AI-driven adaptive interaction system and method for Internet of Vehicles

Abstract

The invention belongs to the technical field of Internet of Vehicles, and discloses an AI drive-based Internet of Vehicles self-adaptive interaction system and an AI drive-based Internet of Vehicles self-adaptive interaction method. The method comprises the steps that signal middleware is constructed, a CAN bus and a service layer of a vehicle are in butt joint through a unified interface of the signal middleware, and CAN signals are collected and subjected to unified conversion and packaging; acquiring customized service data converted and packaged by the signal middleware; performing feature extraction on the user data to obtain user features; taking the user characteristics as input of an instruction analysis model to obtain an initial control instruction; the initial control instruction, the vehicle data and the environment data serve as input of a vehicle control model, and a self-adaptive control instruction is obtained; according to the invention, various types of control instructions are adaptively adjusted by comprehensively considering various vehicle data and environment data, the overall adaptive control of voice interaction and the vehicle driving state can be realized, and an optimization decision can be adaptively made according to the continuously changing running state of the vehicle networking system.

Description

Technical Field

[0001] The present invention relates to the field of vehicle networking technology, and more specifically, to an AI-driven vehicle networking adaptive interaction system and method. Background Art

[0002] Traditional Internet of Vehicles interactions are often based on preset fixed rules and patterns. The commands issued by users are basically executed according to fixed processes, and it is difficult to make flexible adjustments based on real-time vehicle conditions and environmental changes.

[0003] The Chinese patent with the publication number CN118398014A discloses a vehicle-mounted human-machine voice interaction method, device, electronic device and storage medium based on a collaborative adaptive policy gradient algorithm, including obtaining user voice and determining an acoustic action vector based on the user voice; determining a machine action command based on the acoustic action vector and the human-machine interaction strategy, wherein the human-machine interaction strategy is determined based on a Markov decision process and a policy gradient algorithm; and determining a response result for the user voice based on the machine action command. The invention can determine the human-machine interaction strategy based on the Markov decision process, so that the vehicle machine can make the optimal strategy to match the user voice according to the user data and the current state, that is, the static mapping relationship in the traditional human-machine interaction is transformed into a dynamic mapping relationship through the Markov decision process, so that the machine can adjust its response mode according to the real-time situation and demand, so that the interactive system has the ability to learn the driver's experience trajectory from the interactive environment and scene requirements, so that the vehicle-mounted interaction effect is closer to the driver's habits and improves the interaction efficiency. At the same time, the policy gradient algorithm can enable the vehicle machine to spontaneously find the interaction mode closest to the user's habits and find the optimal solution in a specific interaction space.

[0004] Although the above method can meet most scenarios, research and practical application of the above method and existing technology have found that the above method and existing technology have at least the following defects:

[0005] It mainly collects voice data and optimizes voice responses through a policy gradient algorithm, but is unable to achieve adaptive adjustment of other non-voice-related control commands and system state changes of the vehicle; it does not uniformly collect relevant data such as the vehicle's internal state data and external environment data, and the comprehensive utilization of the vehicle's internal state data and external environment data is low, which cannot fully guarantee the vehicle's stability and driving safety in various complex situations.

[0006] In view of this, the present invention proposes an AI-driven Internet of Vehicles adaptive interaction system and method to solve the above problems. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: an AI-driven vehicle networking adaptive interaction method, comprising the following steps:

[0008] Build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals, and performs unified conversion and packaging;

[0009] Collect custom service data after signal middleware conversion and encapsulation; custom service data includes user data, vehicle data and environmental data;

[0010] Extracting features from user data to obtain user features, which include user voice features and user command features;

[0011] Taking the user characteristics as the input of the instruction analysis model, an initial control instruction including instruction values ​​is obtained;

[0012] The initial control command, vehicle data and environmental data are used as inputs of the vehicle control model to obtain an adaptive control command including a command value.

[0013] Furthermore, the method for optimizing the execution efficiency of the adaptive control instruction includes:

[0014] Step 1: Collect current system data; system data includes the number of calls to custom services within a preset time, throughput, resource consumption data when calling custom services, and service response time; resource consumption data includes CPU usage and memory usage, disk I / O usage, disk I / O queue length, disk I / O read and write speed, and network bandwidth utilization;

[0015] Step 2: Using the adaptive control instruction and the system data as inputs of the delay analysis model to obtain a delay risk score of delayed execution of the adaptive control instruction;

[0016] Step 3: Compare the delay risk score with a preset threshold;

[0017] Step 4: If the delay risk score does not exceed the preset threshold, the adaptive control instruction is executed; if the delay risk score exceeds the preset threshold, the adaptive control instruction, system data and delay risk score are used as inputs of the resource optimization model to obtain the resource optimization instruction, and the resource optimization instruction is executed.

