Dangerous driving behavior identification method and system based on hybrid deep neural network

Abstract

The invention discloses a dangerous driving behavior identification method and system based on a hybrid deep neural network, and belongs to the technical field of intelligent driving. The method comprises the following steps: S1, building a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gating circulation unit; the hyper-parameters of the GA-CNN-BiGRU hybrid model are optimized by using a genetic algorithm; s2, constructing a data set, and carrying out the training of the GA-CNN-BiGRU hybrid model; s3, acquiring driving data in the vehicle driving process, wherein the driving data comprise speed, acceleration and steering angle; the driving data are converted into speed time-history curve data, acceleration time-history curve data and a direction angle time-history curve; and S4, inputting the driving data into the trained GA-CNN-BiGRU hybrid model to carry out dangerous driving behavior identification. According to the method, multiple advanced technical means are integrated, the integrated process from feature extraction to behavior recognition is achieved, and compared with an existing method, the method shows higher accuracy and reliability in recognition of abnormal driving behaviors.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent driving technology, and in particular relates to a method and system for identifying dangerous driving behaviors based on a hybrid deep neural network. Background Art

[0002] During driving, the driver's attention is mainly focused on maintaining the normal driving and safety of the vehicle, and any activities unrelated to driving (such as using a mobile phone, making phone calls, operating the vehicle control panel, drinking water, or talking to people in the car, etc.) will cause his attention to be distracted. Although these non-main activities do not seem to affect driving in the short term, in emergencies, the driver often cannot quickly return from these activities to pay attention to the road and traffic environment, thereby increasing the risk of traffic accidents. For example, when making a phone call, the driver's line of sight may leave the road ahead, resulting in the inability to detect potential traffic obstacles or traffic signal changes in time, thereby increasing the possibility of a collision.

[0003] When drivers are engaged in non-primary activities, it is often difficult for them to make timely judgments and take effective actions before an accident occurs, which directly leads to dangerous driving behaviors such as sudden braking, speeding or out-of-control operations. To effectively deal with these problems, modern intelligent traffic management systems can monitor drivers' behaviors in real time and identify possible dangerous behaviors by integrating various advanced technologies, such as machine learning algorithms. It can not only analyze whether the driver has dangerous driving behaviors such as sudden braking, speeding or fatigue driving, but also send real-time warning signals to the driver in combination with the vehicle's operating status and environmental parameters to remind him to correct improper behaviors in a timely manner. For example, when the system detects that the driver is overly distracted or inattentive, it can remind the driver to regain concentration through sound warnings or vibrations, or automatically take braking or other safety measures in an emergency to reduce the probability of accidents. Therefore, the application of intelligent driving assistance technology not only improves traffic safety, but also alleviates the problem of driver distraction to a certain extent, providing strong support for more efficient and safe traffic management.

[0004] With the rapid development of intelligent transportation systems and sensor technology, driving data from various sensors installed inside vehicles, such as location data, status data, and video data, can be used to detect abnormal driving behavior. Abnormal vehicle driving behavior can be roughly divided into two categories. One is video-based abnormal driving behavior recognition, and its recognition technology mainly relies on image processing and computer vision technology. This type of technology identifies abnormal driver behavior, such as speeding, running red lights, and driving in the wrong lane, by analyzing the video data of the vehicle in real time. By analyzing the image sequence in the video stream, the vehicle's motion characteristics, the driver's behavior pattern, and the changes in the traffic environment can be extracted to determine whether there is abnormal driving behavior. The advantage of this method is that it can intuitively reflect the driver's behavior and traffic conditions, which is easy to understand and explain, but it is also limited by factors such as video quality, lighting conditions, and viewing angle.

[0005] The other is the abnormal driving behavior recognition technology based on sensor data. Compared with the video-based method, the abnormal driving behavior based on sensor data can effectively help analyze the vehicle driving process and has advantages in identifying abnormal driving. This type of technology mainly collects data through the vehicle's built-in sensors, such as speed sensors, acceleration sensors, GPS, etc., to monitor the vehicle's operating status and the driver's operating behavior. By analyzing the sensor data, the vehicle's abnormal acceleration, sudden braking, sharp turns and other behaviors can be identified, thereby determining whether the driver has abnormal driving behavior. The advantage of this method lies in the accuracy and real-time nature of the data, which can provide more accurate vehicle status information.

[0006] Sensor data can effectively help analyze the vehicle driving process and has advantages in identifying abnormal driving. However, the current sensor data-based methods mainly use machine learning, such as support vector machines, random forests, linear regression algorithms, etc., and their recognition accuracy is low. Compared with machine learning, deep learning technology has shown significant performance in processing large-scale data sets, especially without the need for artificial feature engineering. At present, the application of deep learning methods in the field of abnormal driving behavior recognition based on state data (such as speed, acceleration, and deceleration) is still relatively rare.

