this, the reinforcement learning pricing module constructs an asymmetric action exploration constraint function based on the ratio of the stress residual value to the vehicle purchase price field. As the ratio decreases, the upper limit boundary of the loan-to-value ratio dimension narrows while the lower limit boundary of the mortgage interest rate dimension rises. This completely solves the fundamental technical flaws of existing reinforcement learning pricing schemes, where pricing exploration and risk constraints are separated, and the model frequently crosses the risk control red line, outputting invalid pricing results. It achieves pre-integration of pricing decisions and risk constraints at the action space level, enabling the reinforcement learning model to naturally avoid invalid exploration actions involving high interest rates and high loan-to-value ratios during training, accelerating model convergence and improving processor computing resource utilization.

[0017] In summary, this invention solves the read / write conflict between residual value prediction data and pricing calculation data by using a dual-cache architecture that isolates the first and second physical address spaces, thereby reducing the power consumption of the processor in invalid pricing calculations. Furthermore, by constructing an asymmetric action exploration constraint function based on the ratio of the pressure residual value to the vehicle purchase price field, the upper limit boundary of the collateral ratio is narrowed and the lower limit boundary of the collateral interest rate is raised as the ratio decreases. This completely eliminates the invalid pricing output and model convergence oscillations caused by the separation of reinforcement learning pricing exploration and risk constraints, achieving a pre-integration of pricing decisions and risk constraints at the action space level.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a diagram illustrating the architecture of a vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules in an embodiment of the present invention. Figure 2 This is a simplified data flow diagram of the technical solution of the vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] refer to Figure 1 and Figure 2 This invention provides a vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules, running on a computer device including a processor and memory, comprising: The data acquisition module is configured to acquire static attribute electronic data and dynamic time-series characteristic electronic data of the mortgaged vehicle through the processor control data interface, and load the static attribute electronic data and dynamic time-series characteristic electronic data into the first data cache area in the memory. The first data cache area is the first physical address space in the memory that is configured with exclusive read and write permissions for the data acquisition module and the residual value decay prediction module. The residual value decay prediction module is configured to have the processor call the weight file of the first neural network model with built-in time-series feature encoder, read the dynamic time-series feature electronic data in the first data buffer, and calculate and generate tensor data representing the probability distribution of the predicted disposal residual value of the mortgaged vehicle at the time of the expiration of the preset mortgage period based on the duration of the preset mortgage period. The risk constraint construction module is configured to have the processor execute a quantile extraction algorithm on the tensor data to extract the pressure residual value that meets the preset risk confidence threshold condition, and write the pressure residual value into the second data cache in the memory. The second data cache is a second physical address space in the memory that is configured with exclusive read and write permissions for the risk constraint construction module and the reinforcement learning pricing module. The second physical address space is isolated from the first physical address space. The reinforcement learning pricing module is configured to have the processor call the weight file of the second neural network model, read the pressure residual value from the second data cache, and construct an asymmetric action exploration constraint function based on the ratio of the pressure residual value to the vehicle purchase price field in the static attribute electronic data. As the ratio decreases, the asymmetric action exploration constraint function synchronously narrows the upper limit boundary of the mortgage ratio dimension and raises the lower limit boundary of the mortgage interest rate dimension. The asymmetric action exploration constraint function limits the lower limit boundary of the mortgage interest rate dimension and the upper limit boundary of the mortgage ratio dimension in the action space of the second neural network model, respectively, and outputs the floating-point values ​​of the mortgage interest rate and the mortgage ratio. The pricing output interface is configured to encapsulate the floating-point values ​​of the mortgage rate and the mortgage percentage into electronic messages and send them to the front-end interactive device.

[0023] In this embodiment, the static attribute electronic data and dynamic time-series characteristic electronic data of the mortgaged vehicle are obtained through the processor control data interface. Specifically, the data acquisition module initiates a structured query language request to an external vehicle credit database through the processor control data interface to obtain the static attribute electronic data and dynamic time-series characteristic electronic data.

[0024] In this embodiment, the static attribute electronic data of the mortgaged vehicle includes vehicle identification number, brand and model, initial registration date, vehicle purchase price field, engine displacement, transmission type, and body color; the dynamic time-series characteristic electronic data includes time-series encoding of the vehicle's historical maintenance records, monthly cumulative mileage increment numerical sequence, and slope parameter of the used car price index curve of the same model.

[0025] In this embodiment, the first data buffer is a first physical address space in the memory that is configured with exclusive read and write permissions for the data acquisition module and the residual value decay prediction module. Specifically, a continuous physical address space is allocated in the memory, and the read and write permissions are granted only to the data acquisition module and the residual value decay prediction module through the memory management unit.

[0026] In this embodiment, static attribute electronic data and dynamic time-series characteristic electronic data are loaded into the first data buffer in the memory. Specifically, the data acquisition module serializes the acquired data into a byte stream and writes it into the first physical address space corresponding to the first data buffer through direct memory access.

[0027] In this embodiment, a first neural network model weight file with a time-series feature encoder is built in. The time-series feature encoder uses a long short-term memory network computing unit. The first neural network model weight file contains the connection weight matrix and bias vector of each gating unit in the long short-term memory network computing unit.

[0028] In this embodiment, the processor calls the weight file of the first neural network model with built-in time-series feature encoder and reads the dynamic time-series feature electronic data in the first data buffer. Specifically, the residual value decay prediction module loads the weight file of the first neural network model from the memory, reads the dynamic time-series feature electronic data from the first data buffer, and inputs it into the long short-term memory network computing unit for processing.

[0029] In this embodiment, the preset mortgage period is the loan term stipulated in the mortgage loan contract, in the form of days or months, which is input by the user through the front-end interactive device or preset by the system according to the vehicle type.

[0030] In this embodiment, tensor data representing the probability distribution of the predicted disposal residual value of the mortgaged vehicle at the expiration of the preset mortgage period is calculated and generated based on the preset mortgage period duration. Specifically, the residual value decay prediction module inputs the preset mortgage period duration as a time step parameter into the long short-term memory network computing unit, and the long short-term memory network computing unit outputs the tensor data representing the probability distribution of the predicted disposal residual value. The tensor data includes a mean parameter and a variance parameter. The mean parameter represents the expected value of the residual value, and the variance parameter represents the degree of uncertainty of the residual value.

[0031] In this embodiment, the processor performs a quantile extraction algorithm on the tensor data and extracts the pressure residual value that meets the preset risk confidence threshold condition. Specifically, the risk constraint construction module calculates the risk value function value corresponding to the preset tail probability threshold based on the mean parameter and variance parameter in the tensor data, and determines the risk value function value as the pressure residual value.

[0032] In this embodiment, the preset risk confidence threshold condition corresponds to the preset tail probability threshold, which is a pre-set probability value used to determine the confidence level corresponding to the pressure residual value.

[0033] In this embodiment, the pressure residual value is written into the second data cache in the memory. Specifically, the risk constraint construction module writes the pressure residual value into the second physical address space corresponding to the second data cache through direct memory access.