[0018] Further, the user data includes voice text and user instructions;

[0019] Methods for collecting user instructions include:

[0020] For physical buttons, obtain the name of the physical button pressed by the user and the corresponding time, and organize it into data in a unified format through the signal middleware;

[0021] For touch screen operations, the user's touch type, starting coordinates, ending coordinates, and corresponding duration are collected and sorted into data in a unified format by the signal middleware;

[0022] The control signal sent by the user via Bluetooth is also connected through the business layer.

[0023] Furthermore, the method for obtaining the user voice feature includes:

[0024] Use the Chinese word segmentation tool to segment each user's voice text, remove duplicate words, obtain the first u words in each voice text, use the pre-trained word vector model to convert the obtained words into word vectors, and concatenate the word vectors as user voice features.

[0025] Furthermore, the method for obtaining the user instruction feature includes:

[0026] The user command features include user command features of physical keys, interactive touches, and control signals; and the method for obtaining the user command features of physical keys includes:

[0027] Encode the physical keys and operation types to obtain user instruction features including key codes, operation type codes and operation time;

[0028] The method for obtaining the user instruction feature of the interactive touch includes:

[0029] Encode the user touch type to obtain user instruction features including touch type code, start coordinates, end coordinates and duration;

[0030] The method for obtaining the user instruction feature of the control signal includes:

[0031] The control signal is spectrally analyzed to collect the signal spectrum peak and signal duration, and the modulation mode of the Bluetooth signal is identified based on the cyclostationary algorithm.

[0032] Furthermore, the method for identifying the modulation mode of the Bluetooth signal based on the cyclostationary algorithm includes:

[0033] The control signal is sampled using a signal sampling device to obtain a discrete control signal sequence x[n], the sampled control signal is preprocessed to remove noise and normalize, and then the cyclic autocorrelation function is used Calculate the discrete control signal sequence x[n] to obtain the cyclic autocorrelation function Regarding the relationship between the discrete control signal sequence x[n], the relationship is as follows:

[0034]

[0035] Wherein, α is the cyclic frequency of the cyclic autocorrelation function; x*[n] is the conjugate of x[n]; n is the discrete point of the control signal sequence; j is the imaginary unit; N is the length of the control signal sequence; m is the discrete time delay; e and π are mathematical constants;

[0036] According to the cyclic autocorrelation function Plotting the cyclic autocorrelation function on the relation of discrete control signal sequence Regarding the change curve of the discrete control signal sequence, the peak features of the change curve are extracted. The peak features include peak position, peak amplitude and peak number. The peak features are used as the input of the pre-trained modulation classification model to obtain the modulation mode label of the control signal. The signal spectrum peak, signal duration and modulation mode label are concatenated to obtain the user command features of the control signal.

[0037] Furthermore, the training method of the modulation classification model includes:

[0038] Collecting a group A of modulation training data in advance, the modulation training data includes peak characteristics and modulation mode labels of control signals;

[0039] Each group of modulation training data is used as input of a modulation classification model, the modulation classification model uses the modulation mode label of the control signal corresponding to each group of peak features as output, and uses the modulation mode label of the actual control signal corresponding to each group of peak features as a prediction target; minimizing the sum of the prediction errors of the modulation mode labels of all control signals is used as a training target; the modulation classification model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the modulation classification model is a deep neural network model;

[0040] The loss function value of the modulation classification model is mean square error;

[0041] Prediction error formula: a is the peak feature of the ath group; A is the number of groups of peak features; is the modulation mode label of the control signal corresponding to the peak feature of the ath group; a is the modulation mode label of the actual control signal corresponding to the peak feature of the ath group.

[0042] Furthermore, the vehicle data includes driving data, vehicle system status and tire pressure monitoring data; the driving data includes the vehicle's current speed, acceleration, steering angle and braking force; the vehicle system status includes engine data, transmission data and battery data; the engine data includes the engine speed and temperature, the transmission data includes the engine gear, temperature and pressure, the battery data includes the battery charge and working status, wherein the working status includes operation, charging and sleep, and the working status is represented by encoding; the tire pressure monitoring data is the vehicle tire pressure data; the environmental data includes ambient temperature, humidity and light intensity.

