Automobile skylight AI global intelligent interaction system

Abstract

The invention relates to the technical field of automobile engineering, in particular to an automobile sunroof AI global intelligent interaction system which comprises a privacy calculation module, an intelligent interaction and edge calculation module, a data safety and authority management module and a self-adaptive control and execution module. Through fusion of privacy calculation, multi-modal intelligent interaction, full-link safety management and control and adaptive optimization control technologies, cross-vehicle data collaborative privacy protection and efficient modeling, real-time robust processing and accurate intention reasoning of multi-source heterogeneous interaction data, dynamic safety protection and fine-grained authority management of data in a full life cycle are realized; and intelligent sunroof control meeting multi-target constraints is realized, so that the safety, privacy, intelligence and reliability of automobile sunroof interaction are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automotive engineering, in particular to an AI global intelligent interaction system for automobile sunroof. BACKGROUND

[0002] The automobile sunroof refers to an openable window installed on the top of the automobile, which is usually composed of a glass panel, a sunshade curtain, a driving motor and a control mechanism, and is an important functional component of the automobile body. Its core function is to realize air circulation, lighting regulation and environmental interaction between the inside and outside of the vehicle through opening or closing actions. With the rapid development of automobile intelligence and networking, the traditional mechanical control sunroof has gradually evolved towards intelligent interaction, and can realize self-adaptive regulation combined with environmental sensor data, greatly improving the convenience of driving and riding. In the prior art, the automobile sunroof interaction has the problems of high privacy leakage risk in data collaboration process, lagging multi-modal interaction response and being easily disturbed, insufficient data security protection and extensive permission management, and control strategy difficult to balance comfort and safety.

[0003] Based on this, the present application provides an AI global intelligent interaction system for automobile sunroof to solve the above-mentioned technical problems. SUMMARY

[0004] The present application aims to provide an AI global intelligent interaction system for automobile sunroof, which integrates privacy computing, multi-modal intelligent interaction, full-link security control and adaptive optimization control technology, realizes privacy protection and efficient modeling of cross-vehicle data collaboration, real-time robust processing and accurate intention reasoning of multi-source heterogeneous interaction data, dynamic security protection and fine-grained permission management of data full life cycle, and intelligent sunroof control meeting multi-objective constraints, and comprehensively improves the safety, privacy, intelligence and reliability of automobile sunroof interaction.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: The present application provides an AI global intelligent interaction system for automobile sunroof, which includes a privacy computing module, an intelligent interaction and edge computing module, a data security and permission management module, and an adaptive control and execution module, wherein: The privacy computing module is used for efficient collaborative modeling and auditable computing of cross-vehicle intelligent interaction data under the premise of protecting user privacy and data security by integrating lightweight encryption, distributed collaborative learning and trusted execution mechanism; The intelligent interaction and edge computing module is used for dynamically weighting multi-modal features through a spatio-temporal attention fusion mechanism, and enhancing the robustness of the model through adversarial training, and real-time processing of voice instructions, ToF gesture recognition signals and environmental sensor data at the edge, and outputting high-reliability sunroof control intention; The data security and permission management module is used for constructing a unified security framework integrating privacy protection and dynamic permission control, and performing full-link security protection and fine-grained access management of data in a multi-party cooperation process through multi-level encryption, adaptive privacy enhancement and behavior-driven access control mechanism. The adaptive control and execution module is based on a proximal policy optimization algorithm, uses encrypted user behavior data provided by the privacy calculation module to train a reinforcement learning model, and generates a sunroof control strategy meeting multi-target constraints.

[0006] The privacy calculation module includes a lightweight encryption unit, a distributed collaborative learning unit and a trusted execution environment unit. The lightweight encryption unit uses a CKKS homomorphic encryption algorithm to perform low-delay encryption processing on sensitive data such as voice instructions and biological features. The distributed collaborative learning unit performs secure aggregation of cross-vehicle model parameters based on a FedAvg framework and supports differential privacy protection collaborative optimization. The trusted execution environment unit is used to isolate sensitive computing through a vehicle-level TEE chip and provides hardware-level data integrity and audit tracking.

[0007] The lightweight encryption unit uses a CKKS homomorphic encryption algorithm to perform low-delay encryption processing on sensitive data such as voice instructions and biological features, and the specific operations are as follows: A1: Vector encoding is performed on voice instructions or biological feature data to generate a plaintext vector ; A2: Based on a polynomial ring , the plaintext vector is encoded into a polynomial ; A3: A key generation module is used to generate a public key , a private key and an evaluation key ; A4: An encryption function is used to generate a ciphertext ; A5: A compression encoding strategy and low-order parameter configuration are used to control the encryption delay within 50ms; the low-order parameter configuration includes modulus , polynomial order and precision control parameter .

