Non-inductive payment system and method based on adaptive multi-modal fusion and behavior prediction

Abstract

The invention discloses a non-inductive payment system and method based on adaptive multi-modal fusion and behavior prediction, and the system comprises a non-inductive payment system, a user module, and a non-inductive payment machine body, and the non-inductive payment system is installed in the non-inductive payment machine body. The non-inductive payment system comprises a user module, an edge computing module, a cloud computing module, a block chain network module and an inter-module communication module, a biological feature collection device and an environment feature collection device are arranged in the user module, the self-adaptive multi-modal fusion algorithm is different from a fixed weight method, the weight is dynamically adjusted, and the recognition accuracy is improved; a behavior prediction model of reinforcement learning predicts customer behaviors in real time, and shopping experience is improved; the edge of the distributed transaction processing mechanism is coordinated with the cloud, the efficiency is improved, and high concurrency is adapted; and data security policy homomorphic encryption is combined with the block chain, so that data security is ensured, and laws and regulations are met.

Description

Technical Field

[0001] The present invention relates to the fields of payment technology and artificial intelligence technology, and specifically to a contactless payment system and method based on adaptive multimodal fusion and behavior prediction. Background Art

[0002] The contactless payment system provides a variety of charging modes, including user payment, manual charging, automatic payment by equipment and automatic deduction, and can provide discounts or points to car owners. This diversified charging mode is more humane. The contactless payment system can directly count the number of parking spaces in the parking lot, avoiding the inconvenience of manually searching for parking spaces. It is very convenient for parking lot managers and can realize unmanned management mode. The system charges by intelligently identifying parking time, avoiding disputes over parking fees caused by parking time, and can realize accurate and efficient charging mode. The payment process of the contactless payment system is very simple. Car owners do not need to install any APP. They only need to Bind the license plate number and vehicle model in Fubao or WeChat, and activate password-free payment to complete automatic deductions. The contactless payment system can automatically calculate taxes and invoices, reducing the time and cost of manual operations and repeated calculations, avoiding human errors and mistakes, and improving work efficiency. The contactless payment system uses encryption technology to ensure the safe storage of corporate data and avoid security risks of data storage. The system has an intelligent search function, and companies can search according to information such as invoice number or invoice type, which is very convenient. The contactless payment system can convert industry platform traffic into merchant's own traffic, realize refined private domain operations, and the cost of traffic acquisition is lower, more accurate, and more effective.

[0003] However, traditional payment methods have the following disadvantages:

[0004] Traditional contactless payment systems have problems with accuracy and security in identity authentication and shopping behavior recognition. Single-modality biometrics such as facial recognition are easily affected by environmental factors, and the tracking of shopping behavior lacks intelligence, making it difficult to meet actual needs. Summary of the Invention

[0005] The purpose of the present invention is to provide a contactless payment system and method based on adaptive multimodal fusion and behavior prediction, so as to solve the problems of insufficient accuracy and security of traditional contactless payment systems in identity authentication and shopping behavior recognition proposed in the above background technology, single-modal biometric recognition such as face recognition is easily affected by environmental factors, and the tracking of shopping behavior lacks intelligence, making it difficult to meet actual needs.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a contactless payment system based on adaptive multimodal fusion and behavior prediction, comprising a contactless payment system, a user module and a contactless payment machine body, wherein the contactless payment system is installed in the contactless payment machine body, the contactless payment system comprises a user module, an edge computing module, a cloud computing module, a blockchain network module and an inter-module communication module, the user module is provided with a biometric feature collection device and an environmental feature collection device, the biometric feature collection device comprises a camera, an iris collector, a voiceprint collector and a display, the camera, iris collector, voiceprint collector and display are installed on the top of the contactless payment machine body, the edge computing module is provided with a multimodal fusion identity authentication module, a distributed transaction processor and a server, the cloud computing module is provided with a behavior prediction module, a complex computing unit, a model training unit and a data storage unit, the blockchain network module is provided with a homomorphic encryption unit, an encryption computing unit and a regional chain recording unit, and the inter-module communication module is provided with a data transmission protocol, an interface design and a message queue.