[0034] In this embodiment, the second data cache is a second physical address space in the memory that is configured with exclusive read and write permissions for the risk constraint construction module and the reinforcement learning pricing module. Specifically, another continuous physical address space is allocated in the memory, and the read and write permissions are granted only to the risk constraint construction module and the reinforcement learning pricing module through the memory management unit.

[0035] In this embodiment, the second physical address space is isolated from the first physical address space. Specifically, the address range of the first physical address space does not overlap with the address range of the second physical address space, and the memory management unit prohibits direct data transmission between the first physical address space and the second physical address space.

[0036] In this embodiment, the processor calls the second neural network model weight file and reads the pressure residual value from the second data cache. Specifically, the reinforcement learning pricing module loads the second neural network model weight file from the memory and reads the pressure residual value from the second data cache.

[0037] In this embodiment, the vehicle purchase price field in the static attribute electronic data is the manufacturer's suggested retail price or the actual transaction price at the time of the initial registration of the mortgaged vehicle, which is extracted from the static attribute electronic data.

[0038] In this embodiment, an asymmetric action exploration constraint function is constructed based on the ratio of the pressure residual value to the vehicle purchase price field in the static attribute electronic data. Specifically, the reinforcement learning pricing module calculates the quotient of the pressure residual value divided by the vehicle purchase price field, and uses this quotient as the independent variable of the asymmetric action exploration constraint function.

[0039] In this embodiment, the asymmetric action exploration constraint function narrows the upper limit boundary of the loan-to-value ratio dimension and raises the lower limit boundary of the mortgage rate dimension as the ratio decreases. Specifically, the asymmetric action exploration constraint function adopts a piecewise linear function. When the ratio is less than the first inflection point threshold, the upper limit boundary of the loan-to-value ratio decreases with the ratio at a first slope. When the ratio is greater than the second inflection point threshold, the lower limit boundary of the mortgage rate increases with the ratio at a second slope.

[0040] In this embodiment, narrowing the upper limit boundary of the collateral ratio dimension means reducing the maximum value of the collateral ratio floating-point value that the reinforcement learning pricing module is allowed to output in the action space.

[0041] In this embodiment, raising the lower bound of the collateral interest rate dimension refers to raising the minimum value of the collateral interest rate floating-point value that the reinforcement learning pricing module is allowed to output in the action space.

[0042] In this embodiment, the asymmetric action exploration constraint function limits the lower boundary of the collateral interest rate dimension and the upper boundary of the collateral ratio dimension in the action space of the second neural network model, and outputs floating-point values ​​for the collateral interest rate and the collateral ratio. Specifically, the reinforcement learning pricing module, based on the original action value output by the second neural network model, compares the original collateral interest rate value with the lower boundary of the collateral interest rate dimension and takes the larger value as the floating-point value of the collateral interest rate, and compares the original collateral ratio value with the upper boundary of the collateral ratio dimension and takes the smaller value as the floating-point value of the collateral ratio.

[0043] In this embodiment, the floating-point values ​​of the mortgage interest rate and the mortgage ratio are encapsulated into electronic messages and sent to the front-end interactive device. Specifically, the pricing output interface encapsulates the floating-point values ​​of the mortgage interest rate and the mortgage ratio into response messages according to the Hypertext Transfer Protocol, encrypts them through the Secure Transmission Layer Protocol, and then sends them to the front-end interactive device for display.

[0044] In this embodiment, taking a mortgaged vehicle with a purchase price of 200,000 yuan in its static attribute electronic data as an example, the preset mortgage period is 12 months. The data acquisition module obtains the vehicle's dynamic time-series characteristic electronic data, including the time-series encoding of the vehicle's historical maintenance records over the past 3 years, the monthly cumulative mileage increment sequence over the past 12 months, and the slope parameter of the used car price index curve for the same model. The long short-term memory network calculation unit in the residual value decay prediction module converts the dynamic time-series characteristic electronic data into a vehicle life cycle feature vector and outputs tensor data predicting the probability distribution of the disposal residual value, where the mean parameter is 120,000 yuan and the variance parameter is 20,000 yuan. The risk constraint construction module sets a preset tail probability threshold of 5%, and performs a quantile extraction algorithm based on the mean and variance parameters to calculate a pressure residual value of 87,000 yuan. The reinforcement learning pricing module calculates the ratio of the residual value to the vehicle purchase price field to be 0.435. Based on this ratio, an asymmetric action exploration constraint function is constructed, determining the lower bound of the mortgage interest rate dimension as 10.5% and the upper bound of the loan-to-value ratio dimension as 0.65. The second neural network model outputs an initial mortgage interest rate adjustment of +2% and an initial loan-to-value ratio adjustment of +0.1. After being constrained by the asymmetric action exploration constraint function, the output mortgage interest rate floating-point value is 12.5% ​​and the loan-to-value ratio floating-point value is 0.65. The pricing output interface encapsulates these values ​​into an electronic message and sends it to the front-end interactive device. The mortgage loan matures in 12 months, with an actual residual value of 90,000 yuan. The loan amount corresponding to the loan-to-value ratio floating-point value is 130,000 yuan, a ratio of 1.44, which is greater than the preset bad debt threshold constant of 1, triggering a negative gradient backpropagation signal and updating the weight parameters of the second neural network model.

[0045] Furthermore, the dynamic time-series characteristic electronic data acquired by the data acquisition module includes the time-series encoding of the vehicle's historical maintenance and repair records, the monthly cumulative mileage increment numerical sequence, and the slope parameter of the used car price index curve for the same model. The residual value decay prediction module has a built-in time-series feature encoder, which is a long short-term memory network computing unit. The long short-term memory network computing unit is used to convert dynamic time-series feature electronic data into a fixed-dimensional vehicle life cycle feature vector. The tensor data generated by the residual value decay prediction module, which characterizes the probability distribution of the predicted disposal residual value, includes mean and variance parameters. The risk constraint construction module executes a quantile extraction algorithm that calculates the risk value function value corresponding to the preset tail probability threshold based on the mean parameter and variance parameter, and determines the risk value function value as the stress residual value.

[0046] In this embodiment, the time-series encoding of the vehicle's historical maintenance records is achieved by arranging the time of each maintenance, the type of maintenance item, the name of the replaced parts, and the amount of maintenance cost in chronological order, and then converting them into a fixed-length numerical vector sequence using a one-hot encoding method; the monthly cumulative mileage increment numerical sequence is a numerical sequence composed of the mileage increase counted for each natural month; and the slope parameter of the used car price index curve for the same model is the slope value of the curve of the average transaction price of the same brand and model of used cars over time obtained from the used car trading platform in each time interval.

[0047] In this embodiment, the Long Short-Term Memory (LSTM) network computing unit is used to convert dynamic temporal feature electronic data into a fixed-dimensional vehicle lifecycle feature vector. Specifically, the LSM network computing unit sequentially reads the time-series encoding of the vehicle's historical maintenance records, the monthly cumulative mileage increment numerical sequence, and the slope parameter of the used car price index curve for the same model. At each time step, it updates the memory cell state and hidden state, and outputs the hidden state of the last time step as a fixed-dimensional vehicle lifecycle feature vector. This vehicle lifecycle feature vector integrates the temporal correlation information in the vehicle's historical maintenance records, the monthly cumulative mileage increment numerical sequence, and the slope parameter of the used car price index curve for the same model.