[0008] The intelligent interaction and edge computing module includes a multi-modal sensor fusion unit, a space-time attention mechanism unit, an adversarial training defense unit and an intent reasoning decision unit. The multi-modal sensor fusion unit is used for real-time synchronous processing of multi-source heterogeneous sensor data of voice, ToF gesture, ambient light / temperature and humidity. The space-time attention mechanism unit is used for dynamically weighting asynchronous multi-modal features through an STAM network to eliminate sensor timing deviation of ≤100 ms. The adversarial training defense unit is used for adopting a generative adversarial network architecture to enhance the robustness of the model to adversarial attacks. The intent reasoning decision unit maps multi-modal features to sunroof control intents such as "fully open" and "sunshade tilt" based on a lightweight BERT.

[0009] The space-time attention mechanism unit in the space-time attention mechanism unit is used for dynamically weighting asynchronous multi-modal features through an STAM network to eliminate sensor timing deviation of ≤100 ms, and the specific operation is as follows: B1: Time stamp alignment is performed on multi-modal sensor data from voice, ToF gesture, ambient light / temperature and humidity to construct asynchronous feature sequences . B2: A sliding time window mechanism is adopted to perform timing compensation on asynchronous features, and the time window length is . B3: The dynamic weight of each modal feature is calculated through a space-time attention module, and the weight calculation formula is: . Among them, , are the query and key vectors of the i-th modal, is a scaling factor; B4: The multi-modal features are weighted and fused according to the weight to output a fused feature vector .

[0010] The adversarial training defense unit in the adversarial training defense unit adopts a generative adversarial network architecture to enhance the robustness of the model to adversarial attacks, and the specific operation is as follows: C1: Adversarial attack sample generation: ① For voice instruction features, an FGSM algorithm is used to generate adversarial perturbations: . Among them, is the perturbation strength, L is the cross-entropy loss, and y is the true label; ② For ToF gesture trajectories, an affine transformation is used to generate adversarial samples: , . ③ A mixed adversarial sample set , ambient feature + random noise}; C2: Generative adversarial network training: ① Generator G: input clean samples , output simulated adversarial samples , using residual network architecture "3 layers of convolution + 2 layers of deconvolution"; ② Discriminator D: input sample S or or ", output classification probability P = D(S); ③ Joint loss function: ; where, is the standard loss of GAN, is the classification loss of the model on the adversarial sample, is the weight coefficient; C3: defense model optimization: ① Mix and in a ratio of 1:4 to train the intelligent interaction model f; ② Early stopping strategy is adopted, and when the accuracy of the adversarial sample on the validation set is ≥95%, the training is stopped; ③ Output robust model , ensuring that the defense success rate against the above adversarial attacks is ≥99%.

[0011] The data security and permission management module comprises a multi-level encryption protection unit, a self-adaptive privacy enhancement unit, a behavior-driven access control unit and a security situation awareness unit, wherein: The multi-level encryption protection unit is used for encrypting the full-link data of the transmission layer, the storage layer and the computing layer; The self-adaptive privacy enhancement unit is used for dynamically injecting differential privacy noise according to the data sensitivity to balance the model accuracy and privacy protection; The behavior-driven access control unit: based on ABAC policy engine, real-time analysis of user operation mode, realize millisecond level dynamic permission rise and fall; The security situation awareness unit is used for real-time monitoring of system security state and triggering response mechanism through abnormal detection algorithm and threat intelligence library.

[0012] The behavior-driven access control unit based on ABAC policy engine analyzes the user operation mode in real time, realizes millisecond level dynamic permission rise and fall, and the specific operation is as follows: D1: multi-dimensional attribute extraction: User attributes U = {identity ID, role level, biometric template}; Resource attribute R = {data type, sensitivity level, access history}; Environmental attribute E = {timestamp, geographic location, vehicle status}; Behavioral attributes A = {operation frequency, operation timing, operation context}; D2: Behavior pattern modeling: ① Construct user operation sequence n = the last 100 operations; ② Generate behavior feature vector through bidirectional LSTM network The calculation formula is: ; ③ Calculate the similarity score of the current operation with the historical pattern based on Gaussian mixture model: Where, is the GMM parameter, and K = 5 is the number of mixed components; D3: Dynamic permission decision: ① Define the permission function f: m is the permission dimension, specifically: Where, e is the environmental feature vector, is the weight matrix, is the Sigmoid activation function; ② When the similarity score and the vehicle speed > 80 km / h, automatically trigger the permission downgrade rule: I. Turn off the "full open" permission in the voice command; II. Limit the gesture operation range to sunshade adjustment; D4: Permission execution and audit: ① Permission decision response time ≤ 5ms; ② Generate an unalterable audit log L = {t, U, R, E, f(U, R, E, A), permission change record}; ③ Use blockchain technology to synchronize the log to the vehicle safety module to ensure data integrity.