[0007] As a preferred technical solution of the present invention, the multimodal fusion identity authentication module is provided with a data preprocessing unit, an environmental parameter acquisition unit, a dynamic weight calculation unit, a feature fusion unit and an identity authentication unit. The data preprocessing unit is electrically connected to the camera, iris collector and voiceprint collector, and the data preprocessing unit performs feature extraction on the face, iris and voiceprint respectively.

[0008] As a preferred technical solution of the present invention, the distributed transaction processor is provided with a transaction request classification unit, a task scheduling strategy unit and a transaction processing unit.

[0009] As a preferred technical solution of the present invention, the behavior prediction module uses the deep Q network algorithm in reinforcement learning to predict customer behavior in real time.

[0010] As a preferred technical solution of the present invention, the homomorphic encryption unit adopts the Paillier homomorphic encryption algorithm public key (n, g), private key λ, μ, the regional chain recording unit smart contract design defines transaction verification and data access rights, the consensus mechanism adopts the proof of stake (PoS) algorithm to improve efficiency, and the transaction data storage transaction hash value and encrypted data are stored on the blockchain.

[0011] As a preferred technical solution of the present invention, the data transmission protocol uses the HTTPS protocol encrypted by TLS, the interface design adopts RESTful API to define the request and response format, and the message queue uses RabbitMQ or Kafka to support high-concurrency data transmission.

[0012] The present invention provides a method for using a contactless payment system based on adaptive multimodal fusion and behavior prediction, comprising the following steps:

[0013] Step 1: Adaptive multimodal fusion algorithm: Dynamically adjust the weights of biometric modalities based on the environment and user characteristics to improve the accuracy and robustness of identity verification;

[0014] Step 2: Behavior prediction model based on reinforcement learning: Utilize reinforcement learning algorithms to predict customers’ shopping behavior in real time and automatically generate shopping lists;

[0015] Step 3: Efficient distributed transaction processing mechanism: Combine edge computing and cloud computing to improve payment processing efficiency, meet high concurrency requirements, and define MDP elements;

[0016] Step 4. Data security and privacy protection strategy: Use homomorphic encryption and blockchain technology to ensure the security and privacy of user data.

[0017] In the step 1, the feature vector F is obtained by extracting the feature of the biometric modality. f , F i , F v , normalize the feature vector to eliminate scale differences, collect the ambient light intensity L, noise level N, and distance D between the user and the device through the environmental parameter acquisition unit, and define the weight function through the dynamic weight calculation unit:

[0018]

[0019] βf, βi, βv are adjustment parameters,

[0020] L0, D0 are the optimal lighting and distance,

[0021] is the normalization coefficient;

[0022] The weighted fusion calculation formula of the feature fusion unit is:

[0023] F=w f ·F f +w i ·F i +w v ·F v

[0024] The identity authentication unit uses a deep neural network (DNN) as a classifier through classifier design, inputs the fusion feature vector F, outputs the identity matching probability, and uses the cross entropy loss function. The calculation formula is:

[0025] L=-[y·log(p)+(1-y)·log(1-p)]

[0026] y is the true label, p is the predicted probability,

[0027] Model training uses gradient descent method to optimize network parameters.

[0028] The transaction request classification unit defines the priority P based on the transaction amount m and the urgency e. The calculation formula is:

[0029] P=λ1m+λ2e

[0030] λ1, λ2 are weight coefficients,

[0031] The task scheduling strategy unit controls the edge server to process high-priority tasks and the cloud server to process low-priority tasks. When the edge server load Le is less than the threshold Lmax, it processes transaction requests with priority P>P0. Otherwise, it forwards the transaction requests to the cloud server.

[0032] The transaction processing unit receives the transaction request, calculates the priority P, selects the processing node, executes the transaction to complete the payment deduction, and returns the transaction result to the user.