[0007] Prior art CN201710232371.6 discloses a driving behavior recognition method based on deep sparse filtering convolutional neural network. It solves the problem that a large amount of hyperparameter tuning is required for feature matching or feature design tasks in traditional deep learning methods. It includes the following steps: using the built-in acceleration sensor of the mobile phone to collect the original three-axis acceleration data signal of the vehicle during driving, extracting the eigenvalues ​​of the time domain and frequency domain as the pre-processing training samples of the model; directly optimizing the sparsity of the sample feature mapping through sparse filtering of the training samples to obtain better feature expression - the weight matrix is ​​used as the input of the convolutional neural network, which can effectively identify driving behaviors such as ignition, shutdown, constant speed driving, sudden speed change, sharp turn, and stationary.

[0008] The above-mentioned existing technology uses deep sparse filtering convolutional neural network to directly optimize the sparsity of sample feature mapping through sparse filtering to obtain better feature expression, which simplifies the tedious hyperparameter tuning task in traditional deep learning methods to a certain extent. However, the overall structure of the model is relatively simple and lacks the ability to model deep time series data. For driving behaviors with obvious time-dependent characteristics, such as continuous rapid acceleration and deceleration, it may not be able to accurately grasp the development trend of the behavior, thereby affecting the recognition accuracy. Summary of the invention

[0009] The purpose of the present invention is to provide a method and system for identifying dangerous driving behaviors based on a hybrid deep neural network, which can partially solve or alleviate the above-mentioned deficiencies in the prior art and can efficiently and accurately identify dangerous driving behaviors through sensor data.

[0010] In order to solve the above-mentioned technical problems, the present invention specifically adopts the following technical solutions: A first aspect of the present invention is to provide a method for identifying dangerous driving behaviors based on a hybrid deep neural network, comprising: S1 builds a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm; S2 constructs a data set to train the GA-CNN-BiGRU hybrid model; S3 collects driving data of the vehicle during driving, including speed, acceleration and steering angle; and converts the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve; S4 inputs the driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors; The GA-CNN-BiGRU hybrid model includes: Convolutional layer, used to extract spatial features from driving data and output feature maps; The pooling layer is used to reduce the dimension of the feature map output by the convolutional layer; A bidirectional gated recurrent component is used to obtain hidden states in feature maps from the forward and reverse directions of the time series and fuse them into time series features; The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features; Softmax function, used for classification prediction based on spatiotemporal features.

[0011] As an improvement, the steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model using a genetic algorithm include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration; S102 builds the corresponding GA-CNN-BiGRU hybrid model based on chromosomes; S103 evaluates the fitness of the GA-CNN-BiGRU hybrid model; S104 screens chromosomes in the population based on fitness; S105 performs crossover mutation on the screened chromosomes to form a new population; S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

[0012] As an improvement, the hyperparameters of the GA-CNN-BiGRU hybrid model optimized by genetic algorithm include: the number of convolutional layers, kernel size, number of channels, and activation function; the type and number of pooling layers; the number of layers, number of hidden units, and activation function of the bidirectional gated recurrent component.

[0013] As an improvement, the step of extracting spatial features from driving data by the convolutional layer includes: The convolution kernel S201 slides on the velocity time history curve data, the acceleration time history curve data and the direction angle time history curve according to the set step size, and obtains the local eigenvalue by performing convolution operation with the local data points; S202 concatenates the obtained local eigenvalues ​​to obtain a feature map.

[0014] As an improvement, the step of obtaining the timing characteristics of the bidirectional gated loop component includes: S301 forward gated recurrent unit combines the hidden state at the previous moment with the current information to output the positive hidden state at the current moment; S302 The backward gated recurrent unit combines the hidden state at the next moment with the current information to output the reverse hidden state at the current moment; S303 merges the current forward hidden state and the reverse hidden state into the current hidden state.

[0015] As an improvement, the number of convolutional layers is 1 to 30, the size of the convolution kernel in each convolutional layer is any one of 3, 5, 7, and 9, the number of channels in each convolutional layer is any one of 32, 64, 128, and 256, and the activation function is any one of ReLU, LeakyReLU, and ELU; the type of the pooling layer is max or average, and the number of pooling layers is 2 or 3; the number of layers of the bidirectional gated recurrent component is 1 to 5, the number of hidden units is any one of 64, 128, 256, and 512, and the activation function is any one of ReLU, LeakyReLU, and ELU.

