Model training method, prediction method and device based on neural network search

Abstract

The present disclosure provides a model training method, prediction method and device based on neural network search, including: obtaining model training samples; establishing a probability distribution model corresponding to a parameter search space, wherein the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters; determining a target model structure based on a preset proxy model and a probability distribution model; using a preset loss function, determining a model loss value corresponding to the target model structure based on the model training samples; updating the parameter search space, the probability distribution model and the current number of iterations based on the model loss value corresponding to the target model structure, until the current number of iterations meets the model iteration condition, and then obtaining a braking intention prediction model. Thus, the time cost in the model design process is greatly reduced, the stability of the model structure is ensured, and the model prediction efficiency is effectively improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of intelligent automobile technology, and in particular, to a model training method, a prediction method and a device applicable to neural network search-based model training method, a prediction method and a device. Background Art

[0002] With the promulgation of relevant national policies and the upgrading of the automobile industry, the research on braking intention recognition methods conforms to the development trend of electrification and intelligence in the automotive field. Advanced braking intention recognition methods can improve the intelligence level of vehicles and effectively avoid some collision accidents.

[0003] In order to reduce the number of motor vehicle traffic accidents and protect the lives and property of drivers and passengers, research on braking intention recognition methods has become a trend in the automotive industry. The existing braking intention recognition method is mainly predicted through the existing Transformer model. However, the existing Transformer model mainly relies on the designer's experience when designing the model structure, and the model performance is unstable, which affects the model prediction efficiency. Summary of the invention

[0004] The embodiments described herein provide a model training method, prediction method and device based on neural network search, which overcome the above-mentioned problems.

[0005] In a first aspect, according to the content of the present disclosure, a model training method based on neural network search is provided, comprising:

[0006] Obtaining a model training sample, wherein the model training sample includes: an electromyographic signal corresponding to a driving user of the intelligent vehicle and driving state information of the intelligent vehicle;

[0007] Establish a probability distribution model corresponding to the parameter search space, the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters, each group of model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure;

[0008] Based on the preset proxy model and the probability distribution model, the target model structure is determined, and the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure;

[0009] Using a preset loss function, determine the model loss value corresponding to the target model structure based on the model training samples;

[0010] Based on the model loss value corresponding to the target model structure, the parameter search space, probability distribution model and current iteration number are updated until the current iteration number meets the model iteration condition, and the braking intention prediction model is obtained.

[0011] In a second aspect, according to the present disclosure, a method for predicting a driver's braking intention is provided, comprising:

[0012] Acquire vehicle driving data, the vehicle driving data including: current electromyographic signals of the vehicle driving user and current driving state information of the vehicle;

[0013] Inputting vehicle driving data into a braking intention prediction model, and determining a braking prediction intensity based on an output result of the braking intention prediction model;

[0014] Among them, the braking intention prediction model is a braking intention prediction model trained and generated by the method in the first aspect.

[0015] In a third aspect, according to the content of the present disclosure, a model training device based on neural network search is provided, comprising:

[0016] A first acquisition module is used to acquire a model training sample, wherein the model training sample includes: an electromyographic signal corresponding to a driving user of the smart vehicle and driving state information of the smart vehicle;

[0017] Establish a module for establishing a probability distribution model corresponding to a parameter search space, wherein the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters, each group of model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure;

[0018] A first determination module is used to determine the target model structure based on a preset proxy model and a probability distribution model, wherein the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure;

[0019] A second determination module is used to determine a model loss value corresponding to a target model structure based on a model training sample using a preset loss function;

[0020] The updating module is used to update the parameter search space, the probability distribution model and the current iteration number based on the model loss value corresponding to the target model structure, until the current iteration number meets the model iteration condition, and the braking intention prediction model is obtained.

[0021] In a fourth aspect, according to the present disclosure, a device for predicting a driver's braking intention is provided, comprising:

[0022] The second acquisition module is used to acquire vehicle driving data, wherein the vehicle driving data includes: current electromyographic signals of the vehicle driving user and current driving state information of the vehicle;

[0023] An input module, used for inputting vehicle driving data into a braking intention prediction model, and determining a braking prediction intensity based on an output result of the braking intention prediction model;

[0024] Among them, the braking intention prediction model is a braking intention prediction model trained and generated by the method in the first aspect.