[0033] In step 3, the reinforcement learning algorithm first defines the Markov decision process MDP elements:

[0034] State space S: includes customer position sp, gesture sh, and historical shopping records sr,

[0035] Action space A: possible shopping behaviors, such as taking a product (apick), putting a product (place), and moving (amove).

[0036] Reward function R(s,a): When taking a product, if the prediction is correct, the reward r>0; if the prediction is wrong, the reward r<0,

[0037] Deep Q network construction, the Q value function calculation formula is:

[0038] Q(s;a;θ)=f(s;a;θ)

[0039] f is a deep neural network, θ is a parameter,

[0040] Training process: Experience replay stores the buffer of past experience (s, a, r, s'). The loss function is calculated as:

[0041] L(θ(=E (s,a,r,s') [(yQ(s,a;θ)) 2 ]

[0042] The target value calculation formula is: y = r + γmax a' Q(s',a';θ),

[0043] The gradient update calculation formula is:

[0044]

[0045] α is the learning rate,

[0046] Online prediction, under the current state s, the optimal action calculation formula is:

[0047]

[0048] According to action a*, update the shopping list.

[0049] In step 4, the encryption calculation formula of homomorphic encryption and blockchain technology is:

[0050] E(m)=g m ·r n mod n 2

[0051] m is plain text, r is a random number,

[0052] The decryption calculation formula is:

[0053]

[0054]

[0055] The formula for calculating homomorphic properties is:

[0056] E(m1)·E(m2)=E(m1+m2 mod n)

[0057] The feature fusion is performed in the encrypted domain by the encryption calculation unit (52), and the calculation formula is:

[0058] E(F)=E(w f ·F f )·E(w i ·F i )·E(w v ·F v ).

[0059] Compared with the existing technology, the beneficial effects of the present invention are: the adaptive multimodal fusion algorithm is different from the fixed weight method, dynamically adjusts the weight, and improves the recognition accuracy; the behavior prediction model of reinforcement learning predicts customer behavior in real time and improves the shopping experience; the distributed transaction processing mechanism collaborates between the edge and the cloud to improve efficiency and adapt to high concurrency; the data security strategy combines homomorphic encryption and blockchain to ensure data security and comply with regulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1A side view of the contactless payment system of the present invention;

[0061] Figure 2 This is a schematic diagram of the structure of the contactless payment system of the present invention;

[0062] Figure 3 Schematic diagram of the user module architecture of the present invention;

[0063] Figure 4 This is a schematic diagram of the architecture of the edge computing module of the present invention;

[0064] Figure 5 This is a schematic diagram of the architecture of the multimodal fusion identity authentication module of the present invention;

[0065] Figure 6 A schematic diagram of the architecture of a distributed transaction processor according to the present invention;

[0066] Figure 7 This is a schematic diagram of the architecture of the cloud computing module of the present invention;

[0067] Figure 8 This is a schematic diagram of the architecture of the blockchain network module of the present invention;

[0068] Figure 9 Schematic diagram of the architecture of the inter-module communication module of the present invention;

[0069] Figure 10 Flowchart of the present invention.

[0070] In the figure: 1. Contactless payment system; 2. User module; 21. Biometric feature collection device; 22. Environmental feature collection device; 3. Edge computing module; 31. Multimodal fusion identity authentication module; 311. Data preprocessing unit; 312. Environmental parameter acquisition unit; 313. Dynamic weight calculation unit; 314. Feature fusion unit; 315. Identity authentication unit; 32. Distributed transaction processor; 321. Transaction request classification unit; 322. Task scheduling strategy unit; 323. Transaction processing unit; 33. Server; 4. Cloud computing module; 41. Behavior prediction module; 42. Complex computing unit; 43. Model training unit; 44. Data storage unit; 5. Blockchain network module; 51. Homomorphic encryption unit; 52. Encryption computing unit; 53. Regional chain recording unit; 6. Inter-module communication module; 61. Data transmission protocol; 62. Interface design; 63. Message queue. DETAILED DESCRIPTION