[0016] As an improvement, the GA-CNN-BiGRU hybrid model is trained using speed time history curves, acceleration time history curves and direction angle time history curves of labeled driving behaviors as samples; the driving behaviors include normal driving, sudden acceleration, sudden deceleration, abnormal left lane change, and abnormal right lane change.

[0017] As an improvement, the GA-CNN-BiGRU hybrid model also includes: The input layer before the convolutional layer is used to preprocess the input data; The output layer after the Softmax function is used to output the results of the Softmax function according to the corresponding driving behavior.

[0018] The present invention also provides a dangerous driving behavior recognition system based on a hybrid deep neural network, comprising: A model building module, used to build a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm; A training module, used for constructing a data set to train the GA-CNN-BiGRU hybrid model; The data acquisition module is used to collect driving data of the vehicle during driving, including speed, acceleration and steering angle; and convert the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve; The recognition module is used to input driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors; The GA-CNN-BiGRU hybrid model includes: Convolutional layer, used to extract spatial features from driving data and output feature maps; The pooling layer is used to reduce the dimension of the feature map output by the convolutional layer; A bidirectional gated recurrent component is used to obtain hidden states in feature maps from the forward and reverse directions of the time series and fuse them into time series features; The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features; Softmax function, used for classification prediction based on spatiotemporal features.

[0019] As an improvement, the steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model by the model building module include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration; S102 builds the corresponding GA-CNN-BiGRU hybrid model based on chromosomes; S103 evaluates the fitness of the GA-CNN-BiGRU hybrid model; S104 screens chromosomes in the population based on fitness; S105 performs crossover mutation on the screened chromosomes to form a new population; S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

[0020] Beneficial effects: The GA-CNN-BiGRU hybrid model proposed in the present invention aims to significantly improve the recognition accuracy of abnormal driving behavior. The model combines the advantages of Genetic Algorithm (GA), Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit (BiGRU) to achieve more accurate and efficient abnormal driving behavior detection. First, the genetic algorithm is used to optimize the parameter settings of the CNN and BiGRU models. By simulating natural selection and genetic mechanisms, it can effectively search the optimal solution space, thereby providing the best initialization parameters for the subsequent deep learning process, which helps to improve the model training efficiency and final performance. Then, CNN is used to automatically extract spatiotemporal features from vehicle driving data. These features include but are not limited to indicators closely related to driving behavior, such as speed, acceleration, and steering angle. The convolution layer can capture local features, while the pooling layer can reduce the data dimension and retain important information, so that the model can fully explore the deep structure in the input data while maintaining high computational efficiency. Finally, the BiGRU component is responsible for processing the time series features extracted by CNN. BiGRU not only considers the past information of the time series, but also utilizes future information (in this scenario, it refers to the data at relatively subsequent moments), which is particularly useful for understanding complex driving behavior patterns. It can better model long-term dependencies and alleviate the problem of gradient disappearance, thereby improving the ability to predict abnormal events.

[0021] Therefore, this study proposes a GA-CNN-BiGRU model that integrates a variety of advanced technical means to achieve an integrated process from feature extraction to behavior recognition. Compared with existing methods, it shows higher accuracy and reliability in identifying abnormal driving behaviors. This achievement is of great significance to the development of traffic safety monitoring systems, autonomous driving technology, and the improvement of road safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale. Obviously, the drawings described below are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without paying creative labor.

[0023] Figure 1 This is the architecture diagram of the GA-CNN-BiGRU hybrid model; Figure 2 Schematic diagram of input data for the GA-CNN-BiGRU hybrid model; Figure 3 The process of dividing the dataset for training the model. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0025] Herein, suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present invention, and have no specific meanings by themselves. Therefore, "module", "component" or "unit" can be used mixedly.

[0026] In this document, the terms "upper", "lower", "inner", "outer", "front", "back", "one end", "the other end" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0027] In this document, unless otherwise clearly specified and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0028] Herein "and / or" includes any and all combinations of one or more of the associated listed items.

[0029] Herein, "plurality" means two or more than two, ie, it includes two, three, four, five, etc.

[0030] Embodiment 1: The driver's driving behavior can be divided into longitudinal behavior and lateral behavior. The longitudinal direction is the vehicle's forward direction, including sudden acceleration and sudden braking, while abnormal driving behavior includes rapid acceleration and emergency braking. Longitudinal driving behavior mainly includes left and right lane changes (overtaking behavior can be regarded as double lane changes). When the driver ignores the surrounding vehicles, rapid lane changes can lead to collision accidents. In this process, when the driver performs driving behaviors such as sudden acceleration, sudden braking, rapid right lane changes, and rapid changes, abnormal road section data can be stored by detecting the curve changes of factors such as acceleration, deceleration, braking status, and steering wheel angle. Four abnormal driving modes (sudden acceleration, sudden deceleration, abnormal left lane change, abnormal right lane change) + one normal driving behavior, a total of five driving behaviors, can represent all behaviors in the general driving process.