[0025] In a fifth aspect, a computer device is provided, comprising a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the model training method based on neural network search in any of the above embodiments are implemented, or the steps of the method for predicting driving braking intention in any of the above embodiments are implemented.

[0026] In a sixth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the model training method based on neural network search in any of the above embodiments are implemented, or the steps of the method for predicting driving braking intention in any of the above embodiments are implemented.

[0027] The model training method based on neural network search provided by the embodiment of the present application obtains model training samples, which include: electromyographic signals corresponding to the driving user of the intelligent vehicle and the driving state information of the intelligent vehicle; establishes a probability distribution model corresponding to the parameter search space, which includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters, each group of model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure; based on the preset proxy model and the probability distribution model, the target model structure is determined, and the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure; using the preset loss function, the model loss value corresponding to the target model structure is determined based on the model training samples; based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current number of iterations are updated until the current number of iterations meets the model iteration condition, and the braking intention prediction model is obtained. In this way, by adaptively searching and determining the model structure, the time cost in the model design process is greatly reduced, the stability of the model structure is guaranteed, and the model prediction efficiency is effectively improved.

[0028] The above description is only an overview of the technical solution of the embodiment of the present application. In order to more clearly understand the technical means of the embodiment of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiment of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly described below. It should be noted that the drawings described below only relate to some embodiments of the present disclosure, but are not intended to limit the present disclosure, wherein:

[0030] Figure 1 It is a flowchart of a model training method based on neural network search provided by an embodiment of the present disclosure.

[0031] Figure 2A It is a schematic diagram of a neural network structure search principle provided by an embodiment of the present disclosure.

[0032] Figure 2B It is a signal schematic diagram of the original electrical signal of the driver's leg provided by an embodiment of the present disclosure.

[0033] Figure 2C It is a schematic diagram of automobile driving status information provided by an embodiment of the present disclosure.

[0034] Figure 2D It is a curve diagram of the score value of the Transformer model under different structures provided by an embodiment of the present disclosure.

[0035] Figure 2E-2H It is a schematic diagram showing the braking intention prediction results under different driver braking intensity peaks provided by an embodiment of the present disclosure.

[0036] Figure 3 It is a flow chart of a method for predicting driving braking intention provided by an embodiment of the present disclosure.

[0037] Figure 4 It is a structural diagram of a model training device based on neural network search provided in an embodiment of the present disclosure.

[0038] Figure 5 It is a structural schematic diagram of a device for predicting driving braking intention provided in an embodiment of the present disclosure.

[0039] Figure 6 It is a structural diagram of a computer device provided by an embodiment of the present disclosure.

[0040] It should be noted that the elements in the drawings are schematic and not drawn to scale. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work also fall within the scope of protection of the present disclosure.

[0042] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by a person skilled in the art to which the subject matter of the present disclosure belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the specification and the relevant art, and will not be interpreted in an idealized or overly formal form unless otherwise explicitly defined herein. As used herein, a statement that two or more parts are "connected" or "coupled" together shall mean that the parts are joined together directly or through one or more intermediate components.

[0043] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase "embodiments" in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0044] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists, A and B exist at the same time, and B exists. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. Terms such as "first" and "second" are only used to distinguish one component (or a part of a component) from another component (or another part of a component).

[0045] In the description of the present application, unless otherwise specified, “plurality” means more than two (including two), and similarly, “plurality groups” means more than two (including two).

[0046] This embodiment collects the driver's muscle electrical signals and vehicle driving status information, integrates the "human-vehicle" interaction information represented by multiple sensors, takes the driver's braking intensity within the next 200 milliseconds as the driver's braking intention, and establishes a mapping mechanism from the "human-vehicle" interaction multi-sensor information to the driver's braking intention based on the Transformer algorithm.

[0047] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0048] Figure 1 is a flow chart of a model training method based on neural network search provided by an embodiment of the present disclosure, such as Figure 1 As shown in FIG. 1 , the specific process of the model training method based on neural network search includes:

[0049] S110: Obtain model training samples.