[0071] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0072] See also Figure 1-10 The present invention provides a non-sensing payment system based on adaptive multimodal fusion and behavior prediction, including a non-sensing payment system 1, a user module 2 and a non-sensing payment machine body 7. The non-sensing payment system 1 is installed in the non-sensing payment machine body 7. The non-sensing payment system 1 includes a user module 2, an edge computing module 3, a cloud computing module 4, a blockchain network module 5 and an inter-module communication module 6. The user module 2 is provided with a biometric feature collection device 21 and an environmental feature collection device 22. The biometric feature collection device 21 includes a camera 10, an iris collector 9, a voiceprint collector 11 and a display 8. The camera The camera 10, iris collector 9, voiceprint collector 11 and display 8 are installed on the top of the contactless payment machine body 7. The edge computing module 3 is equipped with a multimodal fusion identity authentication module 31, a distributed transaction processor 32 and a server 33. The cloud computing module 4 is equipped with a behavior prediction module 41, a complex computing unit 42, a model training unit 43 and a data storage unit. The blockchain network module 5 is equipped with a homomorphic encryption unit 51, an encryption computing unit 52 and a regional chain recording unit 53. The inter-module communication module 6 is equipped with a data transmission protocol 61, an interface design 62 and a message queue 63.

[0073] The multimodal fusion identity authentication module 31 is provided with a data preprocessing unit 311, an environmental parameter acquisition unit 312, a dynamic weight calculation unit 313, a feature fusion unit 314 and an identity authentication unit 315. The data preprocessing unit 311 is electrically connected to the camera 10, the iris collector 9 and the voiceprint collector 11. The data preprocessing unit 311 extracts features of the face, iris and voiceprint respectively.

[0074] The distributed transaction processor 32 includes a transaction request classification unit 321 , a task scheduling strategy unit 322 and a transaction processing unit 323 .

[0075] The behavior prediction module 41 uses the deep Q network algorithm in reinforcement learning to predict customer behavior in real time.

[0076] The homomorphic encryption unit 51 adopts the Paillier homomorphic encryption algorithm public key (n, g), private key λ, μ, and the regional chain recording unit 53 smart contract design defines transaction verification and data access rights. The consensus mechanism adopts the Proof of Stake PoS algorithm to improve efficiency. The transaction data storage transaction hash value and encrypted data are stored on the blockchain.

[0077] The data transmission protocol 61 uses the HTTPS protocol encrypted by TLS, the interface design 62 adopts RESTful API, defines the request and response format, and the message queue 63 uses RabbitMQ or Kafka to support high-concurrency data transmission.

[0078] The present invention provides a method for using a contactless payment system based on adaptive multimodal fusion and behavior prediction, comprising the following steps:

[0079] Step 1: Adaptive multimodal fusion algorithm: Dynamically adjust the weights of biometric modalities based on the environment and user characteristics to improve the accuracy and robustness of identity verification;

[0080] Step 2: Behavior prediction model based on reinforcement learning: Utilize reinforcement learning algorithms to predict customers’ shopping behavior in real time and automatically generate shopping lists;

[0081] Step 3: Efficient distributed transaction processing mechanism: Combining edge computing and cloud computing to improve payment processing efficiency and meet high concurrency requirements;

[0082] Step 4. Data security and privacy protection strategy: Use homomorphic encryption and blockchain technology to ensure the security and privacy of user data.

[0083] In step 1, the feature vector F is obtained by extracting the feature of the biometric modality. f , F i , F v , normalize the feature vector to eliminate scale differences, collect the ambient light intensity L, noise level N, and distance D between the user and the device through the environmental parameter acquisition unit 312, and define the weight function through the dynamic weight calculation unit 313:

[0084]

[0085] βf, βi, βv are adjustment parameters,

[0086] L0, D0 are the optimal lighting and distance,

[0087] is the normalization coefficient;

[0088] The weighted fusion calculation formula of the feature fusion unit 314 is:

[0089] F=w f ·F f +w i ·F i +w v ·F v

[0090] The identity authentication unit 315 adopts a deep neural network (DNN) as a classifier through classifier design, inputs the fusion feature vector F, outputs the identity matching probability, and uses the cross entropy loss function as the loss function, and the calculation formula is:

[0091] L=-[y·log(p)+(1-y)·log(1-p)]

[0092] y is the true label, p is the predicted probability,

[0093] Model training uses gradient descent method to optimize network parameters.