[0031] In order to achieve accurate classification and detection of the above driving behaviors, the present invention provides a dangerous driving behavior identification method based on a hybrid deep neural network, and the specific steps include: S1 builds a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm.

[0032] Specifically, Figure 1 As shown, the GA-CNN-BiGRU hybrid model in this embodiment includes: 1. Input layer, used to preprocess the input data.

[0033] Ensure that the format of the input data meets the processing requirements of the subsequent components of the model. Organize the multi-dimensional driving behavior data into the tensor form required for internal calculations in the model, so that the data can be smoothly transferred and calculated between subsequent convolutional layers, BiGRU and other components, laying the foundation for feature extraction and analysis, and ensuring that the various parts of the model can work together efficiently.

[0034] The input layer can clean and denoise the data. The raw data collected from various sensors of the vehicle are often mixed with a lot of interference information. The input layer will first clean and denoise it. For example, the acceleration sensor may produce small and frequent fluctuation noise due to factors such as the vehicle's own vibration and road bumps. If these noise data are not eliminated, they will interfere with the subsequent extraction of the acceleration characteristics of the real driving behavior. The input layer sets reasonable thresholds to identify and remove outliers that are obviously deviated from the normal data range and caused by noise, ensuring that the input data reflects the driving behavior as realistically as possible. At the same time, for missing data points that occasionally appear due to sensor failures, the input layer will reasonably interpolate and fill them based on the surrounding valid data points to ensure the continuity of the data and provide a reliable basis for subsequent precise analysis.

[0035] The input layer can also integrate time series. Driving behavior data essentially has time series characteristics, and the input layer is responsible for accurately integrating scattered sensor data in chronological order. Taking acceleration, speed, and angular data as an example, they come from different sensors, and the collection frequency may be slightly different. The input layer will pair and align the different dimensional data collected at the same time or similar time based on the timestamp information to construct a complete time series sample. This ensures that the subsequent model can accurately capture the dynamic changes of driving behavior from a coherent time dimension when processing data. For example, when analyzing vehicle lane change behavior, it can accurately judge the start, process, and end stages of lane change based on the continuous speed and angular change sequence, thereby improving the model's ability to understand complex driving scenarios.

[0036] 2. Convolutional layer, used to extract spatial features from driving data and output feature maps.

[0037] Convolutional layers are generally used for image processing to extract meaningful features from two-dimensional images. It consists of multiple convolution kernels, which are small weight matrices whose shapes and parameter values ​​are continuously learned and optimized during the training process. In essence, the convolutional layer is like an intelligent filter that can scan the input data and capture the hidden local patterns. Taking the field of image recognition as an example, convolution kernels can identify features such as edges and textures in images.

[0038] In the driving behavior recognition scenario used by the GA-CNN-BiGRU hybrid model in this embodiment, the driving data (including speed, acceleration and steering angle) needs to be converted into speed time curve data, acceleration time curve data and direction angle time curve before inputting into the model, and then the convolution layer is used to extract the features in the above curves. The convolution kernel can focus on the key change information in driving monitoring indicators such as speed, acceleration, and direction angle, and convert the original data into a feature representation that is more valuable for subsequent judgment.

[0039] In this embodiment, the convolution layer is mainly used for feature extraction of spatial features, that is, to mine key spatial features from complex driving behavior data. Information such as speed changes, acceleration mutations, and direction adjustments during vehicle driving are all contained in the original monitoring data. The convolution layer slides the convolution kernel on these data to accurately capture local features such as the acceleration curve characteristics of the vehicle acceleration phase and the changing trend of the direction angle when turning, providing basic elements for a comprehensive understanding of driving behavior.

[0040] Specifically, the steps of extracting spatial features from driving data by the convolutional layer include: The S201 convolution kernel slides on the velocity time history curve data, acceleration time history curve data and direction angle time history curve data according to the set step size, and obtains the local eigenvalue by performing convolution operation with the local data points.

[0041] The convolution kernel in the convolution layer slides in a specific direction on the input data tensor according to the set step size. For two-dimensional driving behavior data tensors (such as speed-time, acceleration-time, steering angle-time plane data), the convolution kernel starts from the left and moves to the right with a fixed step size. During the movement, the convolution kernel multiplies the covered data points element by element, and the product results are summed. This sum is the feature value extracted by the convolution kernel at the current position. Different convolution kernels have different weight parameters and can extract different types of features. Small convolution kernels are good at capturing detailed changes, and large convolution kernels can obtain more macroscopic feature patterns.