[0050] The model training samples include: electromyographic signals corresponding to the driving user of the smart vehicle and the driving status information of the smart vehicle.

[0051] EMG is a bioelectric signal generated by muscle contraction, that is, the bioelectric current generated by the contraction of the surface muscles of the human body. Different muscle fiber motor units on the surface skin generate different signals at the same time. It is generated in any tissue or organ and is usually a time and a series of amplitude, frequency and waveform functions.

[0052] The driving status information of the smart vehicle may include but is not limited to: vehicle speed, acceleration, pedal opening, steering wheel speed and accelerator pedal force.

[0053] S120. Establish a probability distribution model corresponding to the parameter search space.

[0054] Among them, the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters, each group of model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure.

[0055] See also Figure 2A As shown, this embodiment can define the parameter search space according to prior knowledge, and use the sequential model-based optimization via tree-parzen estimator (SMBO-TPE) to optimize the model structure. To optimize, is the parameter search space of the model structure, The performance of is estimated by the evaluation function, and the result obtained by the estimation determines the model structure to be sampled in the model search space in the next step.

[0056] The number of encoder modules stacked has a great impact on the performance of Transformer. Generally, the more encoder layers, the stronger the model's ability to extract features and learn high-level semantics. However, too many encoder stacking times will result in low computational efficiency of the model. Therefore, in 5 to 8 encoder modules, residual connections ( identity ) to determine whether these modules should be retained or discarded.

[0057] The parameter data and model structure of the parameter search space are shown in Table 1.

[0058] Table 1 Search space and optimal model structure

[0059]

[0060] In Table 1, represents the dimension of the latent space, is the number of heads of the multi-head attention mechanism, is the ratio of the dimensions of the multi-head attention mechanism layer and the feed-forward layer. T indicates the presence of residual connection, and F indicates the absence of residual connection.

[0061] In some embodiments, before establishing the probability distribution model corresponding to the parameter search space, the method further includes:

[0062] Obtain historical parameter search information, the historical parameter search information including: model structure hyperparameters corresponding to the historical model structure, the historical model structure corresponding to the candidate model structure; determine the model loss value corresponding to the historical model structure; and construct a parameter search space based on the model structure hyperparameters corresponding to the historical model structure and the model loss value corresponding to the historical model structure.

[0063] Among them, the constructed parameter search space is , is the model structure hyperparameter corresponding to the historical model structure, is the model loss value corresponding to the historical model structure.

[0064] S130: Determine a target model structure based on a preset proxy model and a probability distribution model.

[0065] Among them, the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure, and the performance probability distribution value corresponding to each candidate model structure can be determined according to the corresponding probability distribution evaluation function.

[0066] The model performance is defined by the loss function Estimation: Neural network architecture search aims to find a specific hyperparameter in the parameter search space To realize the loss function Minimize the hyperparameters The expression of is shown in formula (1).

[0067] (1);

[0068] In formula (1), is the parameter search space for candidate model structures, with an order of magnitude of about 1.4×10 15 , The model structure hyperparameters sampled in the parameter search space for each search, is the optimal model hyperparameter.

[0069] Due to the complexity of Transformer, the search space for model structure is huge. Therefore, when searching for the optimal Transformer model structure, the efficiency of the search operation must be considered. The amount of calculation in the search process is large, and the SMBO algorithm is used to reduce the amount of calculation, that is, to establish a proxy model by observing historical search results to fit .

[0070] In some embodiments, the preset proxy model is determined based on the conditional probability of model performance; if the performance probability distribution value of the candidate model structure is less than or equal to the preset threshold, the model performance conditional probability is a first value; if the performance probability distribution value of the candidate model structure is greater than the preset threshold, the model performance conditional probability is a second value.

[0071] For example, the default proxy model S The role of is to improve the search efficiency of the model structure. In this embodiment, the expected improvement (EI) function shown in formula (2) is selected as the preset proxy model. S .

[0072] (2);

[0073] In formula (2), For parameters The structure of the model performance l The conditional probability of is shown in formula (3).