[0013] The security posture awareness unit monitors the system security state in real time through the anomaly detection algorithm and threat intelligence library and triggers the response mechanism. The specific operation is as follows: E1: Real-time aggregation of security logs from CAN bus, vehicle firewall, and sensor interface, standardized into a unified time series format; E2: Composite analysis through the integration of the following detection engines: ① LSTM-based time series anomaly detection; ② loC matching based on threat intelligence library; ③ Vehicle-specific attack pattern recognition based on rule engine; E3: Calculate the comprehensive risk value using fuzzy logic algorithm, and divide the threat level combined with the vehicle state; E4: Triggering hierarchical response according to threat level: I, 1st level "risk value 30-60": record log and pre-warning; II, 2nd level "risk value 60-80": limit non-critical functions; III, 3rd level "> 80": cut off sunroof power and activate backup mechanical control.

[0014] The adaptive control and execution module comprises a PPO reinforcement learning engine unit, a multi-objective constraint optimization unit and a real-time control execution unit, wherein: The PPO reinforcement learning engine unit: based on the proximal policy optimization algorithm, training the sunroof control policy network in the encrypted state; The multi-objective constraint optimization unit: used for solving the Pareto optimal solution set of the indicators of comfort, safety and energy consumption by using NSGA-II algorithm, and dynamically balancing multiple optimization objectives; The real-time control execution unit: for realizing 10ms-level sunroof motor control through CAN FD bus, supporting emergency shutdown.

[0015] Compared with the prior art, the beneficial effects of the present application are: The present application realizes the privacy protection and efficient modeling of cross-vehicle data collaboration, real-time robust processing and accurate intention reasoning of multi-source heterogeneous interactive data, dynamic security protection and fine-grained permission management of data full life cycle, and intelligent sunroof control meeting multi-objective constraints by fusing privacy computing, multi-modal intelligent interaction, full-link security management and adaptive optimization control technology, and comprehensively improves the safety, privacy, intelligence and reliability of automobile sunroof interaction. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a system diagram of the automobile sunroof AI global intelligent interaction system of the present application.

[0017] Figure 2 It is a multi-modal fusion flowchart in the automobile sunroof AI global intelligent interaction system of the present application.

[0018] Figure 3 It is a safety control closed-loop flowchart in the automobile sunroof AI global intelligent interaction system of the present application.

[0019] EXPLANATION OF REFERENCE NUMBERS: 1. Privacy Computing Module; 11. Lightweight Encryption Unit; 12. Distributed Collaborative Learning Unit; 13. Trusted Execution Environment Unit; 2. Intelligent Interaction and Edge Computing Module; 21. Multimodal Sensor Fusion Unit; 22. Spatiotemporal Attention Mechanism Unit; 23. Adversarial Training Defense Unit; 24. Intent Reasoning and Decision-Making Unit; 3. Data Security and Access Management Module; 31. Multi-Level Encryption Protection Unit; 32. Adaptive Privacy Enhancement Unit; 33. Behavior-Driven Access Control Unit; 34. Security Situation Awareness Unit; 4. Adaptive Control and Execution Module; 41. PPO Reinforcement Learning Engine Unit; 42. Multi-Objective Constraint Optimization Unit; 43. Real-Time Control Execution Unit. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: like Figures 1-3 As shown, this embodiment provides an AI-powered intelligent interaction system for automotive sunroofs, including a privacy computing module 1, an intelligent interaction and edge computing module 2, a data security and access control module 3, and an adaptive control and execution module 4. Specifically: the privacy computing module 1 integrates lightweight encryption, distributed collaborative learning, and trusted execution mechanisms to perform efficient collaborative modeling and auditable computation of cross-vehicle intelligent interaction data while ensuring user privacy and data security; the intelligent interaction and edge computing module 2 dynamically weights multimodal features through a spatiotemporal attention fusion mechanism and enhances model robustness through adversarial training, processing voice commands, ToF gesture recognition signals, and environmental sensor data in real time at the edge to output highly reliable sunroof control intentions; the data security and access control module 3 constructs a unified security framework integrating privacy protection and dynamic access control, using multi-level encryption, adaptive privacy enhancement, and behavior-driven access control mechanisms to ensure end-to-end security and fine-grained access management of data during multi-party collaboration; and the adaptive control and execution module 4, based on a near-end policy optimization algorithm, uses encrypted user behavior data provided by the privacy computing module 1 to train a reinforcement learning model and generate a sunroof control strategy that satisfies multi-objective constraints.