[0094] The transaction request classification unit 321 defines the priority P according to the transaction amount m and the urgency e. The calculation formula is:

[0095] P=λ1m+λ2e

[0096] λ1, λ2 are weight coefficients,

[0097] The task scheduling strategy unit 322 controls the edge server to process high-priority tasks and the cloud server to process low-priority tasks. When the edge server load Le is less than the threshold Lmax, the transaction request with priority P>P0 is processed. Otherwise, the transaction request is forwarded to the cloud server.

[0098] The transaction processing unit 323 receives the transaction request, calculates the priority P, selects the processing node, executes the transaction to complete the payment deduction, and returns the transaction result to the user.

[0099] In step 3, the reinforcement learning algorithm first defines the Markov decision process MDP elements:

[0100] State space S: includes customer position sp, gesture sh, and historical shopping records sr,

[0101] Action space A: possible shopping behaviors, such as taking a product (apick), putting a product (place), and moving (amove).

[0102] Reward function R(s,a): When taking a product, if the prediction is correct, the reward r>0; if the prediction is wrong, the reward r<0,

[0103] Deep Q network construction, the Q value function calculation formula is:

[0104] Q(s;a;θ)=f(s;a;θ)

[0105] f is a deep neural network, θ is a parameter,

[0106] Training process: Experience replay stores the buffer of past experience (s, a, r, s'). The loss function is calculated as:

[0107] L(θ(=E (s,a,r,s') [(yQ(s,a;θ)) 2 ]

[0108] The target value calculation formula is: y = r + γmax a' Q(s',a';θ),

[0109] The gradient update calculation formula is:

[0110]

[0111] α is the learning rate,

[0112] Online prediction, under the current state s, the optimal action calculation formula is:

[0113]

[0114] According to action a*, update the shopping list.

[0115] In step 4, the encryption calculation formula of homomorphic encryption and blockchain technology is:

[0116] E(m)=g m ·r n mod n 2

[0117] m is plain text, r is a random number,

[0118] The decryption calculation formula is:

[0119]

[0120]

[0121] The formula for calculating homomorphic properties is:

[0122] E(m1)·E(m2)=E(m1+m2 mod n)

[0123] The feature fusion is performed in the encrypted domain by the encryption calculation unit (52), and the calculation formula is:

[0124] E(F)=E(w f ·F f )·E(w i ·F i )·E(w v ·F v ).

[0125] Example 1:

[0126] In the present invention, data collection: the user enters the recognition area, collects Ff , F i , F v ; Environmental parameter acquisition; Get L, N, D; Dynamic weight calculation: Calculate w f ,w i ,w v ; Feature fusion: Calculate the fused feature F; Authentication: Input F to DNN and output the matching probability; Decision: If the match is successful, allow entry; otherwise deny.

[0127] Example 2:

[0128] In this invention, shopping behavior prediction, state acquisition: obtain the current state s, action selection, through the DQN model, select action a * ; List updated according to a * , update the shopping list; model update: based on actual feedback, update the experience pool and train the model.

[0129] Example 3:

[0130] In the present invention, payment processing is performed, a transaction request is generated, and when the user leaves the store, the system automatically generates a transaction request; task scheduling selects edge or cloud server processing based on priority; transaction execution completes the payment and records the transaction hash; result feedback notifies the user that the payment is successful.