[0074] (3);

[0075] In formula (3), is the preset threshold, is the probability distribution of the model structure with better performance, is the probability distribution of the model structures that perform poorly.

[0076] In some embodiments, determining a target model structure based on a preset proxy model and a probability distribution model includes:

[0077] Based on the preset proxy model, the performance probability distribution value corresponding to each candidate model structure is obtained from the probability distribution model; among the performance probability distribution values ​​corresponding to all candidate model structures, the candidate model parameters corresponding to the minimum performance probability distribution value are determined as the hyperparameter adjustment reference model; the model structure hyperparameters corresponding to the hyperparameter adjustment reference model are searched from the parameter search space; the initial model structure is optimized based on the model structure hyperparameters corresponding to the hyperparameter adjustment reference model to obtain the target model structure.

[0078] Among them, in establishing the probability distribution model M i-1 After that, you can use the preset proxy model Minimization to determine future model structure .

[0079] Refer to the above formula (3), the fractional digit can be used To determine the threshold ,Right now , formula (3) can be rewritten as shown in formula (4).

[0080] (4);

[0081] Formula (4) can be transformed according to the Bayesian formula to obtain formula (5).

[0082] (5);

[0083] Can get, ,when After confirmation, The value of Therefore, the search strategy of the model structure can also be expressed as shown in formula (6).

[0084] (6);

[0085] S140: Using a preset loss function, determine a model loss value corresponding to the target model structure based on the model training samples.

[0086] Among them, the model output corresponding to the target model structure can be determined through the model training samples, and then based on the preset loss function Calculate the model loss value between the model output (braking intensity) and the predicted output.

[0087] Preset loss function You can choose between Root Mean Square Error (RMSE) and Coefficient of Determination (R 2 ) to evaluate the performance of the proposed model structure. RMSE and R 2 The definition of is shown in formula (7).

[0088] (7);

[0089] In formula (7), and are the true value and predicted value of the driver’s braking intensity respectively; and l s are the average value and length of the predicted driver's braking intensity, respectively.

[0090] S150. Based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current number of iterations are updated until the current number of iterations meets the model iteration condition, thereby obtaining a braking intention prediction model.

[0091] Among them, during the first training, the current iteration number is 0. Based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current iteration number are updated until the current iteration number meets the model iteration condition, and then the braking intention prediction model is obtained, that is, based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current iteration number are updated, and steps S120-S140 are continued until the current iteration number is the set maximum iteration number (i.e., the model iteration condition). At this time, the target model structure under the current iteration number can be used as the braking intention prediction model.

[0092] In some embodiments, based on the model loss value corresponding to the target model structure, updating the parameter search space, the probability distribution model and the current number of iterations includes:

[0093] Obtain the model structure hyperparameters corresponding to the target model structure; update the parameter search space based on the model structure hyperparameters corresponding to the target model structure and the model loss value corresponding to the target model structure; determine the performance probability distribution value corresponding to the target model structure, and update the probability distribution model based on the performance probability distribution value corresponding to the target model structure; increase the current number of iterations by 1.

[0094] Among them, by adding the model structure hyperparameters corresponding to the target model structure and the model loss value corresponding to the target model structure to the parameter search space, adding the performance probability distribution value corresponding to the target model structure to the probability distribution model, and increasing the value of the current number of iterations, the parameter search space, the probability distribution model and the current number of iterations can be updated.

[0095] In this embodiment, a model training sample is obtained, and the model training sample includes: the electromyographic signal corresponding to the driving user of the intelligent vehicle and the driving state information of the intelligent vehicle; a probability distribution model corresponding to the parameter search space is established, and the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters, and each group of model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure; based on the preset proxy model and the probability distribution model, the target model structure is determined, and the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure; a preset loss function is used to determine the model loss value corresponding to the target model structure based on the model training sample; based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current iteration number are updated until the current iteration number meets the model iteration condition, and the braking intention prediction model is obtained. In this way, by adaptively searching and determining the model structure, the time cost in the model design process is greatly reduced, the stability of the model structure is guaranteed, and the model prediction efficiency is effectively improved.