[0042] S202 concatenates the obtained local eigenvalues ​​to obtain a feature map.

[0043] As the convolution kernel continuously slides through the entire input data, each time a sliding calculation is completed, a feature value is obtained. Many such feature values ​​are arranged according to the sliding track of the convolution kernel to form a feature map. Each convolution layer has multiple convolution kernels, and the feature maps generated by each of them together constitute the output of the layer. These outputs carry various local features extracted from the original data and are passed to the next layer for further processing or enter the subsequent pooling layer.

[0044] 3. Pooling layer, used to reduce the dimension of the feature map output by the convolutional layer.

[0045] The main function of the pooling layer is to reduce the dimension and compress the data. The pooling layer usually contains multiple pooling methods, among which the most common are maximum pooling and average pooling. Maximum pooling is to select the maximum value in a given local area as the representative feature value of the area; average pooling is to calculate the average value of all values ​​in this local area to represent the characteristics of the area. Regardless of the pooling method, the object of their operation is the feature map output by the convolution layer. Through a specific window sliding on the feature map, key information is extracted according to established rules to achieve the purpose of simplifying the data. Taking the image field as an example, the pooling layer can reduce the number of pixels while retaining the main features of the image, making the image data more compact and conducive to subsequent processing; in the driving behavior recognition scenario, the pooling layer also optimizes the driving behavior feature map extracted by the convolution layer.

[0046] 4. A bidirectional gated recurrent component is used to obtain the hidden states in the feature maps from the forward and reverse directions of the time series and fuse them into time series features.

[0047] Bidirectional Gated Recurrent Unit (BiGRU) is a deep learning architecture based on the improvement of recurrent neural network (RNN). It is a clever combination of forward gated recurrent unit (FGRU) and backward gated recurrent unit (BGRU). These two units each have unique information processing capabilities. The forward unit gradually accumulates past information along the time series; the backward unit traces back from the end of the time series to collect key information at subsequent moments. They both contain update gates and reset gates to accurately control the flow and retention of information, allowing BiGRU to perform well in processing time series data.

[0048] Since traditional convolutional neural networks are mainly used to process two-dimensional images, they lack the ability to grasp temporal features. However, the three waveforms that require feature extraction in the present invention actually have extremely strong temporal features. Therefore, the present invention adds a bidirectional gated loop component to the existing convolutional neural network to extract the temporal features hidden in the feature graph.

[0049] In the entire model architecture, BiGRU is in a key position to connect the previous and the next, and it receives the feature map output from the convolution layer (after processing by the pooling layer). It gives full play to the advantages of bidirectional processing and deeply integrates past and future information. In the driving behavior recognition scenario, the vehicle's previous acceleration and deceleration history and subsequent possible lane changes, turning trends and other information can be effectively integrated by BiGRU, thereby providing a comprehensive feature representation that includes contextual associations, laying a solid foundation for subsequent accurate judgment of driving behavior categories. That is, the hidden state is the compression and encoding of the current information and historical (or future) information in the sequence by the BiGRU neural network.

[0050] Compared with unidirectional recurrent neural networks, BiGRU has more diverse gradient propagation paths during the back propagation process due to the bidirectional information flow. The gradient of the forward unit can assist the backward unit, and vice versa, which effectively breaks the dilemma of gradient attenuation during unidirectional propagation, allowing the model to maintain good learning ability when processing long-term series data, continuously optimize model parameters, and accurately capture the complex changing patterns of driving behavior over time.

[0051] Specifically, the step of obtaining the timing characteristics of the bidirectional gated loop component includes: S301 forward gated recurrent unit combines the hidden state of the previous moment with the current information to output the positive hidden state of the current moment.

[0052] The forward gated recurrent unit processes the input data in the forward order of the time series. When receiving the feature map extracted from the convolutional layer, it receives the input features of the current moment and the hidden state of the previous moment at each time step. Among them, the update gate uses the current input and the hidden state of the previous moment to calculate a value between 0 and 1 through a specific activation function (such as the sigmoid function). This value determines how much hidden state information of the previous moment should be retained; the reset gate also calculates a coefficient based on the current input and the hidden state of the previous moment to determine how to use the current input to reset part of the hidden state information of the previous moment. Based on the fine operation of these two gates, combined with the current input features, a series of mathematical operations (such as matrix multiplication, addition, etc.) are used to calculate the hidden state of the current moment. By gradually advancing in this way, past information from the starting moment to the current moment can be accumulated, clearly reflecting the previous changing trend of driving behavior.