[0048] In this embodiment, the tensor data generated by the residual value decay prediction module, which characterizes the probability distribution of the predicted disposal residual value, includes a mean parameter and a variance parameter. The mean parameter represents the expected disposal residual value of the mortgaged vehicle at the expiration of the preset mortgage period, and the variance parameter represents the degree of uncertainty in the predicted disposal residual value. The larger the value of the variance parameter, the higher the uncertainty of the prediction.

[0049] In this embodiment, the preset tail probability threshold is a pre-set probability value used to determine the confidence level corresponding to the pressure residual value. The preset tail probability threshold is set to 1% or 5%, with 1% corresponding to a 99% confidence level and 5% corresponding to a 95% confidence level.

[0050] In this embodiment, the quantile extraction algorithm executed by the risk constraint construction module calculates the risk value function value corresponding to the preset tail probability threshold based on the mean parameter and variance parameter, and determines the risk value function value as the stress residual value. Specifically, the risk constraint construction module inputs the mean parameter, variance parameter, and preset tail probability threshold into the quantile extraction algorithm. The quantile extraction algorithm calculates the disposal residual value amount at the quantile position corresponding to the preset tail probability threshold according to the normal distribution assumption or the log-normal distribution assumption, uses the disposal residual value amount as the risk value function value, and determines the risk value function value as the stress residual value.

[0051] Furthermore, the input layer state vector dimension of the reinforcement learning pricing module includes the encoding dimension of static attribute electronic data, the feature dimension of vehicle life cycle feature vector, and the scalar dimension of pressure residual value. The action output layer of the second neural network model includes output nodes for the mortgage rate adjustment range and the mortgage ratio adjustment range.

[0052] In this embodiment, the encoding dimension of the static attribute electronic data is a vector dimension formed by concatenating the vehicle identification number, brand and model, initial registration date, vehicle purchase price field, engine displacement, transmission type, and body color after numerical processing; the feature dimension of the vehicle life cycle feature vector is the dimension of the fixed-dimensional vehicle life cycle feature vector output by the long short-term memory network computing unit; the scalar dimension of the pressure residual value is a one-dimensional scalar; the reinforcement learning pricing module concatenates the encoding dimension of the static attribute electronic data, the feature dimension of the vehicle life cycle feature vector, and the scalar dimension of the pressure residual value as the state vector of the input layer of the second neural network model.

[0053] In this embodiment, the mortgage rate adjustment range output node and the mortgage ratio adjustment range output node are two output nodes of the action output layer of the second neural network model. The mortgage rate adjustment range output node outputs a continuous value representing the adjustment range relative to the benchmark interest rate, and the mortgage ratio adjustment range output node outputs a continuous value representing the adjustment range relative to the benchmark ratio. The reinforcement learning pricing module adds the output value of the mortgage rate adjustment range output node to the benchmark interest rate to obtain the floating-point value of the mortgage rate, and adds the output value of the mortgage ratio adjustment range output node to the benchmark ratio to obtain the floating-point value of the mortgage ratio.

[0054] Furthermore, the reinforcement learning pricing module has a built-in reward signal calculation submodule; In offline training mode, the reward signal calculation submodule is used to calculate the instantaneous reward value; When the ratio of the collateral ratio floating-point value output by the second neural network model to the electronic data of the actual residual value of the mortgaged vehicle obtained by the data acquisition module at the expiration of the mortgage period is greater than the preset bad debt threshold constant, the reward signal calculation submodule outputs a negative gradient back transmission signal, triggering the update process of the weight parameters of the second neural network model to suppress the power consumption of the processor performing invalid pricing calculations. When the ratio of the collateral ratio floating-point value output by the second neural network model to the actual residual value electronic data at the end of the collateral term is less than or equal to the preset bad debt threshold constant, and the collateral interest rate floating-point value meets the preset income threshold condition, the reward signal calculation submodule outputs a positive gradient feedback signal to solidify the weight parameters of the second neural network model.

[0055] In this embodiment, calculating the immediate reward value refers to the reward signal calculation submodule built into the reinforcement learning pricing module calculating the reward value based on the comparison result between the electronic data of the actual residual value obtained after the expiration of the mortgage period and the floating value of the mortgage ratio after each output of the floating value of the mortgage interest rate and the floating value of the mortgage ratio. This reward value is used to guide the optimization direction of the weight parameters of the second neural network model.

[0056] In this embodiment, the preset bad debt threshold constant is a pre-set numerical constant used to define the bad debt judgment boundary of the ratio of the floating-point value of the collateral percentage to the electronic data of the actual residual value. The preset bad debt threshold constant is set to the value one. When the ratio is greater than the value one, it means that the floating-point value of the collateral percentage exceeds the electronic data of the actual residual value, and a bad debt occurs.

[0057] In this embodiment, the ratio of the floating-point value of the mortgage percentage output by the second neural network model to the electronic data of the actual residual value of the mortgaged vehicle obtained by the data acquisition module at the time of the mortgage term is specifically the ratio obtained by dividing the floating-point value of the mortgage percentage by the electronic data of the actual residual value. The electronic data of the actual residual value is the amount actually recovered by the mortgaged vehicle through auction or disposal channels after the expiration of the mortgage term.

[0058] In this embodiment, the negative gradient backpropagation signal is the negative gradient signal output by the reward signal calculation submodule when the ratio is greater than the preset bad debt threshold constant. The negative gradient backpropagation signal is transmitted to the weight parameters of each layer of the second neural network model through the backpropagation algorithm, triggering the weight parameters to be updated in the direction of reducing bad debt risk.

[0059] In this embodiment, the update process of the weight parameters of the second neural network model is triggered to suppress the power consumption of the processor performing invalid pricing calculations. Specifically, the weight parameters of the second neural network model are adjusted in the direction of narrowing the upper limit of the collateral ratio by using a negative gradient backpropagation signal, thereby reducing the number of invalid exploration calculations for high-collateral-ratio and low-interest-rate combinations in the subsequent pricing process, and thus reducing the power consumption of the processor performing invalid forward propagation calculations.

[0060] In this embodiment, the floating-point value of the mortgage rate is the annualized mortgage rate finally output by the reinforcement learning pricing module, stored in floating-point format, with a value range between 5% and 24%.

[0061] In this embodiment, the preset return threshold condition is the minimum return requirement of the pre-set floating value of the mortgage interest rate. When the floating value of the mortgage interest rate is greater than or equal to the preset minimum return interest rate threshold, it is determined that the preset return threshold condition is met. The preset minimum return interest rate threshold is set to eight percent.

[0062] In this embodiment, the positive gradient backpropagation signal is a positive gradient signal output by the reward signal calculation submodule when the ratio is less than or equal to the preset bad debt threshold constant and the floating-point value of the collateral interest rate meets the preset return threshold condition. The positive gradient backpropagation signal is transmitted to the weight parameters of each layer of the second neural network model through the backpropagation algorithm, triggering the weight parameters to solidify in the direction of maintaining the current pricing strategy.

[0063] In this embodiment, the weight parameters of the solidified second neural network model refer to the gradient magnitude of the positive gradient backpropagation signal being close to zero, the update amount of the weight parameters of each layer of the second neural network model being less than the preset convergence threshold, and the weight parameters tending to stabilize and no longer changing significantly.