[0096] Figure 3 is a flow chart of a method for predicting a driving braking intention provided by an embodiment of the present disclosure, such as Figure 3 As shown, the specific process of the method for predicting the driving braking intention includes:

[0097] S310: Acquire vehicle driving data.

[0098] The vehicle driving data includes: the current electromyographic signal of the vehicle driving user and the current driving status information of the vehicle.

[0099] The vehicle's current driving state information may include, but is not limited to: the vehicle's current driving speed, current acceleration, current pedal opening, current steering wheel speed, and current accelerator pedal force.

[0100] S310 inputs the vehicle driving data into a braking intention prediction model, and determines the braking prediction intensity based on the output result of the braking intention prediction model.

[0101] The braking intention prediction model is the braking intention prediction model trained and generated by the method in the above embodiment, so as to accurately and quickly predict the braking intention of the intelligent vehicle.

[0102] The predicted braking intensity may include but is not limited to: level 1 intensity, level 2 intensity, and level 3 intensity. The intensity value of level 1 intensity is smaller than the intensity value of level 2 intensity, and the intensity value of level 2 intensity is smaller than the intensity value of level 3 intensity.

[0103] In addition, this embodiment performs a training example of a braking intention prediction model, such as Figure 2B and Figure 2C As shown in the figure, the input of the braking intention prediction model includes the driver's leg muscle electrical signals, vehicle acceleration, brake pedal force, brake master cylinder pressure and speed, and the output of the braking intention prediction model is the driver's braking intensity in the next 200 ms. The training set and validation set are established in a ratio of 8:2, and the neural network structure search process consists of pre-training and fine-tuning stages.

[0104] Considering the accuracy and efficiency of the Transformer model, this embodiment designs a loss function for the model structure search process and uses 20% of the total data to build a test set. After each search, the NAS algorithm evaluates the Transformer model under the structure on the test set. The loss function is shown in the following formula (8).

[0105] (8);

[0106] In formula (8), Represents the average time required to execute each epoch. Represents the RMSE of the model on the test set.

[0107] The score value defined in formula (8) is used to consider the accuracy and efficiency of the Transformer model. Figure 2D As shown in Figure 1, after 350 rounds of searching, the model score is 99.384, which is close to 100. The Transformer model structure corresponding to this score is shown in Table 1, which shows that the Transformer model has the best accuracy and efficiency under this structure.

[0108] Figure 2E-2H The figure shows the prediction results of the driver's braking intention corresponding to the peak braking intensity of different drivers within 3~8 MPa. The solid line represents the driver's braking intensity (i.e., the actual braking intensity), and the dotted line represents the driver's braking intention predicted by the algorithm (i.e., the predicted braking intensity). It can be found that the prediction error of the driver's braking intention corresponding to the increase and decrease of the braking intensity is small. When the braking intensity reaches near the peak, the prediction error of the braking intention increases. The peak area of ​​the braking intensity curve usually corresponds to the maximum activation degree of the driver's leg muscles, and the maximum activation degree of the muscles of different drivers cannot be exactly the same, which causes the braking intention prediction algorithm to have a large error in the peak area of ​​the braking intensity. Despite this, the change trend of the braking intention prediction value and the braking intensity remains consistent.

[0109] In addition, this embodiment compares the braking intention prediction performance of different algorithms (as shown in Table 2), and finds that the prediction accuracy of the Transformer model is much higher than that of FNN and LSTM. The Transformer model structure automatically designed using NAS is significantly better than the traditional Transformer model in performance.

[0110] Table 2 Comparison of braking intention prediction performance under different algorithms

[0111]

[0112] In practical applications, there is no need to use NAS to design the Transformer model. The input of the Transformer model is the "human-vehicle" interaction information represented by multiple sensors, and the output is the driver's braking intensity within the next 200 ms. Therefore, the Transformer model obtained based on the neural network structure search strategy can accurately predict the driver's braking intention under different braking intensities.

[0113] Figure 4 A structural schematic diagram of a model training device based on neural network search is provided for this embodiment, wherein the model training device based on neural network search may include: a first acquisition module 410, an establishment module 420, a first determination module 430, a second determination module 440 and an update module 450.