[0022] It should be noted that the privacy computing module 1 achieves data security collaboration through lightweight encryption and federated learning, providing a privacy-protected multimodal data processing foundation for the intelligent interaction and edge computing module 2. The highly reliable control intent output by the intelligent interaction and edge computing module 2 is dynamically controlled and threatened by the data security and permission management module 3, and then the adaptive control and execution module 4 generates an optimized strategy that meets security constraints. Finally, it is securely executed through the end-to-end encryption mechanism of the data security and permission management module 3.

[0023] In this embodiment, it should also be noted that the privacy computing module 1 includes a lightweight encryption unit 11, a distributed collaborative learning unit 12, and a trusted execution environment unit 13, wherein: the lightweight encryption unit 11: uses the CKKS homomorphic encryption algorithm to perform low-latency encryption processing on sensitive data such as voice commands and biometric features; the specific operation is as follows: A1: performs vector encoding on the voice commands or biometric feature data to generate plaintext vectors. A2: Based on polynomial rings , plaintext vector Encoding as a polynomial A3: Use the key generation module to generate a public key. Private key and evaluation key A4: Through encryption function Generate ciphertext A5: Employs compression coding strategies and low-order parameter configurations to keep encryption latency below 50ms; low-order parameter configurations include: modulus polynomial order Precision control parameters Distributed collaborative learning unit 12: Based on the FedAvg framework, it performs secure aggregation of cross-vehicle model parameters and supports differential privacy-preserving collaborative optimization; Trusted execution environment unit 13: It is used to isolate sensitive computations through automotive-grade TEE chips and provide hardware-level data integrity and audit trail.

[0024] It should be noted that the lightweight encryption unit 11 provides real-time encryption protection for sensitive data, providing a secure data foundation for the distributed collaborative learning unit 12. Meanwhile, the trusted execution environment unit 13 provides hardware-level secure computing isolation and auditing support for the former two.

[0025] Furthermore, it's important to explain the workflow of the FedAvg framework: First, the aggregation server sends global model parameters to each vehicle node; then, each vehicle node trains its model using local data to obtain local model parameters; next, each vehicle node uploads its local model parameters to the aggregation server; finally, the aggregation server aggregates the model parameters of all vehicles using a weighted average method to obtain new global model parameters. The automotive-grade TEE chip uses ARM TrustZone technology, such as the NXP i.MX8M Plus. ARM TrustZone technology is a hardware-level security technology that achieves hardware-level security isolation by dividing the device's memory and processor resources into a secure world and a normal world.

[0026] In this embodiment, it should also be noted that the intelligent interaction and edge computing module 2 includes a multimodal sensor fusion unit 21, a spatiotemporal attention mechanism unit 22, an adversarial training defense unit 23, and an intent reasoning decision unit 24, wherein: the multimodal sensor fusion unit 21 is used to process multi-source heterogeneous sensor data of voice, ToF gesture, and ambient light / temperature and humidity in real time; the spatiotemporal attention mechanism unit 22 is used to dynamically weight asynchronous multimodal features through a STAM network to eliminate sensor timing deviations of ≤100ms; the specific operation is as follows: B1: timestamp alignment is performed on the multimodal sensor data from voice, ToF gesture, and ambient light / temperature and humidity to construct an asynchronous feature sequence. B2: A sliding time window mechanism is used to perform timing compensation for asynchronous features, with a time window length of [missing information]. B3: The dynamic weights of each modality feature are calculated using the spatiotemporal attention module. The weight calculation formula is as follows: ; in, , Let be the query and key vectors for the i-th modality, respectively. B4: Scaling factor; B5: Weighted fusion of multimodal features based on weights, outputting a fused feature vector. Adversarial Training Defense Unit 23: Used to enhance the model's robustness against adversarial attacks using a generative adversarial network architecture; specific operations are as follows: C1: Adversarial attack sample generation: ① Generate adversarial perturbations based on voice command features using the FGSM algorithm: ; in, 1. Let L be the perturbation intensity, L be the cross-entropy loss, and y be the true label; 2. For ToF gesture trajectories, generate adversarial examples through affine transformation: , ③ Construct a hybrid adversarial sample set Environmental features + random noise}; C2: Generative Adversarial Network Training: ① Generator G: Input clean samples Output simulated adversarial examples The residual network architecture "3 convolutional layers + 2 deconvolutional layers" is adopted; ② Discriminator D: Input sample S " or or ", Output classification probability P=D(S); ③ The joint loss function is: ; in, For GAN standard loss, The classification loss of the model on adversarial examples, For weighting coefficients; C3: Defense model optimization: ① and ① Train the intelligent interaction model f by mixing samples in a 1:4 ratio; ② Employ an early stopping strategy, stopping training when the accuracy of adversarial examples on the validation set is ≥95%; ③ Output a robust model. This ensures a defense success rate of ≥99% against attacks. Intent reasoning decision unit 24: Based on lightweight BERT, it maps multimodal features to sunroof control intents, such as "fully open" and "sunshade tilt".