[0064] Furthermore, it also includes a risk rule fusion module, which runs in the processor; The risk rule fusion module loads a set of expert risk hard boundary rules pre-stored in the memory. The set of expert risk hard boundary rules includes the maximum allowable vehicle age threshold variable, the minimum allowable mortgage residual value multiple threshold variable, and the upper and lower limit threshold variables of the mortgage interest rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. The risk rule fusion module calls the comparator logic circuit to perform the following judgments: compare the vehicle age value of the mortgaged vehicle with the maximum allowable vehicle age threshold variable, and compare the ratio of the residual value of the pressure to the vehicle purchase price field with the minimum allowable residual value multiple threshold variable. When the vehicle age value exceeds the maximum allowable vehicle age threshold, an interest rate adjustment is triggered. When the ratio of the residual value of the pressure to the vehicle purchase price field is lower than the minimum allowable residual value multiple threshold variable, the upper limit boundary of the mortgage ratio dimension will be forcibly set to zero. After the reinforcement learning pricing module outputs the original floating-point value of the mortgage rate, the risk rule fusion module determines whether the original floating-point value of the mortgage rate conforms to the upper and lower limit threshold variables of the mortgage rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. If it does not conform, the original floating-point value of the mortgage rate is truncated and corrected to generate a corrected floating-point value of the mortgage rate that meets the hard boundary conditions.

[0065] In this embodiment, the maximum allowable vehicle age threshold variable is a pre-set upper limit value for the age of the mortgaged vehicle, which is set to ten years; the minimum allowable residual value multiple threshold variable is a pre-set minimum allowable multiple of the ratio of the residual value to the vehicle purchase price field, which is set to 0.5 times; the upper and lower limit threshold variables of the mortgage interest rate corresponding to the vehicle age threshold variable and the residual value multiple threshold variable include the interest rate increase corresponding to each year of vehicle age exceeding one year and the interest rate increase corresponding to each 0.1 times decrease in the residual value multiple, with the upper limit threshold variable of the mortgage interest rate set to 24 percent and the lower limit threshold variable of the mortgage interest rate set to 5 percent.

[0066] In this embodiment, the risk rule fusion module calls the comparator logic circuit to perform the following judgments: comparing the vehicle's age value with the maximum allowable vehicle age threshold variable, and comparing the ratio of the residual value to the vehicle purchase price field with the minimum allowable residual value multiple threshold variable. When the vehicle age value exceeds the maximum allowable vehicle age threshold variable, an interest rate adjustment is triggered, and the lower limit threshold variable of the mortgage interest rate is adjusted according to the rule of increasing by a certain amount for each year the vehicle age exceeds the threshold variable. When the ratio of the residual value to the vehicle purchase price field is lower than the minimum allowable residual value multiple threshold variable, the upper limit boundary of the loan-to-value ratio dimension is forcibly set to zero, indicating that the mortgaged vehicle does not meet the mortgage conditions.

[0067] In this embodiment, the original floating-point value of the mortgage rate is the floating-point value of the mortgage rate output by the reinforcement learning pricing module after applying the asymmetric action exploration constraint function, which has not been verified by the risk rule fusion module.

[0068] In this embodiment, after the reinforcement learning pricing module outputs the original floating-point value of the mortgage rate, the risk rule fusion module determines whether the original floating-point value of the mortgage rate conforms to the upper and lower limit threshold variables of the mortgage rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. Specifically, it determines whether the original floating-point value of the mortgage rate is located between the lower limit threshold variable and the upper limit threshold variable of the mortgage rate.

[0069] In this embodiment, the original floating-point value of the mortgage rate is truncated to generate a modified floating-point value of the mortgage rate that satisfies the hard boundary conditions. Specifically, if the original floating-point value of the mortgage rate is less than the lower limit threshold variable of the mortgage rate, the original floating-point value of the mortgage rate is modified to the lower limit threshold variable of the mortgage rate. If the original floating-point value of the mortgage rate is greater than the upper limit threshold variable of the mortgage rate, the original floating-point value of the mortgage rate is modified to the upper limit threshold variable of the mortgage rate. If the original floating-point value of the mortgage rate is between the lower limit threshold variable of the mortgage rate and the upper limit threshold variable of the mortgage rate, the original floating-point value of the mortgage rate remains unchanged. The modified value is used as the modified floating-point value of the mortgage rate.

[0070] Furthermore, the residual value decay prediction module and the reinforcement learning pricing module are configured with an asynchronous clock signal triggering mechanism; The weight file of the first neural network model in the residual decay prediction module is updated with parameters at the end of the first clock cycle. The input layer of the second neural network model in the reinforcement learning pricing module receives the new state vector dimension boundary generated by the updated weight file of the first neural network model after the first clock cycle ends, and performs policy iterative update calculation in the second clock cycle.

[0071] In this embodiment, the residual value decay prediction module and the reinforcement learning pricing module are configured with an asynchronous clock signal triggering mechanism. Specifically, the processor allocates a first clock signal source to the residual value decay prediction module and a second clock signal source to the reinforcement learning pricing module. The first clock cycle of the first clock signal source and the second clock cycle of the second clock signal source are independent of each other. The first clock cycle is used to trigger the parameter update calculation of the residual value decay prediction module, and the second clock cycle is used to trigger the policy iteration calculation of the reinforcement learning pricing module.

[0072] In this embodiment, the weight file of the first neural network model in the residual decay prediction module performs parameter updates at the end of the first clock cycle. Specifically, whenever the rising edge of the first clock signal arrives, the residual decay prediction module starts a forward and backward propagation calculation. At the end of the first clock cycle, the gradient update of the connection weight matrix and bias vector of each gate unit of the first neural network model is completed, and the updated weight parameters are written into the weight file of the first neural network model.

[0073] In this embodiment, the input layer of the second neural network model in the reinforcement learning pricing module receives the new state vector dimension boundary generated by the updated first neural network model weight file after the first clock cycle ends. In the second clock cycle, it performs a policy iteration update calculation. Specifically, after the first clock cycle ends, the residual value decay prediction module sends a parameter update completion signal to the reinforcement learning pricing module. After receiving the parameter update completion signal, the reinforcement learning pricing module reads the pressure residual value recalculated based on the updated first neural network model weight file from the second data buffer, reconstructs the asymmetric action exploration constraint function to generate a new state vector dimension boundary, and when the rising edge of the second clock signal arrives, the reinforcement learning pricing module performs a policy iteration update calculation based on the new state vector dimension boundary.

[0074] In this embodiment, the new state vector dimension boundary generated by the updated first neural network model weight file refers to the tensor data of the predicted disposal residual probability distribution recalculated based on the new first neural network model weight file after the residual value decay prediction module completes the parameter update in the first clock cycle. The risk constraint construction module re-extracts the pressure residual value based on the new tensor data, and the reinforcement learning pricing module reconstructs the asymmetric action exploration constraint function based on the re-extracted pressure residual value, thereby updating the lower boundary of the mortgage rate dimension and the upper boundary of the mortgage ratio dimension.