[0043] Furthermore, the training method of the instruction analysis model includes:

[0044] Collecting a group B of instruction training data in advance, the instruction training data includes user characteristics and initial control instructions;

[0045] Each group of instruction training data is used as the input of the instruction analysis model. The instruction analysis model uses the initial control instructions corresponding to each group of user features as output, and uses the actual initial control instructions corresponding to each group of user features as the prediction target; minimizing the sum of the prediction errors of all initial control instructions is used as the training target; the instruction analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the instruction analysis model is a deep neural network model;

[0046] The loss function value of the instruction analysis model is mean square error;

[0047] Prediction error formula: b is the user feature of the bth group; B is the number of user feature groups; is the initial control instruction corresponding to the user characteristics of group b; b is the actual initial control instruction corresponding to the user characteristics of group b.

[0048] Furthermore, the training method of the vehicle control model includes:

[0049] Collecting a group C of control training data in advance, the control training data includes initial control instructions, vehicle data, environmental data and adaptive control instructions;

[0050] Each group of control training data is used as the input of the vehicle control model, the vehicle control model uses the adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as output, and uses the actual adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as prediction targets; minimizing the sum of prediction errors of all adaptive control instructions is used as the training target; the vehicle control model is trained until the sum of prediction errors reaches convergence and the training is stopped; the vehicle control model is a deep neural network model;

[0051] The vehicle control model loss function value is the mean square error;

[0052] Prediction error formula: c is the cth group of initial control instructions, vehicle data and environmental data; C is the number of groups of initial control instructions, vehicle data and environmental data; is the adaptive control instruction corresponding to the initial control instruction, vehicle data and environmental data of the cth group; c is the actual adaptive control instruction corresponding to the initial control instruction, vehicle data and environmental data of the cth group.

[0053] The AI-driven Internet of Vehicles adaptive interaction system is used to implement the AI-driven Internet of Vehicles adaptive interaction method, including:

[0054] Unified docking module: used to build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals and performs unified conversion and packaging;

[0055] Data acquisition module: used to collect customized service data after signal middleware conversion and encapsulation; customized service data includes user data, vehicle data and environmental data;

[0056] The first analysis module is used to extract features from user data to obtain user features, which include user voice features and user command features;

[0057] The second analysis module is used to use the user characteristics as the input of the instruction analysis model to obtain the initial control instruction including the instruction value;

[0058] Adaptive adjustment module: used to take the initial control instructions, vehicle data and environmental data as inputs of the vehicle control model to obtain adaptive control instructions containing instruction values.

[0059] Technical effects and advantages of the AI-driven Internet of Vehicles adaptive interaction system and method of the present invention:

[0060] The present invention can realize the centralized and standardized collection of various key data in the Internet of Vehicles by constructing a signal middleware to uniformly connect the CAN bus and business layer of the vehicle, and uniformly convert and encapsulate the collected CAN signals, thereby improving the efficiency of data collection and ensuring the integrity and accuracy of the data. Comprehensively considering various types of vehicle data and environmental data to adaptively adjust various types of control instructions, it can realize the overall adaptive control of voice interaction and vehicle driving status; through deep integration of the real-time status information and environmental data of the vehicle, it can adaptively make optimization decisions according to the changing operating status of the Internet of Vehicles system to ensure driving safety and stability; and through the analysis of system data, it can realize the refined adjustment of resource allocation, etc., thereby optimizing resource scheduling without affecting the execution of adaptive control instructions, shortening the service response time corresponding to the control instructions, and thus improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the AI-driven Internet of Vehicles adaptive interaction method mentioned in Embodiment 1 of the present invention;

[0062] Figure 2 A flow chart of a method for uniformly converting and packaging CAN signals mentioned in Embodiment 1 of the present invention;

[0063] Figure 3 This is a flow chart of the resource adaptive scheduling strategy mentioned in Example 2 of the present invention;

[0064] Figure 4 This is a block diagram of the AI-driven Internet of Vehicles adaptive interaction system mentioned in Example 3 of the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] Example 1

[0067] See also Figure 1 As shown, this embodiment provides an AI-driven Internet of Vehicles adaptive interaction method, including the following steps:

[0068] Build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals, and performs unified conversion and packaging;

[0069] Reference Figure 2, the methods for uniformly converting and packaging CAN signals include:

[0070] Step S1, collecting and analyzing the original signals of different vehicle models, including the data type, unit, encoding method and composite signal of the signal;

[0071] Step S2: Based on the results of signal analysis, a signal mapping table is established to match the original signals of different vehicle models with the standard format of the middleware. The signal mapping table includes information such as the name of the signal, the original encoding method, the standard format of the middleware, and the conversion algorithm. For example, for vehicle speed signals, the encoding methods of different vehicle models may be different. Some vehicle models directly transmit values ​​in kilometers per hour, while some vehicle models may use the frequency of pulse signals to represent the vehicle speed. In the mapping table, these different encoding methods can be matched with the standard speed format of the middleware (such as the value of kilometers per hour).

[0072] Step S3, design a conversion algorithm, including a conversion algorithm for the data type, unit, and encoding method of the signal. For example, convert the integer type in the original signal to the floating-point type required by the middleware, or convert the string type to a numeric type, etc. When performing data type conversion, it is necessary to pay attention to the accuracy and range of the data to ensure that the converted result is accurate. For example, convert the unit miles per hour in the original signal to the standard unit kilometers per hour of the middleware, or convert pounds per square inch to Pascals, etc. Convert the binary code in the original signal to decimal code, or convert a specific encoding format to a general encoding format, etc.

[0073] Step S4, physical quantity conversion: convert the CAN signal according to the conversion algorithm.

[0074] Step S5, design a unified data structure: create a structure of the CAN signal after application encapsulation conversion, the structure includes information such as signal name, signal source identification, converted signal and timestamp; fill the structure according to the CAN signal.

[0075] Step S6, encapsulating into objects or data structures: further organizing the single encapsulated CAN signal into a data set, such as an array, a linked list or other container-type objects, to facilitate batch processing and management.

[0076] By building a signal middleware to uniformly connect the vehicle's CAN bus and business layer, and uniformly convert and encapsulate the collected CAN signals, it is possible to achieve centralized and standardized collection of various key data in the Internet of Vehicles, improve the efficiency of data collection, and ensure the integrity and accuracy of the data.

[0077] Collect custom service data after signal middleware conversion and encapsulation; custom service data includes user data, vehicle data and environmental data;

[0078] User data includes voice text and user commands. User voice is collected through the in-vehicle voice collection device and converted into voice text through the text converter, and user commands are collected through the business layer. Collecting user data can clarify user needs and facilitate multi-faceted optimization and precise adaptation when outputting control commands in the later stage.

[0079] Vehicle data is collected through various sensors on the vehicle, including driving data, vehicle system status and tire pressure monitoring data; driving data includes the vehicle's current speed, acceleration, steering angle and braking force; vehicle system status includes engine data, transmission data and battery data; engine data includes engine speed and temperature, transmission data includes engine gear, temperature and pressure, battery data includes battery power and working status, among which working status includes operation, charging and sleep; the working status is encoded and represented by encoding, such as 2 for operation, 1 for charging, and 0 for sleep. Tire pressure monitoring data is vehicle tire pressure data; environmental data is collected through sensors, including ambient temperature, humidity and light intensity. Collecting vehicle data and environmental data can adaptively adjust control instructions (for example, in an autonomous vehicle scenario, the user issues a control instruction to set the vehicle speed to 80 km / h and head to the destination. The current vehicle initial speed is 0 and the environmental conditions are good. The system will gradually increase the speed to nearly 80 km / h according to the normal acceleration curve. However, if during driving, the vehicle data shows that the tire pressure suddenly drops to a dangerous threshold, the system will automatically adjust the speed instruction previously set by the user. It may limit the speed to 60 km / h or even lower to ensure the vehicle's driving safety in the case of insufficient tire pressure and slippery roads, and at the same time send users prompts of abnormal tire pressure and speed adjustment). While meeting user needs, it maintains the stability of the vehicle and ensures the user's driving safety.

[0080] Methods for collecting user instructions include:

[0081] For physical buttons, the name of the physical button pressed by the user and the corresponding time are obtained, and the signal middleware is used to organize the data into a unified format, such as "button: XX, operation type: press / release, time: XX".