[0027] It should be noted that the multimodal sensor fusion unit 21 realizes real-time synchronous processing of multi-source heterogeneous data, providing aligned feature inputs for the spatiotemporal attention mechanism unit 22; the spatiotemporal attention mechanism unit 22 eliminates sensor timing deviations through dynamic weighting, and the output fused features are robustly enhanced by the adversarial training defense unit 23, and finally generate highly reliable sunroof control intent through the intent reasoning decision unit 24.

[0028] Furthermore, it should be noted that the timestamp alignment algorithm in the multimodal sensor fusion unit 21 is as follows: Speech data preprocessing: The acquired speech data is denoised using methods such as spectral subtraction or wavelet transform to remove background noise. Then, feature extraction is performed, extracting features such as Mel-frequency cepstral coefficients and linear prediction cepstral coefficients, converting the speech signal into a feature vector. ToF gesture data preprocessing: The ToF gesture data is denoised using methods such as median filtering or Gaussian filtering to remove noise points. Then, gesture segmentation and feature extraction are performed, extracting features such as the gesture contour and key points, converting the gesture data into a feature vector. Ambient light / temperature and humidity data preprocessing: The ambient light and temperature / humidity data are normalized, mapping the data values ​​to the [0,1] interval to eliminate dimensional differences between data from different sensors. The control intent mapping table of the intent reasoning decision unit 24.

[0029] Fusion feature range Control instructions Safety constraints [0.0,0.3) Off No [0.3,0.6) Sunshade tilt 15° Disabled at > 60 km / h [0.6,1.0] Fully open Limited in rain / > 30 km / h .

[0030] In this embodiment, it should also be noted that the data security and access control module 3 includes a multi-level encryption protection unit 31, an adaptive privacy enhancement unit 32, a behavior-driven access control unit 33, and a security situation awareness unit 34, wherein: the multi-level encryption protection unit 31 is used to encrypt the entire link data of the transmission layer, storage layer, and computing layer; the adaptive privacy enhancement unit 32 is used to dynamically inject differential privacy noise according to data sensitivity to balance model accuracy and privacy protection; the behavior-driven access control unit 33 is based on the ABAC policy engine to analyze user operation patterns in real time and realize millisecond-level dynamic permission upgrades and downgrades; the specific operations are as follows: D1: Multi-dimensional attribute extraction: User attribute U={identity ID, role level, biometric template}; resource attribute R={data type, sensitivity level, access history}; environmental attribute E={timestamp, geographical location, vehicle status}; behavior attribute A={operation frequency, operation sequence, operation context}; D2: Behavioral pattern modeling: ① Construct user operation sequence ① n = the most recent 100 operations; ② Generate behavioral feature vectors through a bidirectional LSTM network. The calculation formula is: ③ Calculate the current operation based on the Gaussian mixture model. Similarity score with historical patterns: ,in, For GMM parameters, K=5 represents the number of mixture components; D3: Dynamic permission decision: ① Define permission function f: 'm' represents the permission dimension, specifically: Where e is the environmental feature vector, This is the weight matrix. ① Use the Sigmoid activation function; ② When the similarity score When the vehicle speed is >80km / h, the following permission downgrade rules are automatically triggered: Ⅰ. Disable the "fully open" permission in voice commands; Ⅱ. Limit the gesture operation range to sunshade adjustment; D4: Permission execution and auditing: ① Permission decision response time ≤5ms; ② Generate an immutable audit log L={t, U, R, E, f(U, R, E, A), permission change record}; ③ Use blockchain technology to synchronize the log to the vehicle security module to ensure data integrity. Security situation awareness unit 34: Used to monitor the system security status in real time and trigger response mechanisms through anomaly detection algorithms and threat intelligence databases. The specific operations are as follows: E1: Real-time aggregation of security logs from the CAN bus, vehicle firewall, and sensor interfaces, standardized into a unified time series format; E2: Composite analysis through the integration of the following detection engines: ① LSTM-based time-series anomaly detection; ② LoC matching based on the threat intelligence database; ③ Vehicle-specific attack pattern identification based on a rule engine; E3: Calculation of a comprehensive risk value using a fuzzy logic algorithm, combined with vehicle status to classify threat levels; E4: Triggering graded responses based on threat level: I. Level 1 "Risk value 30-60": Log recording and warning; II. Level 2 "Risk value 60-80": Restrict non-critical functions; III. Level 3 ">80": Cut off sunroof power and activate backup mechanical controls.