[0114] The first acquisition module 410 is used to acquire model training samples, where the model training samples include: electromyographic signals corresponding to the driving user of the smart vehicle and driving state information of the smart vehicle.

[0115] Establish module 420, which is used to establish a probability distribution model corresponding to the parameter search space. The parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of model structure hyperparameters. Each group of model structure hyperparameters corresponds to a candidate model structure. The probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure.

[0116] The first determination module 430 is used to determine the target model structure based on a preset proxy model and a probability distribution model, where the preset proxy model is used to describe the performance probability distribution value corresponding to each candidate model structure.

[0117] The second determination module 440 is used to determine the model loss value corresponding to the target model structure based on the model training samples by using a preset loss function.

[0118] The updating module 450 is used to update the parameter search space, the probability distribution model and the current iteration number based on the model loss value corresponding to the target model structure, until the current iteration number meets the model iteration condition, thereby obtaining the braking intention prediction model.

[0119] In this embodiment, optionally, the first determining module 430 is specifically configured to:

[0120] Based on the preset proxy model, the performance probability distribution value corresponding to each candidate model structure is obtained from the probability distribution model; among the performance probability distribution values ​​corresponding to all candidate model structures, the candidate model parameters corresponding to the minimum performance probability distribution value are determined as the hyperparameter adjustment reference model; the model structure hyperparameters corresponding to the hyperparameter adjustment reference model are searched from the parameter search space; the initial model structure is optimized based on the model structure hyperparameters corresponding to the hyperparameter adjustment reference model to obtain the target model structure.

[0121] In this embodiment, optionally, it also includes: a third determining module and a constructing module.

[0122] The first acquisition module 410 is further used to acquire historical parameter search information, where the historical parameter search information includes: model structure hyperparameters corresponding to the historical model structure, where the historical model structure corresponds to the candidate model structure.

[0123] The third determination module is used to determine the model loss value corresponding to the historical model structure.

[0124] A building module is used to build a parameter search space based on the model structure hyperparameters corresponding to the historical model structure and the model loss value corresponding to the historical model structure.

[0125] In this embodiment, optionally, the preset proxy model is determined based on the conditional probability of model performance; if the performance probability distribution value of the candidate model structure is less than or equal to the preset threshold, the model performance conditional probability is a first value; if the performance probability distribution value of the candidate model structure is greater than the preset threshold, the model performance conditional probability is a second value.

[0126] In this embodiment, optionally, the updating module 450 is specifically configured to:

[0127] Obtain the model structure hyperparameters corresponding to the target model structure; update the parameter search space based on the model structure hyperparameters corresponding to the target model structure and the model loss value corresponding to the target model structure; determine the performance probability distribution value corresponding to the target model structure, and update the probability distribution model based on the performance probability distribution value corresponding to the target model structure; increase the current number of iterations by 1.

[0128] The model training device based on neural network search provided in the present disclosure can execute the above method embodiments. Its specific implementation principles and technical effects can be found in the above method embodiments, and the present disclosure will not repeat them here.

[0129] Figure 5 This is a schematic diagram of the structure of a device for predicting a driving braking intention provided in this embodiment, wherein the device for predicting a driving braking intention may include: a second acquisition module 510 and an input module 520 .

[0130] The second acquisition module 510 is used to acquire vehicle driving data, which includes: the current electromyographic signal of the vehicle driving user and the current driving state information of the vehicle,

[0131] The input module 520 is used to input the vehicle driving data into the braking intention prediction model, and determine the braking prediction intensity based on the output result of the braking intention prediction model.

[0132] Among them, the braking intention prediction model is a braking intention prediction model trained and generated by the method in the above embodiment.

[0133] The device for predicting driving braking intention provided by the present disclosure can execute the above method embodiments. Its specific implementation principles and technical effects can be found in the above method embodiments, and the present disclosure will not repeat them here.

[0134] The present application also provides a computer device. Figure 6 , Figure 6 This is a basic structural block diagram of the computer device in this embodiment.