[0131] Example 4:

[0132] In the present invention, the process of the adaptive multimodal fusion identity authentication module is specifically as follows: biometric data collection: when the user enters the recognition area, the system simultaneously collects the face image, iris image and voiceprint signal; environmental parameter acquisition: the sensor measures the ambient light intensity L, noise level N, and user distance D; feature extraction and preprocessing: feature extraction of face, iris and voiceprint data to obtain the feature vector F f , F i , F v ; Dynamic weight calculation: According to the environmental parameters, the weight function is used to calculate the weight w of each mode f ,w i ,w v ; Feature fusion: The weighted feature vectors are fused to obtain the comprehensive feature F; Identity verification; The fused feature F is input into the deep neural network model and the identity matching probability is output; Result judgment: Based on the matching probability, it is determined whether the user identity is successfully verified and whether entry is authorized or denied.

[0133] Example 5:

[0134] In the present invention, the process of the behavior prediction module based on reinforcement learning is as follows: state information acquisition: the system obtains the customer's location sp, gesture sh, historical shopping record sr, and state information in real time; action space definition: defines possible customer behavior actions, such as taking goods, putting goods, and moving positions; deep Q network (DQN) model: inputs the current state s and outputs the Q value Q(s, a) of each action; action selection strategy: adopts a greedy strategy to select the action a* with the maximum Q value, or selects random actions with probability to balance exploration and utilization; shopping list update: based on the predicted action a* and actual behavior, the customer's shopping list is updated in real time; model training and update: the actual state transition (s, a, r, s') is stored in the experience replay pool, and the DQN model is trained by sampling regularly.

[0135] Example 6:

[0136] In this invention, the architecture of the distributed transaction processing mechanism is as follows: transaction request generation module When a customer finishes shopping and is ready to leave the store, the system automatically generates a transaction request, including a shopping list and payment information; task scheduling module; priority calculation based on the transaction amount m and the urgency e to calculate the priority P; edge and cloud server selection, based on the load L of the edge server e and threshold L max ,determines whether the transaction request is processed by the edge server or the cloud server,Transaction execution module,Edge server: processes high-priority transaction requests and completes payments quickly; cloud server processes low-priority transaction requests to ensure system stability,Result feedback module, returns the transaction results to the user and updates the transaction records in the blockchain network.

[0137] Example 7:

[0138] In the present invention, homomorphic encryption processing: data encryption is performed at the biometric data collection end, and the feature vector F f , F i , F v Perform homomorphic encryption to obtain F f , F i , F v :Encrypted domain computing**: Complete feature fusion F in the encrypted state; Authentication and transaction processing: Complete authentication and transaction processing in the encrypted domain to ensure that the original data is not leaked; Blockchain records; Smart contracts define data access rights and transaction verification rules; Transactions are stored on the chain to store transaction hashes and encrypted data in the blockchain to ensure that the data cannot be tampered with; Data decryption (when necessary) can only be decrypted by the private key when it meets the authority to ensure data security.

[0139] Example 8:

[0140] In the present invention, the weight function w is shown f,w i ,w v Curve varying with environmental parameters. The horizontal axis represents the change in environmental parameters such as light intensity L, noise level N, and distance D. The vertical axis represents the weight value of the corresponding modality. When the light intensity decreases, the weight w of face recognition f decreases; when the noise level increases, the weight w of voiceprint recognition v decreases. The system dynamically adjusts the weights to ensure recognition accuracy.

[0141] Example 9:

[0142] In the present invention, the Q-value update of the reinforcement learning model is shown, demonstrating the training process and Q-value update mechanism of the deep Q-network (DQN); the network structure diagram includes an input layer (state s), a hidden layer, and an output layer (Q-value of action a); the experience replay pool shows the storage and sampling process of experience samples; the loss function curve: shows the change trend of the loss function L θ during the training process, indicating the convergence of the model.