[0031] It should be noted that the multi-level encryption protection unit 31 constructs a full-link data protection barrier, providing a secure data foundation for the adaptive privacy enhancement unit 32; the dynamic noise injection of the adaptive privacy enhancement unit 32 and the real-time permission decision of the behavior-driven access control unit 33 form dual protection, and the security situation awareness unit 34 performs global threat monitoring.

[0032] Furthermore, it should be noted that the transport layer encryption employs the TLS 1.3 protocol, supports ECDHE key exchange (curve secp256r1) and the AES-256-GCM encryption algorithm, and controls the handshake latency to within 20ms, ensuring encrypted transmission of real-time streams such as voice commands and gesture data. Storage layer encryption uses AES-256-XTS encryption for in-vehicle databases (such as SQLite), with keys generated and stored by the Hardware Security Module (HSM) and automatically rotated every 30 days. Sensitive data (such as biometric templates) is additionally stored across multiple physical partitions using the Shamir secret sharing algorithm. Computation layer encryption combines CKKS homomorphic encryption with privacy computing module 1, supporting addition and multiplication operations in encrypted form, ensuring that data does not leak original information during model training, with a computational accuracy loss of ≤3%. Data sensitivity grading standards establish a 5-level sensitivity classification system: Level 1 (Very Low): Environmental temperature and humidity data; Level 3 (Medium): Non-sensitive voice commands (such as "close the sunshade"); Level 5 (Very High): Biometric features such as fingerprints and irises. Differential privacy noise dynamic injection strategy: For level 1-2 data: adopt Gaussian noise ( ), ensuring that the basic statistical characteristics of the data are preserved; for level 4-5 data: adopt Laplacian noise of 0.5 (b=0.02), combined with k-anonymization (k=5) to further reduce identification risk; real-time noise parameter adjustment mechanism: when data is detected for model training, The value is automatically incremented by 30% to balance precision; when used for external data sharing. The value is reduced by 50% to enhance protection.

[0033] In this embodiment, it should also be noted that the adaptive control and execution module 4 includes a PPO reinforcement learning engine unit 41, a multi-objective constraint optimization unit 42, and a real-time control execution unit 43, wherein: the PPO reinforcement learning engine unit 41: trains the sunroof control policy network in encrypted state based on the near-end policy optimization algorithm; the multi-objective constraint optimization unit 42: uses the NSGA-II algorithm to solve for the Pareto optimal solution set of comfort, safety, and energy consumption indicators, and dynamically balances multiple optimization objectives; the real-time control execution unit 43: uses the CAN FD bus to realize 10ms-level sunroof motor control and supports emergency closing.

[0034] It should be noted that the PPO reinforcement learning engine unit 41 trains the control policy under privacy protection, providing a basic policy framework for the multi-objective constraint optimization unit 42; the multi-objective constraint optimization unit 42 uses the NSGA-II algorithm to generate the Pareto optimal solution set, which is finally controlled precisely by the real-time control execution unit 43 through the high-speed bus.

[0035] Furthermore, it should be noted that the NSGA-II algorithm parameter configuration optimization objectives are as follows: Comfort objective: minimize the deviation between the in-vehicle temperature and the set value + the sunroof vibration noise in decibels; Safety objective: maximize the emergency closing response speed + the anti-pinch detection sensitivity; Energy consumption objective: minimize the motor operating current × duration. Key parameters: Population size = 100, crossover probability = 0.9, mutation probability = 0.01, maximum number of iterations = 50, and population diversity is maintained by sorting by crowding degree.

[0036] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0037] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-powered, fully intelligent interactive system for car sunroofs, characterized in that: It includes a privacy computing module (1), an intelligent interaction and edge computing module (2), a data security and access control module (3), and an adaptive control and execution module (4), wherein: The privacy computing module (1) is used to perform efficient collaborative modeling and auditable computation of cross-vehicle intelligent interaction data by integrating lightweight encryption, distributed collaborative learning and trusted execution mechanisms, while ensuring user privacy and data security. The intelligent interaction and edge computing module (2) is used to dynamically weight multimodal features through a spatiotemporal attention fusion mechanism and enhance the robustness of the model by using adversarial training. It processes voice commands, ToF gesture recognition signals and environmental sensor data in real time at the edge and outputs highly reliable sunroof control intentions. The data security and access management module (3) is used to build a unified security framework that integrates privacy protection and dynamic access control. Through multi-level encryption, adaptive privacy enhancement and behavior-driven access control mechanisms, it provides full-link security protection and fine-grained access management for data in the process of multi-party collaboration. The adaptive control and execution module (4) is based on the near-end policy optimization algorithm and uses the encrypted user behavior data provided by the privacy computing module (1) to train a reinforcement learning model and generate a sunroof control strategy that satisfies multi-objective constraints.