[0135] The computer device includes a memory 610 and a processor 620 that are connected to each other through a system bus. It should be noted that the figure only shows a computer device with a memory 610 and a processor 620, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0136] Computer devices can be computing devices such as desktop computers, notebooks, PDAs, and cloud servers. Computer devices can interact with users through keyboards, mice, remote controls, touch pads, or voice control devices.

[0137] The memory 610 includes at least one type of readable storage medium, and the readable storage medium includes a non-volatile memory or a volatile memory, such as a flash memory, a hard disk, a multimedia card, a card-type memory (such as an SD or DX memory, etc.), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc., and the RAM may include a static RAM or a dynamic RAM. In some embodiments, the memory 610 may be an internal storage unit of a computer device, such as a hard disk or a memory of the computer device. In other embodiments, the memory 610 may also be an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital (SD) card, or a flash card (Flash Card), etc., equipped on the computer device. Of course, the memory 610 may also include both an internal storage unit of the computer device and an external storage device thereof. In this embodiment, the memory 610 is generally used to store an operating system and various application software installed on the computer device, such as the program code of the above method, etc. In addition, the memory 610 may also be used to temporarily store various data that have been output or are to be output.

[0138] The processor 620 is generally used to perform the overall operation of the computer device. In this embodiment, the memory 610 is used to store program codes or instructions, the program code includes computer operation instructions, and the processor 620 is used to execute the program codes or instructions stored in the memory 610 or process data, such as running the program code of the above method.

[0139] In this article, the bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0140] Another embodiment of the present application also provides a computer-readable medium, which may be a computer-readable signal medium or a computer-readable medium. A processor in a computer reads a computer-readable program code stored in the computer-readable medium, so that the processor can execute the functional actions specified in each step or a combination of steps in the above method; and generate a device for implementing the functional actions specified in each block or a combination of blocks in the block diagram.

[0141] Computer-readable media include but are not limited to electronic, magnetic, optical, electromagnetic, infrared memory or semiconductor systems, devices or apparatuses, or any appropriate combination of the foregoing, the memory is used to store program codes or instructions, the program codes include computer operating instructions, and the processor is used to execute the program codes or instructions of the above methods stored in the memory.

[0142] For the definitions of memory and processor, please refer to the description of the aforementioned computer device embodiment and will not be repeated here.

[0143] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0144] Each functional unit or module in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0145] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., and other media that can store program codes.

[0146] In the claims, any reference symbols placed between brackets shall not be construed as limiting the claims. The word "comprising" described in the present application does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application can be implemented with the aid of hardware comprising several different elements and with the aid of a suitably programmed computer. In a unit claim that lists a number of devices, several units of these devices may be embodied by the same hardware item. The use of first, second, and third, etc. does not indicate any order, and these words may be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be understood as limitations on the order of execution.

[0147] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A model training method based on neural network search, characterized in that: include: Acquire a model training sample, wherein the model training sample includes: an electromyographic signal corresponding to a driving user of the smart vehicle and driving state information of the smart vehicle; Establishing a probability distribution model corresponding to a parameter search space, wherein the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of the model structure hyperparameters, each group of the model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure; Based on the preset proxy model and the probability distribution model, the target model structure is determined, and the preset proxy model is determined based on the model performance conditional probability; the target model structure is determined based on the preset proxy model and the probability distribution model, including: based on the preset proxy model, obtaining the performance probability distribution value corresponding to each of the candidate model structures from the probability distribution model; determining the candidate model parameter corresponding to the minimum performance probability distribution value among the performance probability distribution values ​​corresponding to all the candidate model structures as the hyper-parameter adjustment reference model; searching for the model structure hyper-parameters corresponding to the hyper-parameter adjustment reference model from the parameter search space; optimizing the parameters of the initial model structure based on the model structure hyper-parameters corresponding to the hyper-parameter adjustment reference model to obtain the target model structure; Using a preset loss function, based on the model training samples, determine a model loss value corresponding to the target model structure; Based on the model loss value corresponding to the target model structure, the parameter search space, the probability distribution model and the current number of iterations are updated until the current number of iterations meets the model iteration condition, thereby obtaining a braking intention prediction model.