[0075] Furthermore, the data acquisition module is also equipped with an on-board diagnostic system signal receiving port, which is used to acquire in real time the on-board diagnostic system fault code data and the cumulative number of traffic violations of the mortgaged vehicle during the mortgage period; The risk constraint construction module obtains the first attenuation coefficient by querying a preset fault code attenuation coefficient mapping table based on the fault code data of the vehicle diagnostic system, and obtains the second attenuation coefficient by querying a preset violation number attenuation coefficient mapping table based on the cumulative number of traffic violations. The product of the first attenuation coefficient and the second attenuation coefficient is used as the dynamic attenuation coefficient. The pressure residual value stored in the second data cache is dynamically attenuated and corrected by multiplication. The corrected pressure residual value is then updated in real time to the address space of the second data cache of the reinforcement learning pricing module through atomic synchronization operation. This triggers the reinforcement learning pricing module to recalculate the floating-point values ​​of the mortgage rate and the mortgage ratio based on the corrected pressure residual value.

[0076] In this embodiment, the vehicle diagnostic system fault code data is the vehicle fault diagnosis code that the mortgaged vehicle obtains in real time through the vehicle diagnostic system signal receiving port during the mortgage period. The fault code data includes power system fault codes, chassis system fault codes, body electronic system fault codes and network communication system fault codes, and each fault code corresponds to a specific vehicle component fault type.

[0077] In this embodiment, the cumulative number of traffic violations is the cumulative number of unprocessed traffic violation records obtained by the mortgaged vehicle through the traffic management data interface during the mortgage period.

[0078] In this embodiment, the preset fault code attenuation coefficient mapping table is a two-dimensional mapping table pre-stored in the memory. The first column of the mapping table is the fault code type identifier, and the second column is the corresponding first attenuation coefficient value. The fault code type identifier is divided into three levels according to the severity of the fault: minor fault, general fault, and serious fault. The first attenuation coefficient corresponding to minor fault is 0.98, the first attenuation coefficient corresponding to general fault is 0.95, and the first attenuation coefficient corresponding to serious fault is 0.90.

[0079] In this embodiment, the first attenuation coefficient is obtained by querying a preset fault code attenuation coefficient mapping table based on the fault code data of the vehicle diagnostic system. Specifically, the risk constraint construction module matches the fault code type identifier in the received fault code data of the vehicle diagnostic system with the first column of the preset fault code attenuation coefficient mapping table, and reads the first attenuation coefficient value in the second column of the matching row.

[0080] In this embodiment, the preset violation count attenuation coefficient mapping table is a two-dimensional mapping table pre-stored in the memory. The first column of the mapping table is the range of the cumulative number of traffic violations, and the second column is the corresponding second attenuation coefficient value. Zero cumulative violations correspond to a second attenuation coefficient of 1.00, one to two cumulative violations correspond to a second attenuation coefficient of 0.97, three to five cumulative violations correspond to a second attenuation coefficient of 0.93, and six or more cumulative violations correspond to a second attenuation coefficient of 0.88.

[0081] In this embodiment, the second attenuation coefficient is obtained by querying the preset violation attenuation coefficient mapping table based on the cumulative number of traffic violations. Specifically, the risk constraint construction module matches the cumulative number of traffic violations with the range of the first column of the preset violation attenuation coefficient mapping table and reads the value of the second attenuation coefficient in the second column of the matching row.

[0082] In this embodiment, the pressure residual value stored in the second data buffer is corrected by dynamic attenuation coefficient multiplication. Specifically, the risk constraint construction module uses the product of the first attenuation coefficient and the second attenuation coefficient as the dynamic attenuation coefficient, multiplies the pressure residual value stored in the second data buffer with the dynamic attenuation coefficient, and uses the product result as the corrected pressure residual value.

[0083] In this embodiment, the corrected residual pressure value is synchronized in real time to the address space of the second data cache of the reinforcement learning pricing module, and the reinforcement learning pricing module is triggered to recalculate the floating-point values ​​of the mortgage interest rate and the mortgage ratio based on the corrected residual pressure value. Specifically, the risk constraint construction module writes the corrected residual pressure value into the storage address corresponding to the residual pressure value in the second data cache, overwriting the original residual pressure value, and sends a residual pressure update trigger signal to the reinforcement learning pricing module. After receiving the residual pressure update trigger signal, the reinforcement learning pricing module reads the corrected residual pressure value from the second data cache, reconstructs the asymmetric action exploration constraint function based on the ratio of the corrected residual pressure value to the vehicle purchase price field in the static attribute electronic data, and recalculates and outputs the floating-point values ​​of the mortgage interest rate and the mortgage ratio.

[0084] Furthermore, it also includes a module for generating a visual interactive interface for mortgage risk assessment; The module for generating a visual interactive interface for mortgage risk assessment is used to drive the display device to render a graphical user interface that includes a pressure residual value decay curve layer and a pricing confidence interval dashboard layer; The pressure residual value decay curve layer uses the time axis as the horizontal axis and the mean of the pressure residual value value and the probability distribution of the predicted disposal residual value as the vertical axis. It displays the safety boundary of vehicle value during the mortgage period in real time and responds to the user's drag operation of the vertex control on the time axis by sending a mortgage period change signal to the risk constraint construction module to trigger the recalculation of the pressure residual value value. The pricing confidence interval dashboard layer is used to respond to the floating-point value of the collateral rate output by the reinforcement learning pricing module. It marks the risk-return balance range of the current floating-point value of the collateral rate on the dashboard, and responds to the user's drag operation on the target interest rate pointer on the dashboard. It sends an interest rate adjustment signal to the reinforcement learning pricing module to trigger the recalculation of the floating-point value of the collateral ratio.

[0085] In this embodiment, the pressure residual value decay curve layer is a two-dimensional curve drawing layer in the graphical user interface. The horizontal axis represents the mortgage term time axis, and the vertical axis represents the amount value axis. The layer draws the decay curve of the pressure residual value over time and the expected curve of the mean of the predicted disposal residual value probability distribution over time.

[0086] In this embodiment, the pricing confidence interval dashboard layer is a semi-circular dashboard drawing layer in the graphical user interface. The dashboard scale range from 5% to 24% corresponds to the allowable range of floating point values ​​of the mortgage rate. The dashboard is divided into three fan-shaped areas: low-risk return range, balanced risk return range, and high-risk return range.

[0087] In this embodiment, the driver display device renders a graphical user interface that includes a pressure residual value decay curve layer and a pricing confidence interval dashboard layer. Specifically, the mortgage risk assessment visualization interactive interface generation module calls the graphics processor rendering pipeline to synthesize the curve data of the pressure residual value decay curve layer and the dashboard data of the pricing confidence interval dashboard layer into a frame of image buffer data, and outputs the image buffer data to the display device for display through the display interface.

[0088] In this embodiment, the mean of the predicted residual value probability distribution is the mean parameter value in the tensor data of the predicted residual value probability distribution generated by the residual value decay prediction module, which represents the expected residual value of the mortgaged vehicle at the time of expiration of the preset mortgage period.

[0089] In this embodiment, the cursor control on the time axis is a sliding selection control on the pressure residual value decay curve layer that can respond to user drag operations. The position of the cursor control on the time axis corresponds to a mortgage term value, and the user can change the mortgage term value by dragging the cursor control.