2. The automotive sunroof AI-powered full-domain intelligent interaction system according to claim 1, characterized in that, The privacy computing module (1) includes a lightweight encryption unit (11), a distributed collaborative learning unit (12), and a trusted execution environment unit (13), wherein: The lightweight encryption unit (11) uses the CKKS homomorphic encryption algorithm to perform low-latency encryption processing on sensitive data such as voice commands and biometric features. The distributed collaborative learning unit (12) performs secure aggregation of cross-vehicle model parameters based on the FedAvg framework and supports collaborative optimization with differential privacy protection. The Trusted Execution Environment (13) unit is used to isolate sensitive computing through automotive-grade TEE chips, providing hardware-level data integrity and audit trail.

3. The automotive sunroof AI-powered omni-channel intelligent interaction system according to claim 2, characterized in that, The lightweight encryption unit (11) uses the CKKS homomorphic encryption algorithm to perform low-latency encryption processing on sensitive data such as voice commands and biometric features. The specific operation is as follows: A1: Perform vector encoding on voice commands or biometric data to generate plaintext vectors. ; A2: Based on polynomial rings The plaintext vector Encoding as a polynomial ; A3: Use the key generation module to generate a public key. Private key and evaluation key ; A4: Through encryption function Generate ciphertext ; A5: Employing a compression coding strategy and low-order parameter configuration, the encryption latency is controlled within 50ms; the low-order parameter configuration includes: modulus. polynomial order Precision control parameters .

4. The automotive sunroof AI-powered omni-channel intelligent interaction system according to claim 1, characterized in that, The intelligent interaction and edge computing module (2) includes a multimodal sensor fusion unit (21), a spatiotemporal attention mechanism unit (22), an adversarial training defense unit (23), and an intent reasoning decision unit (24), wherein: The multimodal sensor fusion unit (21) is used to process multi-source heterogeneous sensor data such as voice, ToF gesture, and ambient light / temperature and humidity in real time. The spatiotemporal attention mechanism unit (22) is used to eliminate sensor timing deviations of ≤100ms by dynamically weighting asynchronous multimodal features through the STAM network; The adversarial training defense unit (23) is used to enhance the robustness of the model against adversarial attacks by adopting a generative adversarial network architecture. The intent reasoning decision unit (24) is based on lightweight BERT and maps multimodal features to sunroof control intents, such as "fully open" and "tilt sunshade".

5. The automotive sunroof AI-powered omni-channel intelligent interaction system according to claim 4, characterized in that, The spatiotemporal attention mechanism unit (22) is used to dynamically weight asynchronous multimodal features through the STAM network to eliminate sensor timing deviations of ≤100ms. The specific operation is as follows: B1: Timestamp-aligned multimodal sensor data from voice, ToF gestures, and ambient light / temperature / humidity to construct asynchronous feature sequences. ; B2: A sliding time window mechanism is used to perform timing compensation for asynchronous features, with a time window length of [missing information]. ; B3: The dynamic weights of each modality feature are calculated using the spatiotemporal attention module. The weight calculation formula is as follows: ； in, , Let be the query and key vectors for the i-th modality, respectively. This is the scaling factor; B4: Weighted fusion of multimodal features based on weights, outputting a fused feature vector. .

6. The automotive sunroof AI-powered full-domain intelligent interaction system according to claim 4, characterized in that, The adversarial training defense unit (23) adopts a generative adversarial network architecture to enhance the robustness of the model against adversarial attacks. The specific operation is as follows: C1: Generation of adversarial attack samples: ① Based on the characteristics of voice commands, adversarial perturbations are generated using the FGSM algorithm: ； in, Let L be the perturbation strength, L be the cross-entropy loss, and y be the true label; ② For Time-of-Flight (ToF) gesture trajectories, generate adversarial examples through affine transformation: , ; ③ Construct a hybrid adversarial sample set Environmental characteristics + random noise; C2: Generative Adversarial Network Training: ① Generator G: Input clean samples Output simulated adversarial examples It adopts a residual network architecture of "3 convolutional layers + 2 deconvolutional layers"; ② Discriminator D: Input sample S" or or Output the classification probability P=D(S); ③ The joint loss function is: ； in, For GAN standard loss, The classification loss of the model on adversarial examples, These are the weighting coefficients; C3: Defense Model Optimization: ① and Mix them at a 1:4 ratio and train the intelligent interaction model f; ② Adopt an early stopping strategy: stop training when the accuracy of the adversarial examples on the validation set is ≥95%; ③ Output robust model This ensures a defense success rate of ≥99% against the aforementioned counterattacks.