2. The method according to claim 1, characterized in that Before establishing the probability distribution model corresponding to the parameter search space, the method further includes: Acquire historical parameter search information, wherein the historical parameter search information includes: model structure hyperparameters corresponding to a historical model structure, wherein the historical model structure corresponds to the candidate model structure; Determining a model loss value corresponding to the historical model structure; The parameter search space is constructed based on the model structure hyperparameters corresponding to the historical model structure and the model loss value corresponding to the historical model structure.

3. The method according to claim 1, characterized in that If the performance probability distribution value of the candidate model structure is less than or equal to a preset threshold, the model performance conditional probability is a first value; if the performance probability distribution value of the candidate model structure is greater than the preset threshold, the model performance conditional probability is a second value.

4. The method according to claim 1, characterized in that: The updating of the parameter search space, the probability distribution model and the current number of iterations based on the model loss value corresponding to the target model structure includes: Obtaining model structure hyperparameters corresponding to the target model structure; Based on the model structure hyperparameters corresponding to the target model structure and the model loss value corresponding to the target model structure, updating the parameter search space; Determine a performance probability distribution value corresponding to the target model structure, and update the probability distribution model based on the performance probability distribution value corresponding to the target model structure; The current iteration number is increased by 1.

5. A method for predicting a driver's braking intention, characterized in that: include: Acquiring vehicle driving data, wherein the vehicle driving data includes: current electromyographic signals of the vehicle driving user and current driving state information of the vehicle; Inputting the vehicle driving data into a braking intention prediction model, and determining a braking prediction intensity based on an output result of the braking intention prediction model; Wherein, the braking intention prediction model is a braking intention prediction model trained and generated by the method described in any one of claims 1-4.

6. A model training device based on neural network search, characterized in that: include: A first acquisition module is used to acquire a model training sample, wherein the model training sample includes: an electromyographic signal corresponding to a driving user of the smart vehicle and driving state information of the smart vehicle; An establishment module is used to establish a probability distribution model corresponding to a parameter search space, wherein the parameter search space includes: multiple groups of model structure hyperparameters and model loss values ​​corresponding to each group of the model structure hyperparameters, each group of the model structure hyperparameters corresponds to a candidate model structure, and the probability distribution model is used to describe the performance probability distribution value corresponding to each candidate model structure; A first determination module is used to determine the target model structure based on a preset proxy model and the probability distribution model, wherein the preset proxy model is determined based on the model performance conditional probability; the first determination module is specifically used to: obtain the performance probability distribution value corresponding to each of the candidate model structures from the probability distribution model based on the preset proxy model; determine that the candidate model parameter corresponding to the minimum performance probability distribution value among the performance probability distribution values ​​corresponding to all the candidate model structures is the hyperparameter adjustment reference model; search for the model structure hyperparameter corresponding to the hyperparameter adjustment reference model from the parameter search space; optimize the parameters of the initial model structure based on the model structure hyperparameter corresponding to the hyperparameter adjustment reference model to obtain the target model structure; A second determination module is used to determine a model loss value corresponding to the target model structure based on the model training sample by using a preset loss function; An updating module is used to update the parameter search space, the probability distribution model and the current number of iterations based on the model loss value corresponding to the target model structure, until the current number of iterations meets the model iteration condition, thereby obtaining a braking intention prediction model.

7. The device according to claim 6, characterized in that If the performance probability distribution value of the candidate model structure is less than or equal to a preset threshold, the model performance conditional probability is a first value; if the performance probability distribution value of the candidate model structure is greater than the preset threshold, the model performance conditional probability is a second value.

8. A device for predicting a driver's braking intention, characterized in that: include: A second acquisition module is used to acquire vehicle driving data, wherein the vehicle driving data includes: a current electromyographic signal of a vehicle driving user and current driving state information of the vehicle; An input module, used for inputting the vehicle driving data into a braking intention prediction model, and determining a braking prediction intensity based on an output result of the braking intention prediction model; Wherein, the braking intention prediction model is a braking intention prediction model trained and generated by the method described in any one of claims 1-4.

Images