[0082] For touch screen operations, the user's touch type, starting coordinates, ending coordinates and corresponding duration are collected and organized into data in a unified format by the signal middleware, such as "touch type: slide / click, starting coordinates: (X1, Y1), ending coordinates: (X2, Y2), duration: XX".

[0083] The control signal sent by the user via Bluetooth is also connected through the business layer.

[0084] Feature extraction is performed on user data to obtain user features, which include user voice features and user command features.

[0085] Methods for obtaining user voice features include:

[0086] Use the Chinese word segmentation tool to segment each voice text of the user, remove duplicate words, and obtain the first u words in each voice text respectively (u is preferably 30, representing the first 30 words in the voice text). Use the pre-trained word vector model to convert the obtained words into word vectors, and concatenate the word vectors as user voice features.

[0087] The method for obtaining the user instruction feature includes:

[0088] The user command features include user command features of physical keys, interactive touches, and control signals; and the method for obtaining the user command features of physical keys includes:

[0089] Encode the physical keys and operation types to obtain user instruction features including key codes, operation type codes and operation time;

[0090] The method for obtaining the user instruction feature of interactive touch includes:

[0091] Encode the user touch type to obtain user instruction features including touch type code, start coordinates, end coordinates and duration;

[0092] The method for obtaining the user instruction characteristic of the control signal includes:

[0093] Perform spectrum analysis on the control signal, collect the signal spectrum peak and signal duration, and identify the modulation mode of the Bluetooth signal based on the cyclostationary algorithm;

[0094] The method for identifying the modulation mode of the Bluetooth signal based on the cyclostationary algorithm includes:

[0095] The control signal is sampled using a signal sampling device to obtain a discrete control signal sequence x[n], the sampled control signal is preprocessed to remove noise and normalize, and then the cyclic autocorrelation function is used Calculate the discrete control signal sequence x[n] to obtain the cyclic autocorrelation function Regarding the relationship between the discrete control signal sequence x[n], the relationship is as follows:

[0096]

[0097] Where α is the cyclic frequency of the cyclic autocorrelation function; x *[n] is the conjugate of x[n]; n is a discrete point of the control signal sequence; j is an imaginary unit; N is the length of the control signal sequence; m is a discrete time delay; e and π are mathematical constants.

[0098] According to the cyclic autocorrelation function Plotting the cyclic autocorrelation function on the relation of discrete control signal sequence Regarding the change curve of the discrete control signal sequence, the peak features of the change curve are extracted. The peak features include peak position, peak amplitude and peak number. The peak features are used as the input of the pre-trained modulation classification model to obtain the modulation mode label of the control signal (such as amplitude shift keying (ASK), frequency shift keying (FSK) and phase shift keying (PSK), etc.). The signal spectrum peak, signal duration and modulation mode label are spliced ​​to obtain the user command feature of the control signal.

[0099] The training method of the modulation classification model includes:

[0100] A group A of modulation training data is collected in advance, and the modulation training data includes peak characteristics and modulation mode labels of control signals.

[0101] Each group of modulation training data is used as the input of the modulation classification model. The modulation classification model takes the modulation mode label of the control signal corresponding to each group of peak features as the output, and takes the modulation mode label of the actual control signal corresponding to each group of peak features as the prediction target; minimizing the sum of the prediction errors of the modulation mode labels of all control signals is used as the training target; the modulation classification model is trained until the sum of the prediction errors reaches convergence, and the training is stopped; the modulation classification model is a deep neural network model.

[0102] The loss function value of the modulation classification model is the mean square error.

[0103] Prediction error formula: a is the peak feature of the ath group; A is the number of groups of peak features; is the modulation mode label of the control signal corresponding to the peak feature of the ath group; a is the modulation mode label of the actual control signal corresponding to the peak feature of the ath group.

[0104] Taking the user characteristics as the input of the instruction analysis model, an initial control instruction including instruction values ​​is obtained;

[0105] The training method of the instruction analysis model includes:

[0106] A group B of instruction training data is collected in advance, and the instruction training data includes user characteristics and initial control instructions.

[0107] Each group of instruction training data is used as the input of the instruction analysis model. The instruction analysis model takes the initial control instructions corresponding to each group of user features as output, and the actual initial control instructions corresponding to each group of user features as prediction targets; minimizing the sum of prediction errors of all initial control instructions is used as the training target; the instruction analysis model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the instruction analysis model is a deep neural network model.