[0053] S302 The backward gated recurrent unit combines the hidden state at the next moment with the current information to output the reverse hidden state at the current moment.

[0054] The backward gated recurrent unit processes the time series in reverse, going back from the time step at the end of the sequence. Similarly, at each time step, it receives the input features of the current moment (although the time step is the same here, the data meaning is different due to the reverse processing) and the hidden state of the next moment. The update gate and reset gate of the backward unit use a similar mechanism to calculate the corresponding retention coefficient and reset coefficient based on the current input and the hidden state of the next moment, and then calculate the hidden state of the current moment, gradually backtracking to collect information at future moments, and knowing the situation of subsequent driving behavior in advance.

[0055] S303 merges the current forward hidden state and the reverse hidden state into the current hidden state.

[0056] At each time step, BiGRU fuses the hidden state obtained by the forward unit with the hidden state obtained by the backward unit. Common fusion methods include splicing (such as splicing two hidden states together along a certain dimension) or summing. Through this two-way processing and fusion, the BiGRU component not only integrates the impact of past driving behavior on the current moment, but also considers the reference provided by future situations for current judgment, so that it can accurately model long-term dependencies, deeply analyze the continuous changes of driving behavior over time, and effectively improve the ability to recognize complex driving behavior patterns.

[0057] 5. The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features.

[0058] The fully connected layer is composed of a large number of neurons, which present a fully connected topology, that is, each neuron is connected to all neurons in the previous layer, and each element in the weight matrix corresponds to the weight value of the connection. It can receive and process feature information from different dimensions and levels, and unify and standardize complex and diverse input data.

[0059] In this embodiment, the fully connected layer receives features output from the BiGRU component, which contain local spatial features extracted by the convolutional layer and temporal features processed by the BiGRU component. The fully connected layer integrates these high-dimensional, abstract features through a fully connected weight matrix between a large number of neurons, summarizes the feature information extracted from different levels and dimensions, and converts it into a feature vector of unified dimension. For example, if the BiGRU component outputs multiple temporal feature dimensions related to driving behavior, the fully connected layer can regularize them into a vector of fixed length, so that subsequent operations can be performed based on this standardized feature expression.

[0060] 6. Softmax function, used for classification prediction based on spatiotemporal features.

[0061] The Softmax function is often used in the output layer of multi-classification problems. Its core function is to convert a vector containing any real numbers into a probability distribution over each category. From a mathematical point of view, for the input vector, the Softmax function calculates each element to obtain the probability value of the corresponding category after conversion. This allows the model output results to be presented intuitively in the form of probability, clearly indicating the probability of each category being selected, which is convenient for subsequent classification judgment.

[0062] In the driving behavior recognition scenario, the GA-CNN-BiGRU hybrid model is processed by the previous convolutional layer and BiGRU components, and the fully connected layer outputs a vector representing the strength of different driving behavior features, but this does not directly correspond to a specific category. After the Softmax function comes on stage, this vector can be converted into a probability distribution of various driving behaviors such as normal driving, sudden acceleration, and abnormal lane changes. For example, the output may be the probabilities corresponding to the above three types of behaviors, allowing the model to clearly give the most likely category judgment of the current driving behavior, providing a key basis for subsequent decision-making.

[0063] Since driving behaviors have many complex manifestations, the model needs to accurately distinguish many different behavior categories. Softmax is designed specifically for multi-classification tasks. It weighs and compares the many candidate categories based on the feature vectors transmitted by the fully connected layer, and highlights the behavior category that best matches the current driving characteristics, ensuring that the model can accurately identify various potential driving behaviors when facing complex and changing road scenes, effectively improving the accuracy and reliability of classification.

[0064] 7. The output layer is used to output the results of the Softmax function according to the corresponding driving behavior.

[0065] The output layer, as the last layer of the model, is a direct window for information output. It is responsible for converting the information processed and refined by multiple complex components (such as convolutional layers, pooling layers, BiGRU, fully connected layers, etc.) into a meaningful result form that can be directly used by the outside world. Usually, in the application scenario of driving behavior recognition, the output layer will output information representing different driving behavior categories in a specific format. This information is based on the model's in-depth analysis of the input driving monitoring data and is the model's final "judgment conclusion" on the current driving state.

[0066] When the model receives a series of data from the vehicle sensor, after feature extraction, time series analysis and model optimization at each internal layer, the output layer will give a clear classification indication, such as judging whether the current driving behavior belongs to normal driving, sudden acceleration, abnormal right lane change, abnormal left lane change, etc. This result is presented to the user in a clear and easy-to-understand form, whether it is a traffic supervisor viewing abnormal driving behavior records or an autonomous driving system making decisions based on judgments, key information can be quickly obtained.