[0143] Example 10:

[0144] In the present invention, the task scheduling process for transaction processing: receiving a transaction request, the system receives a transaction request, including the transaction amount and urgency; calculating the priority: calculating the transaction priority according to the formula P = λ1m + λ2e; monitoring the load of the edge server: obtaining the current load Le of the edge server; selecting a processing node: if Le < Lmax and P > P0, then process by the edge server; otherwise, forward it to the cloud server for processing; transaction execution and result feedback: after completing the transaction, return the result to the user and record the transaction information.

[0145] Example 11:

[0146] In the present invention, the homomorphic encryption calculation process: data encryption, showing the process of immediately encrypting biometric data after collection; calculation in the encrypted domain: completing feature fusion and identity verification calculations in the encrypted state; data transmission and storage: the encrypted data is transmitted through a secure channel and stored in a server or blockchain; data decryption: decrypting the data with a private key when necessary.

[0147] Example 12:

[0148] In the present invention, the blockchain network specifically refers to a node distribution display consisting of multiple nodes, including edge nodes and cloud nodes; the data blockchain displays the chain structure of the blockchain, with each block containing transaction data hash, timestamp, and hash information of the previous block; the consensus mechanism indicates the adopted Proof of Stake (PoS) consensus algorithm, with each node working together to maintain data consistency; and the smart contract execution displays the deployment and execution process of the smart contract in the blockchain network, automatically handling transaction verification and authority control.

[0149] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A non-contact payment system based on adaptive multi-modal fusion and behavior prediction, comprising a non-contact payment system (1), a user module (2) and a non-contact payment machine body (7), characterized in that: The contactless payment system (1) is installed in a contactless payment machine body (7). The contactless payment system (1) comprises a user module (2), an edge computing module (3), a cloud computing module (4), a blockchain network module (5) and an inter-module communication module (6). The user module (2) is provided with a biometric feature collection device (21) and an environmental feature collection device (22). The biometric feature collection device (21) comprises a camera (10), an iris collector (9), a voiceprint collector (11) and a display (8). The camera (10), the iris collector (9), the voiceprint collector (11) and the display (8) are installed in the user module (2). Installed on the top of the body (7) of the contactless payment machine, the edge computing module (3) is provided with a multimodal fusion identity authentication module (31), a distributed transaction processor (32) and a server (33), the cloud computing module (4) is provided with a behavior prediction module (41), a complex computing unit (42), a model training unit (43) and a data storage unit, the blockchain network module (5) is provided with a homomorphic encryption unit (51), an encryption computing unit (52) and a regional chain recording unit (53), and the inter-module communication module (6) is provided with a data transmission protocol (61), an interface design (62) and a message queue (63).

2. The contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 1, characterized in that: The multimodal fusion identity authentication module (31) is provided with a data preprocessing unit (311), an environmental parameter acquisition unit (312), a dynamic weight calculation unit (313), a feature fusion unit (314) and an identity authentication unit (315); the data preprocessing unit (311) is electrically connected to a camera (10), an iris collector (9) and a voiceprint collector (11); and the data preprocessing unit (311) extracts features of a face, an iris and a voiceprint respectively.

3. The contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 1, characterized in that: The distributed transaction processor (32) is provided with a transaction request classification unit (321), a task scheduling strategy unit (322) and a transaction processing unit (323).

4. The contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 1, characterized in that: The behavior prediction module (41) uses the deep Q network algorithm in reinforcement learning to predict customer behavior in real time.

5. The contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 1, characterized in that: The homomorphic encryption unit (51) adopts the Paillier homomorphic encryption algorithm public key (n, g), private key λ, μ, the regional chain recording unit (53) smart contract design defines transaction verification and data access rights, and the consensus mechanism adopts the proof of stake (PoS) algorithm to improve efficiency. The transaction data storage transaction hash value and encrypted data are stored on the blockchain.

6. The contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 1, characterized in that: The data transmission protocol (61) uses the HTTPS protocol encrypted by TLS, the interface design (62) adopts the RESTful API to define the request and response formats, and the message queue (63) uses RabbitMQ or Kafka to support high-concurrency data transmission.