[0108] The loss function value of the instruction analysis model is the mean square error.

[0109] Prediction error formula: b is the user feature of the bth group; B is the number of user feature groups; is the initial control instruction corresponding to the user characteristics of group b; b is the actual initial control instruction corresponding to the user characteristics of group b.

[0110] The initial control command, vehicle data and environmental data are used as inputs of the vehicle control model to obtain an adaptive control command containing a command value, that is, a new control command is obtained by integrating the vehicle data and the environmental data and adaptively adjusting the initial control command.

[0111] The training methods for the vehicle control model include:

[0112] A group C of control training data is collected in advance, and the control training data includes initial control instructions, vehicle data, environmental data and adaptive control instructions.

[0113] Each group of control training data is used as the input of the vehicle control model. The vehicle control model uses the adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as output, and uses the actual adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as prediction targets; minimizing the sum of prediction errors of all adaptive control instructions is used as the training target; the vehicle control model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the vehicle control model is a deep neural network model.

[0114] The vehicle control model loss function value is the mean square error.

[0115] Prediction error formula: c is the cth group of initial control instructions, vehicle data and environmental data; C is the number of groups of initial control instructions, vehicle data and environmental data; is the adaptive control instruction corresponding to the initial control instruction, vehicle data and environmental data of the cth group; c is the actual adaptive control instruction corresponding to the initial control instruction, vehicle data and environmental data of the cth group.

[0116] Example 2

[0117] See also Figure 3 This embodiment provides an AI-driven Internet of Vehicles adaptive interaction method to optimize the execution efficiency of adaptive control instructions; the method comprises the following steps:

[0118] Step 1: Collect current system data; system data includes the number of calls to custom services (such as vehicle health detection and vehicle adaptive cruise control, etc.) within a preset time (such as the number of calls to vehicle adaptive cruise control issued by users within a week), throughput, resource consumption data when calling custom services, and service response time; resource consumption data includes CPU usage and memory occupancy, disk I / O usage, disk I / O queue length, disk I / O read and write speed, and network bandwidth utilization. Collecting system data can fully reflect the performance of the system when implementing custom services, facilitate the analysis of system data by the later delay analysis model, and realize fine-tuning of resource allocation, etc., so as to optimize resource scheduling without affecting the execution of adaptive control instructions, shorten the service response time corresponding to the control instructions, and thus improve the user experience.

[0119] Step 2: Use the adaptive control instruction and system data as inputs of the delay analysis model to obtain a delay risk score for delayed execution of the adaptive control instruction.

[0120] The training method of the latency analysis model includes:

[0121] D groups of delay training data are collected in advance, and the delay training data includes adaptive control instructions, system data and delay risk scores.

[0122] Each group of delay training data is used as the input of the delay analysis model, the delay analysis model takes the delay risk score corresponding to each group of adaptive control instructions and system data as the output, and takes the actual delay risk score corresponding to each group of adaptive control instructions and system data as the prediction target; minimizing the sum of the prediction errors of all delay risk scores is used as the training target; the delay analysis model is trained until the sum of the prediction errors reaches convergence, and the training is stopped; the delay analysis model is a deep neural network model.

[0123] The delay analysis model loss function value is the mean square error.

[0124] Prediction error formula: d is the dth group of adaptive control instructions and system data; D is the number of groups of adaptive control instructions and system data; is the delay risk score corresponding to the dth group of adaptive control instructions and system data; d is the actual delay risk score corresponding to the dth group of adaptive control instructions and system data.

[0125] Step 3: Compare the delay risk score with a preset threshold;

[0126] Step 4: If the delay risk score does not exceed the preset threshold, the adaptive control instruction is executed; if the delay risk score exceeds the preset threshold, the adaptive control instruction, system data and delay risk score are used as inputs of the resource optimization model to obtain the resource optimization instruction, and the resource optimization instruction is executed; the training method of the resource optimization model can refer to the training method of the vehicle control model.

[0127] For example, the current adaptive control instruction is to execute the vehicle adaptive cruise instruction on the highway. Through the delay analysis model, the delay risk score K is calculated; the delay risk score K is compared with the preset threshold T. If K≤T, the adaptive control instruction is executed; if K>T, the adaptive control instruction, system data and delay risk score K are used as inputs of the resource optimization model to obtain and execute resource optimization instructions. The resource optimization instructions may include reducing CPU usage and memory occupancy, etc. The corresponding execution method is to close the programs executed during the current driving process, such as pausing data synchronization and social media applications running in the background.