7. The automotive sunroof AI-powered full-domain intelligent interaction system according to claim 1, characterized in that, The data security and access control module (3) includes a multi-level encryption protection unit (31), an adaptive privacy enhancement unit (32), a behavior-driven access control unit (33), and a security situation awareness unit (34), wherein: The multi-level encryption protection unit (31) is used to encrypt the data across the entire link of the transport layer, storage layer, and computing layer. The adaptive privacy enhancement unit (32) is used to dynamically inject differential privacy noise according to data sensitivity, balancing model accuracy and privacy protection. The behavior-driven access control unit (33) analyzes user operation patterns in real time based on the ABAC policy engine to achieve millisecond-level dynamic permission elevation and scaling; The security situation awareness unit (34) is used to monitor the system security status in real time and trigger the response mechanism through anomaly detection algorithms and threat intelligence database.

8. The automotive sunroof AI-powered full-domain intelligent interaction system according to claim 7, characterized in that, The behavior-driven access control unit (33) analyzes user operation patterns in real time based on the ABAC policy engine to achieve millisecond-level dynamic permission elevation and escalation. The specific operation is as follows: D1: Multi-dimensional attribute extraction: User attribute U = {Identity ID, Role Level, Biometric Template}; Resource attribute R = {data type, sensitivity level, access history}; Environmental attribute E = {timestamp, geographic location, vehicle status}; Behavioral attribute A = {Operation frequency, Operation sequence, Operation context}; D2: Behavioral Pattern Modeling ① Constructing user operation sequences n = the most recent 100 operations; ② Generate behavioral feature vectors using a bidirectional LSTM network The calculation formula is: ; ③ Calculate the current operation based on the Gaussian mixture model Similarity score with historical patterns: ,in, These are GMM parameters, where K=5 represents the number of mixture components; D3: Dynamic permission decision-making: ① Define the permission function f: 'm' represents the permission dimension, specifically: Where e is the environmental feature vector, This is the weight matrix. Use the Sigmoid activation function; ② When similarity score And when the vehicle speed is >80km / h, the permission downgrade rule will be automatically triggered: Ⅰ. Disable the "Full Open" permission in voice commands; II. Limit the range of gesture operation to adjusting the sunshade; D4: Privilege Enforcement and Auditing ① The response time for permission decisions is ≤5ms; ② Generate an immutable audit log L={t, U, R, E, f(U, R, E, A), permission change record}; ③ Use blockchain technology to synchronize logs to the vehicle safety module to ensure data integrity.

9. The automotive sunroof AI-powered intelligent interactive system according to claim 7, characterized in that, The security situation awareness unit (34) monitors the system security status in real time and triggers a response mechanism through anomaly detection algorithms and a threat intelligence database. The specific operations are as follows: E1: Real-time aggregation of security logs from CAN bus, vehicle firewall, and sensor interface, standardized into a unified time series format; E2: Composite analysis is performed by integrating the following detection engines: ① LSTM-based timing anomaly detection; ②LoC matching based on threat intelligence database; ③ Identification of car-specific attack patterns based on a rule engine; E3: Calculate the comprehensive risk value using the fuzzy logic algorithm and divide the threat level in combination with the vehicle status; E4: Trigger hierarchical responses according to the threat level: Ⅰ. Level 1 "risk value 30 - 60": Record logs and give warnings; Ⅱ. Level 2 "risk value 60 - 80": Restrict non-critical functions; Ⅲ. Level 3 ">80": Cut off the skylight power supply and activate the backup mechanical control.

10. The automotive sunroof AI-powered full-domain intelligent interaction system according to claim 1, characterized in that, The adaptive control and execution module (4) includes a PPO reinforcement learning engine unit (41), a multi-objective constraint optimization unit (42), and a real-time control execution unit (43), where: The PPO reinforcement learning engine unit (41): Based on the proximal policy optimization algorithm, train the skylight control policy network in an encrypted state; The multi-objective constraint optimization unit (42): Used to solve the Pareto optimal solution set of the comfort, safety, and energy consumption indicators using the NSGA-II algorithm and dynamically balance multiple optimization goals; The real-time control execution unit (43): Used to achieve skylight motor control at the 10ms level through the CAN FD bus and support emergency closing.

Images