7. A method for using a contactless payment system based on adaptive multimodal fusion and behavior prediction according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step 1: Adaptive multimodal fusion algorithm: Dynamically adjust the weights of biometric modalities according to the environment and user characteristics to improve the accuracy and robustness of identity authentication; Step 2: Behavior prediction model based on reinforcement learning: Use reinforcement learning algorithm to predict customers' shopping behavior in real time and automatically generate shopping lists; Step 3: Efficient distributed transaction processing mechanism: Combine edge computing and cloud computing to improve payment processing efficiency and meet high concurrency requirements; Step 4. Data security and privacy protection strategy: Use homomorphic encryption and blockchain technology to ensure the security and privacy of user data.

8. The method for using the non-contact payment system based on adaptive multi-modal fusion and behavior prediction according to claim 7 is characterized by: In step 1, feature extraction is performed on the biometric recognition modality to obtain a feature vector F f , F i , F v , normalize the feature vector to eliminate scale differences, collect the parameters of ambient light intensity L, noise level N, and distance D between the user and the device through the environmental parameter acquisition unit (312), and define the weight function through the dynamic weight calculation unit (313): Βf, βi, βv are adjustment parameters, L0, D0 are the best lighting and distance, is the normalization coefficient; The weighted fusion calculation formula of the feature fusion unit (314) is: F=w f ·F f +w i ·F i +w v ·F v The identity authentication unit (315) adopts a deep neural network (DNN) as a classifier through classifier design, inputs a fusion feature vector F, outputs an identity matching probability, and uses a cross entropy loss function as a loss function, and the calculation formula is: L=-[y·log(p)+(1-y)·log(1-p)] y is the true label, p is the predicted probability, Model training uses gradient descent method to optimize network parameters. The transaction request classification unit (321) defines the priority level P according to the transaction amount m and the urgency level e. The calculation formula is: P=λ1m+λ2e λ1, λ2 are weight coefficients, The task scheduling strategy unit (322) controls the edge server to process high priority tasks and the cloud server to process low priority tasks. When the edge server load Le is less than the threshold value Lmax, the transaction request with priority P>P0 is processed. Otherwise, the transaction request is forwarded to the cloud server. The transaction processing unit (323) receives the transaction request, calculates the priority P, selects the processing node, executes the transaction to complete the payment deduction, and returns the result to feed back the transaction result to the user.

9. The method for using the non-contact payment system based on adaptive multi-modal fusion and behavior prediction according to claim 7, characterized in that: In step 3, the reinforcement learning algorithm first defines the Markov decision process MDP element: State space S: includes customer position sp, gesture sh, and historical shopping record sr. Action space A: possible shopping behaviors, such as taking a product apick, putting a product aplace, and moving amove, Reward function R(s,a): When taking a product, if the prediction is correct, the reward r>0; if the prediction is wrong, the reward r<0, Deep Q network construction, the Q value function calculation formula is: Q(s;a;θ)=f(s;a;θ) f is a deep neural network, θ is a parameter, Training process: Experience replay stores the buffer of past experience (s, a, r, s'), and the loss function calculation formula is: L(θ(=E (s,a,r,s') [(yQ(s,a;θ)) 2 ] The target value calculation formula is: y = r + γγma a' Q(s',a'; θ), The gradient update calculation formula is: α is the learning rate, Online prediction, under the current state s, the optimal action calculation formula is: According to action a*, update the shopping list.

10. The method for using the contactless payment system based on adaptive multimodal fusion and behavior prediction according to claim 7, characterized in that: In step 4, the encryption calculation formula of homomorphic encryption and blockchain technology is: E(m)=g m ·r n mod n 2 m is plain text, r is a random number, The decryption calculation formula is: The formula for calculating homomorphic properties is: E(m1)·E(m2)=E(m1+m2 mod n) The feature fusion is performed in the encrypted domain by the encryption calculation unit (52), and the calculation formula is: E(F)=E(w f ·F f )·E(w i ·F i )·E(w v ·F v ).

Images