[0128] Step 5: storing the adaptive control instruction, the system data, the delay risk score and the model parameters of the corresponding delay analysis model into the memory unit of the delay analysis model for the next delay analysis.

[0129] Example 3

[0130] See also Figure 4 As shown, this embodiment provides an AI-driven Internet of Vehicles adaptive interaction system, including:

[0131] Unified docking module: used to build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals and performs unified conversion and packaging;

[0132] Data acquisition module: used to collect customized service data after signal middleware conversion and encapsulation; customized service data includes user data, vehicle data and environmental data;

[0133] The first analysis module is used to extract features from user data to obtain user features, which include user voice features and user command features;

[0134] The second analysis module is used to use the user characteristics as the input of the instruction analysis model to obtain the initial control instruction including the instruction value;

[0135] Adaptive adjustment module: used to take the initial control instructions, vehicle data and environmental data as inputs of the vehicle control model to obtain adaptive control instructions containing instruction values.

[0136] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

[0137] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The AI-driven Internet of Vehicles adaptive interaction method is characterized by: The steps include: Build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals, and performs unified conversion and packaging; Collect custom service data after signal middleware conversion and encapsulation; custom service data includes user data, vehicle data and environmental data; Extracting features from user data to obtain user features, which include user voice features and user command features; Taking the user characteristics as the input of the instruction analysis model, an initial control instruction including instruction values ​​is obtained; The initial control command, vehicle data and environmental data are used as inputs of the vehicle control model to obtain an adaptive control command including a command value.

2. The AI-driven Internet of Vehicles adaptive interaction method according to claim 1 is characterized in that: The method for optimizing the execution efficiency of the adaptive control instruction includes: Step 1: Collect current system data; system data includes the number of calls to custom services within a preset time, throughput, resource consumption data when calling custom services, and service response time; resource consumption data includes CPU usage and memory usage, disk I / O usage, disk I / O queue length, disk I / O read and write speed, and network bandwidth utilization; Step 2: Using the adaptive control instruction and the system data as inputs of the delay analysis model to obtain a delay risk score of delayed execution of the adaptive control instruction; Step 3: Compare the delay risk score with a preset threshold; Step 4: If the delay risk score does not exceed the preset threshold, the adaptive control instruction is executed; if the delay risk score exceeds the preset threshold, the adaptive control instruction, system data and delay risk score are used as inputs of the resource optimization model to obtain the resource optimization instruction, and the resource optimization instruction is executed.

3. The AI-driven Internet of Vehicles adaptive interaction method according to claim 1 is characterized in that: The user data includes voice text and user instructions; Methods for collecting user instructions include: For physical buttons, obtain the name of the physical button pressed by the user and the corresponding time, and organize it into data in a unified format through the signal middleware; For touch screen operations, the user's touch type, starting coordinates, ending coordinates, and corresponding duration are collected and sorted into data in a unified format by the signal middleware; The control signal sent by the user via Bluetooth is also connected through the business layer.

4. The AI-driven Internet of Vehicles adaptive interaction method according to claim 3 is characterized in that: The method for obtaining the user voice feature includes: Use the Chinese word segmentation tool to segment each user's voice text, remove duplicate words, obtain the first u words in each voice text, use the pre-trained word vector model to convert the obtained words into word vectors, and concatenate the word vectors as user voice features.

5. The AI-driven Internet of Vehicles adaptive interaction method according to claim 3 is characterized in that: The method for obtaining the user instruction feature includes: The user command features include user command features of physical keys, interactive touches, and control signals; and the method for obtaining the user command features of physical keys includes: Encode the physical keys and operation types to obtain user instruction features including key codes, operation type codes and operation time; The method for obtaining the user instruction feature of the interactive touch includes: Encode the user touch type to obtain user instruction features including touch type code, start coordinates, end coordinates and duration; The method for obtaining the user instruction feature of the control signal includes: The control signal is spectrally analyzed to collect the signal spectrum peak and signal duration, and the modulation mode of the Bluetooth signal is identified based on the cyclostationary algorithm.