[0090] In this embodiment, the mortgage term change signal is a digital signal generated by the mortgage risk assessment visualization interactive interface generation module in response to the user's drag operation of the cursor control on the timeline. The mortgage term change signal includes the new mortgage term value selected by the user through the cursor control.

[0091] In this embodiment, the pressure residual value decay curve layer displays the vehicle value safety boundary during the mortgage period in real time, with the time axis as the horizontal axis and the mean of the pressure residual value and the predicted disposal residual value probability distribution as the vertical axis. In response to the user's dragging operation of the cursor control on the time axis, it sends a mortgage period change signal to the risk constraint construction module to trigger the recalculation of the pressure residual value. Specifically, the mortgage risk assessment visualization interactive interface generation module draws the pressure residual value corresponding to each time point within the preset mortgage period as the first curve and the mean of the predicted disposal residual value probability distribution corresponding to each time point as the second curve. The vertical axis difference area between the first curve and the second curve is filled with a semi-transparent color to represent the vehicle value safety boundary. When the user drags the cursor control, the mortgage risk assessment visualization interactive interface generation module obtains the new mortgage period value corresponding to the current position of the cursor control, generates a mortgage period change signal containing the new mortgage period value, and sends the mortgage period change signal to the risk constraint construction module. After receiving the mortgage period change signal, the risk constraint construction module recalculates the tensor data of the predicted disposal residual value probability distribution based on the new mortgage period value and re-extracts the pressure residual value.

[0092] In this embodiment, in response to the floating-point value of the mortgage interest rate output by the reinforcement learning pricing module, the risk-return balance range of the current floating-point value of the mortgage interest rate is marked on the dashboard. Specifically, the mortgage risk assessment visualization interface generation module receives the floating-point value of the mortgage interest rate output by the reinforcement learning pricing module, maps the floating-point value of the mortgage interest rate to the scale position of the pricing confidence interval dashboard layer, and drives the dashboard pointer to rotate to the corresponding scale position. The sector area where the pointer is located is the risk-return balance range of the current floating-point value of the mortgage interest rate.

[0093] In this embodiment, the risk-return balance range where the current floating value of the mortgage rate is located refers to the sector area that the pointer on the pricing confidence interval dashboard layer is currently pointing to. The low-risk return range means that the floating value of the mortgage rate is at a low level, the institution's return is low, but the default risk is small. The balanced risk-return range means that the floating value of the mortgage rate is at a moderate level, achieving a balance between return and risk. The high-risk return range means that the floating value of the mortgage rate is at a high level, the institution's return is high, but the default risk is large.

[0094] In this embodiment, in response to the user's drag operation on the target interest rate pointer on the dashboard, an interest rate adjustment signal is sent to the reinforcement learning pricing module to trigger the linkage recalculation of the collateral ratio floating-point value. Specifically, the collateral risk assessment visualization interface generation module responds to the user's drag operation on the target interest rate pointer on the dashboard, obtains the target collateral interest rate value pointed to by the pointer after rotation, generates an interest rate adjustment signal containing the target collateral interest rate value, and sends the interest rate adjustment signal to the reinforcement learning pricing module. After receiving the interest rate adjustment signal, the reinforcement learning pricing module uses the target collateral interest rate value as the new lower limit boundary of the collateral interest rate dimension, re-executes the asymmetric action exploration constraint function calculation based on the new lower limit boundary of the collateral interest rate dimension, and outputs the linkage recalculated collateral ratio floating-point value that matches the target collateral interest rate value.

[0095] Furthermore, the augmented learning pricing module includes an online inference working mode and an offline training working mode; When the reinforcement learning pricing module is in online inference mode, the processor only calls the weight parameter file of the second neural network model stored in memory to perform forward propagation calculation to generate the floating-point value of the mortgage rate. It does not perform backpropagation gradient update calculation. It obtains the asymmetric action exploration constraint function parameters corresponding to the current pressure residual value by reading the mapping table data pre-calculated based on the preset risk confidence threshold. Based on the asymmetric action exploration constraint function parameters, it determines the lower boundary of the mortgage rate dimension and the upper boundary of the mortgage ratio dimension.

[0096] In this embodiment, the online inference working mode refers to the real-time response state of the reinforcement learning pricing module when it receives a pricing request from the mortgage application processing unit. In the online inference working mode, the reinforcement learning pricing module does not perform model training, but only uses the trained model parameters to perform forward calculation and output the pricing result.

[0097] In this embodiment, the offline training mode refers to the running state in which the reinforcement learning pricing module uses historical collateral transaction data to batch train the second neural network model during system idle periods or in the background server. In the offline training mode, the reinforcement learning pricing module performs complete forward propagation calculation and backpropagation gradient update calculation to optimize the weight parameters of the second neural network model.

[0098] In this embodiment, the processor only calls the weight parameter file of the second neural network model stored in the memory to perform forward propagation calculation to generate the floating-point value of the mortgage rate, without performing backpropagation gradient update calculation. Specifically, when the reinforcement learning pricing module is in online inference mode, the processor reads the weight parameter file of the second neural network model that has been trained offline from the memory, concatenates the encoding dimension of the static attribute electronic data, the feature dimension of the vehicle life cycle feature vector, and the scalar dimension of the pressure residual value, and inputs them into the input layer of the second neural network model. Matrix multiplication and activation function calculation are performed layer by layer until the output layer outputs the mortgage rate adjustment range and the mortgage ratio adjustment range. During the calculation process, the loss function is not calculated and gradient backpropagation is not performed.

[0099] In this embodiment, the preset risk confidence threshold is the same value as the preset risk confidence threshold condition in claim 1. The preset risk confidence threshold is a pre-set confidence level parameter used to extract the pressure residual value. The preset risk confidence threshold is set to 95% or 99%.

[0100] In this embodiment, the mapping table data pre-calculated based on the preset risk confidence threshold is a two-dimensional lookup table pre-calculated and stored in the memory during the offline training phase. The first column of the mapping table data is the discretized interval of the ratio of the pressure residual value to the vehicle purchase price field, and the second column is the asymmetric action exploration constraint function parameter corresponding to each discretized interval. The mapping table data is pre-calculated based on the preset risk confidence threshold and the historical output statistical distribution of the residual decay prediction model.

[0101] In this embodiment, the asymmetric action exploration constraint function parameters corresponding to the current pressure residual value are obtained by reading the mapping table data pre-calculated based on the preset risk confidence threshold. Specifically, the reinforcement learning pricing module calculates the ratio of the current pressure residual value to the vehicle purchase price field in the static attribute electronic data, matches the ratio value with the discretized interval of the first column of the mapping table data, and reads the asymmetric action exploration constraint function parameters of the second column of the matching row.

[0102] In this embodiment, the parameters of the asymmetric action exploration constraint function corresponding to the current pressure residual value include the slope parameter of the upper boundary of the collateral ratio dimension, the intercept parameter of the upper boundary of the collateral ratio dimension, the slope parameter of the lower boundary of the collateral interest rate dimension, and the intercept parameter of the lower boundary of the collateral interest rate dimension. These parameters together define the specific functional form of the asymmetric action exploration constraint function under the current pressure residual value.