[0067] In addition, in this embodiment, the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm, specifically including the number of convolutional layers, convolution kernel size, number of channels, and activation functions; pooling layer type and number of layers; the number of layers, number of hidden units, and activation functions of the bidirectional gated loop component. The number of convolutional layers is 1 to 30, the size of the convolution kernel in each convolutional layer is any one of 3, 5, 7, and 9, the number of channels in each convolutional layer is any one of 32, 64, 128, and 256, and the activation function is any one of ReLU, LeakyReLU, and ELU; the type of the pooling layer is max or average, and the number of pooling layers is 2 or 3; the number of layers of the bidirectional gated loop component is 1 to 5, the number of hidden units is any one of 64, 128, 256, and 512, and the activation function is any one of ReLU, LeakyReLU, and ELU.

[0068] Specifically, the steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model using a genetic algorithm include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration.

[0069] First, a batch of chromosomes are randomly generated according to the user-specified search space (for example, the maximum number of convolutional layers is 30, the convolution kernel size can only be selected from {3,5,7,9} and other hyperparameters’ genetic encoding information), and each chromosome corresponds to a feasible convolutional neural network + BiGRU configuration.

[0070] The population size is usually set at 20 to 50. It should not be too small, otherwise the diversity will be insufficient; nor should it be too large, otherwise the training will take too long.

[0071] S102 builds the corresponding GA-CNN-BiGRU hybrid model based on the chromosome.

[0072] Decode each chromosome in the population, that is, build the corresponding neural network structure (configure CNN layer, Bi-GRU layer, etc.) according to the genetic information in the chromosome, and set the specified optimizer parameters (such as learning rate, regularization, etc.).

[0073] Use the training data to train each individual, possibly using full training or "early stopping + shortened epoch" to save time. After training is completed, calculate the fitness of the individual on the validation set or cross-validation set, such as accuracy, F1-score, loss function, etc.

[0074] S103 performs fitness evaluation on the GA-CNN-BiGRU hybrid model.

[0075] The key indicators obtained during or after training (such as validation set accuracy / loss, etc.) are used as the "fitness" of the genetic algorithm. A high fitness means that the CNN + Bi-GRU configuration performs better on the current task.

[0076] S104 screens chromosomes in the population based on fitness.

[0077] The current population is screened based on fitness, with common methods such as "Roulette Wheel" or "Tournament Selection". The goal is to allow chromosomes with high fitness (good performance) to have a greater chance of "reproducing" and thus pass on their excellent genes.

[0078] S105 performs crossover mutation on the screened chromosomes to form a new population.

[0079] Specifically, in the selected excellent individuals, two chromosomes undergo genetic recombination (crossover) to produce new offspring. For each chromosome array, strategies such as single-point crossover, multi-point crossover, or uniform crossover can be used. For example, if each chromosome has 10 genes, one or more positions can be randomly selected for crossover operations. After crossover, the offspring will inherit the gene combination of each part of the parent, and new hyperparameter combinations may appear.

[0080] Perturb some genes to prevent the population from falling into local optimality too early. For discrete hyperparameters (such as convolution kernel size, number of layers, activation function, etc.), they can be randomly replaced with feasible values ​​of the same type; for continuous hyperparameters (such as learning rate, dropout), they can be randomly increased or decreased within a small range. The mutation rate is usually low (such as 1%~5%) to ensure the stability of the search direction.

[0081] S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

[0082] The updated new population repeats the process of steps S102 to S105 again.

[0083] The number of iterations is usually between 20 and 50, or the stopping time is determined based on the time budget and indicator performance (such as when fitness stagnates). When the algorithm terminates, the hyperparameter configuration corresponding to the chromosome with the highest fitness is taken as the final CNN + Bi-GRU structure and training plan.

[0084] S2 constructs a dataset to train the GA-CNN-BiGRU hybrid model.

[0085] like Figure 3 As shown in the figure, during the dataset construction phase, the model first extracts the dynamic changes of driving behaviors from the training data. These key driving monitoring indicators include acceleration curves (acceleration time-history curve data), speed curves (speed time-history curve data), and direction angle time curves (direction angle time-history curves). These behaviors are labeled into different categories, such as sudden acceleration, sudden deceleration, abnormal right lane change, and abnormal left lane change are labeled as abnormal driving behaviors, thereby distinguishing them from normal driving behaviors. Subsequently, the dataset is divided into three parts: training data, verification data, and test data. Through data fusion, a complete input dataset is formed to provide support for subsequent model training and performance evaluation.

[0086] S3 collects driving data of the vehicle during driving, including speed, acceleration and steering angle; and converts the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve data.