6. The AI-driven Internet of Vehicles adaptive interaction method according to claim 5 is characterized in that: The method for identifying the modulation mode of the Bluetooth signal based on the cyclostationary algorithm includes: The control signal is sampled using a signal sampling device to obtain a discrete control signal sequence x[n], the sampled control signal is preprocessed to remove noise and normalize, and then the cyclic autocorrelation function is used Calculate the discrete control signal sequence x[n] to obtain the cyclic autocorrelation function Regarding the relationship between the discrete control signal sequence x[n], the relationship is as follows: Where α is the cyclic frequency of the cyclic autocorrelation function; x * [n] is the conjugate of x[n]; n is a discrete point of the control signal sequence; j is an imaginary unit; N is the length of the control signal sequence; m is the discrete time delay; e and π are mathematical constants; According to the cyclic autocorrelation function Plotting the cyclic autocorrelation function on the relation of discrete control signal sequence Regarding the change curve of the discrete control signal sequence, the peak features of the change curve are extracted. The peak features include peak position, peak amplitude and peak number. The peak features are used as the input of the pre-trained modulation classification model to obtain the modulation mode label of the control signal. The signal spectrum peak, signal duration and modulation mode label are concatenated to obtain the user command features of the control signal.

7. The AI-driven Internet of Vehicles adaptive interaction method according to claim 6 is characterized in that: The training method of the modulation classification model includes: Collecting a group A of modulation training data in advance, the modulation training data includes peak characteristics and modulation mode labels of control signals; Each group of modulation training data is used as the input of the modulation classification model. The modulation classification model takes the modulation mode label of the control signal corresponding to each group of peak features as the output, and takes the modulation mode label of the actual control signal corresponding to each group of peak features as the prediction target; minimizing the sum of the prediction errors of the modulation mode labels of all control signals is used as the training target; the modulation classification model is trained until the sum of the prediction errors reaches convergence, and the training is stopped; the modulation classification model is a deep neural network model.

8. The AI-driven Internet of Vehicles adaptive interaction method according to claim 1, characterized in that: The vehicle data includes driving data, vehicle system status and tire pressure monitoring data; the driving data includes the vehicle's current speed, acceleration, steering angle and braking force; The vehicle system status includes engine data, transmission data and battery data; the engine data includes the engine speed and temperature, the transmission data includes the engine gear, temperature and pressure, the battery data includes the battery charge and working status, among which the working status includes operation, charging and sleep, and the working status is represented by encoding; the tire pressure monitoring data is the vehicle tire pressure data; the environmental data includes ambient temperature, humidity and light intensity.

9. The AI-driven Internet of Vehicles adaptive interaction method according to claim 1, characterized in that: The training method of the instruction analysis model includes: Collecting a group B of instruction training data in advance, the instruction training data includes user characteristics and initial control instructions; Each group of instruction training data is used as the input of the instruction analysis model. The instruction analysis model takes the initial control instructions corresponding to each group of user features as output, and the actual initial control instructions corresponding to each group of user features as prediction targets; minimizing the sum of prediction errors of all initial control instructions is used as the training target; the instruction analysis model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the instruction analysis model is a deep neural network model.

10. The AI-driven Internet of Vehicles adaptive interaction method according to claim 1, characterized in that: The training method of the vehicle control model includes: Collecting a group C of control training data in advance, the control training data includes initial control instructions, vehicle data, environmental data and adaptive control instructions; Each group of control training data is used as the input of the vehicle control model. The vehicle control model uses the adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as output, and uses the actual adaptive control instructions corresponding to each group of initial control instructions, vehicle data and environmental data as prediction targets; minimizing the sum of prediction errors of all adaptive control instructions is used as the training target; the vehicle control model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the vehicle control model is a deep neural network model.

11. The AI-driven Internet of Vehicles adaptive interactive system is characterized by: The method for implementing the AI-driven Internet of Vehicles adaptive interaction method according to any one of claims 1 to 10 comprises: Unified docking module: used to build signal middleware, whose unified interface connects the vehicle's CAN bus and business layer, collects CAN signals and performs unified conversion and packaging; Data acquisition module: used to collect customized service data after signal middleware conversion and encapsulation; customized service data includes user data, vehicle data and environmental data; The first analysis module is used to extract features from user data to obtain user features, which include user voice features and user command features; The second analysis module is used to use the user characteristics as the input of the instruction analysis model to obtain the initial control instruction including the instruction value; Adaptive adjustment module: used to take the initial control instructions, vehicle data and environmental data as inputs of the vehicle control model to obtain adaptive control instructions containing instruction values.

Images