[0103] In this embodiment, the lower bound of the mortgage interest rate dimension and the upper bound of the loan-to-value ratio dimension are determined based on the parameters of the asymmetric action exploration constraint function. Specifically, the reinforcement learning pricing module substitutes the ratio of the residual value of the pressure to the vehicle purchase price field into a piecewise linear function defined by the parameters of the asymmetric action exploration constraint function to calculate the lower bound of the mortgage interest rate dimension and the upper bound of the loan-to-value ratio dimension. The original mortgage interest rate value output by the second neural network model is compared with the lower bound of the mortgage interest rate dimension, and the larger value is taken as the floating-point value of the mortgage interest rate. The original loan-to-value ratio value is compared with the upper bound of the loan-to-value ratio dimension, and the smaller value is taken as the floating-point value of the loan-to-value ratio.

[0104] This invention provides an embodiment of a vehicle dynamic pricing platform based on the fusion of reinforcement learning and risk rules, comprising: The mortgage application processing unit is used to parse the received mortgage request data packet containing the vehicle identification number field and the preset mortgage term field; The risk assessment interface unit is used to initiate structured query language requests to external vehicle credit databases to obtain electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records. The pricing engine unit's processor is loaded with computer-executable instructions stored in memory, such as those described above, for a vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules, used to generate floating-point values ​​for the mortgage rate and the mortgage ratio. The result feedback unit is used to encapsulate the floating-point values ​​of the collateral interest rate and the collateral ratio generated by the pricing engine unit into a Hypertext Transfer Protocol response message and send it to the front-end interactive device.

[0105] In this embodiment, parsing the received mortgage request data packet containing the vehicle identification number field and the preset mortgage period field specifically involves the mortgage application processing unit receiving the Hypertext Transfer Protocol request data packet sent by the front-end interactive device, extracting the vehicle identification number field value and the preset mortgage period field value from the request data packet, and storing the vehicle identification number field value and the preset mortgage period field value into the mortgage application data structure.

[0106] In this embodiment, the vehicle identification code field is a unique identification code for the mortgaged vehicle, consisting of seventeen letters and numbers, including the vehicle manufacturer code, model code, production year code, assembly plant code, and serial number.

[0107] In this embodiment, the preset mortgage term field is the loan term value agreed upon in the mortgage loan contract, in units of months, with a value range of one month to thirty-six months.

[0108] In this embodiment, the external vehicle credit database is a vehicle historical information query database provided by a third-party vehicle data service agency, which stores electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records.

[0109] In this embodiment, a structured query language request is initiated to an external vehicle credit database to obtain electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records. Specifically, the risk assessment interface unit uses the vehicle identification number field value as the query condition to concatenate a structured query language statement, sends the query request to the external vehicle credit database through a secure transport layer protocol encrypted connection, and receives the electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records returned by the external vehicle credit database.

[0110] In this embodiment, the vehicle mortgage status record includes a flag indicating whether the vehicle is under mortgage, the name of the mortgagee, the mortgage registration date, and the mortgage release date. The accident record includes the number of accidents the vehicle has ever been involved in, the amount of compensation for each accident, and the type of liability for the accident.

[0111] In this embodiment, the processor of the pricing engine unit is loaded with computer-executable instructions stored in the memory, such as those described above, for a vehicle dynamic pricing system based on reinforcement learning and risk rule fusion, used to generate floating-point values ​​for the mortgage rate and the mortgage ratio. Specifically, the processor of the pricing engine unit reads the computer-executable instructions from the memory, executes the instructions of the data acquisition module to obtain static attribute electronic data and dynamic time-series feature electronic data, executes the instructions of the residual value decay prediction module to generate tensor data of the probability distribution of predicted disposal residual value, executes the instructions of the risk constraint construction module to extract the pressure residual value, executes the instructions of the reinforcement learning pricing module to construct an asymmetric action exploration constraint function and output the floating-point values ​​for the mortgage rate and the mortgage ratio, and executes the instructions of the risk rule fusion module to verify and correct the floating-point values ​​for the mortgage rate.

[0112] In this embodiment, the floating-point values ​​of the collateral interest rate and the collateral ratio generated by the pricing engine unit are encapsulated into a Hypertext Transfer Protocol (HTTP) response message and sent to the front-end interactive device. Specifically, the result feedback unit serializes and encapsulates the floating-point values ​​of the collateral interest rate and the collateral ratio according to the key-value pair format of the HTTP response message, sets the response status code to a success status, and sends the response message to the front-end interactive device for display through a secure transport layer protocol encrypted connection.

[0113] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules, running on a computer device including a processor and memory, characterized in that, include: The data acquisition module is configured to acquire static attribute electronic data and dynamic time-series characteristic electronic data of the mortgaged vehicle through the processor control data interface, and load the static attribute electronic data and dynamic time-series characteristic electronic data into the first data cache area in the memory. The first data cache area is the first physical address space in the memory that is configured with exclusive read and write permissions for the data acquisition module and the residual value decay prediction module. The residual value decay prediction module is configured to have the processor call the weight file of the first neural network model with built-in time-series feature encoder, read the dynamic time-series feature electronic data in the first data buffer, and calculate and generate tensor data representing the probability distribution of the predicted disposal residual value of the mortgaged vehicle at the time of the expiration of the preset mortgage period based on the duration of the preset mortgage period. The risk constraint construction module is configured to have the processor execute a quantile extraction algorithm on the tensor data to extract the pressure residual value that meets the preset risk confidence threshold condition, and write the pressure residual value into the second data cache in the memory. The second data cache is a second physical address space in the memory that is configured with exclusive read and write permissions for the risk constraint construction module and the reinforcement learning pricing module. The second physical address space is isolated from the first physical address space. The reinforcement learning pricing module is configured to have the processor call the weight file of the second neural network model, read the pressure residual value from the second data cache, and construct an asymmetric action exploration constraint function based on the ratio of the pressure residual value to the vehicle purchase price field in the static attribute electronic data. As the ratio decreases, the asymmetric action exploration constraint function synchronously narrows the upper limit boundary of the mortgage ratio dimension and raises the lower limit boundary of the mortgage interest rate dimension. The asymmetric action exploration constraint function limits the lower limit boundary of the mortgage interest rate dimension and the upper limit boundary of the mortgage ratio dimension in the action space of the second neural network model, respectively, and outputs the floating-point values ​​of the mortgage interest rate and the mortgage ratio. The pricing output interface is configured to encapsulate the floating-point values ​​of the mortgage rate and the mortgage percentage into electronic messages and send them to the front-end interactive device.

2. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules as described in claim 1, characterized in that, The dynamic time-series characteristic electronic data acquired by the data acquisition module includes the time-series encoding of vehicle historical maintenance records, the monthly cumulative mileage increment numerical sequence, and the slope parameter of the used car price index curve for the same model. The residual value decay prediction module has a built-in time-series feature encoder, which is a long short-term memory network computing unit. The long short-term memory network computing unit is used to convert dynamic time-series feature electronic data into a fixed-dimensional vehicle life cycle feature vector. The tensor data generated by the residual value decay prediction module, which characterizes the probability distribution of the predicted disposal residual value, includes mean and variance parameters. The risk constraint construction module executes a quantile extraction algorithm that calculates the risk value function value corresponding to the preset tail probability threshold based on the mean parameter and variance parameter, and determines the risk value function value as the stress residual value.

3. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules as described in claim 2, characterized in that, The input layer state vector dimension of the reinforcement learning pricing module includes the encoding dimension of static attribute electronic data, the feature dimension of vehicle life cycle feature vector, and the scalar dimension of pressure residual value. The action output layer of the second neural network model includes output nodes for the mortgage rate adjustment range and the mortgage ratio adjustment range.

4. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 3, characterized in that, The reinforcement learning pricing module has a built-in reward signal calculation submodule; In offline training mode, the reward signal calculation submodule is used to calculate the instantaneous reward value; When the ratio of the collateral ratio floating-point value output by the second neural network model to the electronic data of the actual residual value of the mortgaged vehicle obtained by the data acquisition module at the expiration of the mortgage period is greater than the preset bad debt threshold constant, the reward signal calculation submodule outputs a negative gradient back transmission signal, triggering the update process of the weight parameters of the second neural network model to suppress the power consumption of the processor performing invalid pricing calculations. When the ratio of the collateral ratio floating-point value output by the second neural network model to the actual residual value electronic data at the end of the collateral term is less than or equal to the preset bad debt threshold constant, and the collateral interest rate floating-point value meets the preset income threshold condition, the reward signal calculation submodule outputs a positive gradient feedback signal to solidify the weight parameters of the second neural network model.

5. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 1, characterized in that, It also includes a risk rule fusion module, which runs in the processor; The risk rule fusion module loads a set of expert risk hard boundary rules pre-stored in the memory. The set of expert risk hard boundary rules includes the maximum allowable vehicle age threshold variable, the minimum allowable mortgage residual value multiple threshold variable, and the upper and lower limit threshold variables of the mortgage interest rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. The risk rule fusion module calls the comparator logic circuit to perform the following judgments: compare the vehicle age value of the mortgaged vehicle with the maximum allowable vehicle age threshold variable, and compare the ratio of the residual value of the pressure to the vehicle purchase price field with the minimum allowable residual value multiple threshold variable. When the vehicle age value exceeds the maximum allowable vehicle age threshold, an interest rate adjustment is triggered. When the ratio of the residual value of the pressure to the vehicle purchase price field is lower than the minimum allowable residual value multiple threshold variable, the upper limit boundary of the mortgage ratio dimension will be forcibly set to zero. After the reinforcement learning pricing module outputs the original floating-point value of the mortgage rate, the risk rule fusion module determines whether the original floating-point value of the mortgage rate conforms to the upper and lower limit threshold variables of the mortgage rate corresponding to the vehicle age threshold variable and the mortgage residual value multiple threshold variable. If it does not conform, the original floating-point value of the mortgage rate is truncated and corrected to generate a corrected floating-point value of the mortgage rate that meets the hard boundary conditions.

6. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 1, characterized in that, The residual value decay prediction module and the reinforcement learning pricing module are equipped with an asynchronous clock signal triggering mechanism; The weight file of the first neural network model in the residual decay prediction module is updated with parameters at the end of the first clock cycle. The input layer of the second neural network model in the reinforcement learning pricing module receives the new state vector dimension boundary generated by the updated weight file of the first neural network model after the first clock cycle ends, and performs policy iterative update calculation in the second clock cycle.

7. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 1, characterized in that, The data acquisition module is also equipped with a vehicle diagnostic system signal receiving port, which is used to acquire in real time the vehicle diagnostic system fault code data and the cumulative number of traffic violations of the mortgaged vehicle during the mortgage period; The risk constraint construction module obtains the first attenuation coefficient by querying a preset fault code attenuation coefficient mapping table based on the fault code data of the vehicle diagnostic system, and obtains the second attenuation coefficient by querying a preset violation number attenuation coefficient mapping table based on the cumulative number of traffic violations. The product of the first attenuation coefficient and the second attenuation coefficient is used as the dynamic attenuation coefficient. The pressure residual value stored in the second data cache is dynamically attenuated and corrected by multiplication. The corrected pressure residual value is then updated in real time to the address space of the second data cache of the reinforcement learning pricing module through atomic synchronization operation. This triggers the reinforcement learning pricing module to recalculate the floating-point values ​​of the mortgage rate and the mortgage ratio based on the corrected pressure residual value.

8. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 1, characterized in that, It also includes a module for generating a visual interactive interface for mortgage risk assessment; The module for generating a visual interactive interface for mortgage risk assessment is used to drive the display device to render a graphical user interface that includes a pressure residual value decay curve layer and a pricing confidence interval dashboard layer; The pressure residual value decay curve layer uses the time axis as the horizontal axis and the mean of the pressure residual value value and the probability distribution of the predicted disposal residual value as the vertical axis. It displays the safety boundary of vehicle value during the mortgage period in real time and responds to the user's drag operation of the vertex control on the time axis by sending a mortgage period change signal to the risk constraint construction module to trigger the recalculation of the pressure residual value value. The pricing confidence interval dashboard layer is used to respond to the floating-point value of the collateral rate output by the reinforcement learning pricing module. It marks the risk-return balance range of the current floating-point value of the collateral rate on the dashboard, and responds to the user's drag operation on the target interest rate pointer on the dashboard. It sends an interest rate adjustment signal to the reinforcement learning pricing module to trigger the recalculation of the floating-point value of the collateral ratio.

9. The vehicle dynamic pricing system based on the fusion of reinforcement learning and risk rules according to claim 1, characterized in that, The reinforcement learning pricing module includes an online inference working mode and an offline training working mode; When the reinforcement learning pricing module is in online inference mode, the processor only calls the weight parameter file of the second neural network model stored in memory to perform forward propagation calculation to generate the floating-point value of the mortgage rate. It does not perform backpropagation gradient update calculation. It obtains the asymmetric action exploration constraint function parameters corresponding to the current pressure residual value by reading the mapping table data pre-calculated based on the preset risk confidence threshold. Based on the asymmetric action exploration constraint function parameters, it determines the lower boundary of the mortgage rate dimension and the upper boundary of the mortgage ratio dimension.

10. A vehicle dynamic pricing platform based on the fusion of reinforcement learning and risk rules, characterized in that, include: The mortgage application processing unit is used to parse the received mortgage request data packet containing the vehicle identification number field and the preset mortgage term field; The risk assessment interface unit is used to initiate structured query language requests to external vehicle credit databases to obtain electronic data of vehicle static attributes, electronic data of dynamic time-series characteristics, vehicle mortgage status records, and accident records. The pricing engine unit has a processor loaded with computer-executable instructions stored in memory for a vehicle dynamic pricing system based on reinforcement learning and risk rule fusion as described in any one of claims 1 to 9, used to generate floating-point values ​​for the mortgage rate and the mortgage ratio. The result feedback unit is used to encapsulate the floating-point values ​​of the collateral interest rate and the collateral ratio generated by the pricing engine unit into a Hypertext Transfer Protocol response message and send it to the front-end interactive device.

Images