[0087] First of all, in order to accurately obtain various key data during vehicle driving, it is necessary to select appropriate sensors. For speed measurement, high-precision wheel speed sensors or GPS speed sensors are usually used. The wheel speed sensor is installed near the wheel hub. By sensing the number of rotations of the wheel and combining parameters such as the wheel circumference, it uses a special conversion algorithm to calculate the vehicle's driving speed in real time. Its advantage is that it can quickly feedback speed changes in a short time with high accuracy; the GPS speed sensor uses satellite positioning technology to measure the position movement of the vehicle within a certain time interval and calculate the speed information. Its advantage is that it is not affected by factors such as wheel slippage and can provide more macro speed data.

[0088] For acceleration measurement, an acceleration sensor is selected, which is generally installed near the center of gravity of the vehicle, such as the middle of the chassis, so that it can more accurately reflect the acceleration and deceleration status of the vehicle. Based on the principle of inertia, it detects the displacement change of the mass block under the action of acceleration, converts it into an electrical signal, and then obtains the acceleration value. Whether it is a smooth acceleration of the vehicle or a drastic acceleration change during sudden braking, it can accurately capture it.

[0089] Steering angle measurement relies on an angle sensor installed on the steering column, which fits tightly to the steering system. As the steering wheel turns, it accurately measures the rotation angle of the steering shaft, instantly reflecting the vehicle's steering situation and providing key basis for subsequent analysis of vehicle operations such as lane changing and turning.

[0090] Secondly, it is crucial to set the data collection frequency reasonably. The speed data collection frequency can be set to 10-20 times per second, which can capture the smooth changes in speed during normal driving of the vehicle, such as when cruising at a constant speed on a highway, updating the speed data every 0.05-0.1 seconds to ensure a smooth speed curve; and can also keep up with the rapid fluctuations in speed in road conditions where the vehicle accelerates and decelerates frequently, such as when the vehicle frequently starts and stops in congested sections of the city, without missing any key change nodes.

[0091] The frequency of acceleration data collection is relatively higher, about 20 to 30 times per second, because the change in acceleration is often more sudden. At the moment of emergency braking or rapid start of the vehicle, high-frequency collection can fully record the sudden change of acceleration from zero to maximum value, providing detailed data support for subsequent analysis of the aggressiveness of driving behavior.

[0092] The frequency of collecting steering angle data can be set to 15-20 times per second, ensuring that the dynamic changes of the steering angle can be tracked continuously and accurately when the vehicle is turning quickly or fine-tuning the direction. For example, when the vehicle is making an S-curve, the continuous adjustment of the steering angle is recorded at an appropriate frequency to restore the details of the driving operation.

[0093] Thirdly, while each sensor collects data at a set frequency, it is necessary to ensure data synchronization. Due to differences in the working principles and response times of different sensors, high-precision clock synchronization technology is used to accurately correspond the speed, acceleration and steering angle data to the same moment. Based on the vehicle startup time, the collected data of each sensor is stamped with a millisecond-level timestamp. Through a special synchronization algorithm, in the subsequent data processing stage, the different dimensional data collected at the same time are paired and integrated according to the timestamp.

[0094] For example, when the speed sensor collects a vehicle speed of 50km / h at a certain moment, the acceleration sensor collects an acceleration of 0.5m / s² at the same time, and the steering angle sensor collects a steering angle of 5°, these three data are integrated into a group through timestamp matching, representing the vehicle's driving status at that moment, laying the foundation for subsequent conversion into a time curve.

[0095] Finally, after completing data collection and integration, we will enter the time course curve drawing stage. With time as the horizontal axis, the continuously collected speed data is drawn into a speed time course curve in sequence. The ups and downs of the curve directly reflect the increase and decrease of the vehicle speed. For example, in the acceleration stage, the curve shows an upward trend; in the deceleration stage, it tilts downward.

[0096] Similarly, with time as the horizontal axis, the acceleration data is plotted as an acceleration-time curve. The peak and valley values ​​of the curve clearly show the severity of the vehicle's acceleration and deceleration. Positive values ​​indicate acceleration, and negative values ​​indicate deceleration. The shape of the curve can be used to analyze the smoothness or aggressiveness of driving behavior.

[0097] The steering angle data is plotted as a direction angle time curve. The fluctuation of the curve reflects the frequency and amplitude of the vehicle's steering operation. When the vehicle is traveling in a straight line, the curve is close to horizontal; when turning or changing lanes, the curve shows obvious upward or downward fluctuations, fully presenting the vehicle's steering dynamics.

[0098] It is foreseeable that the above data collection process can also be applied to model training.

[0099] S4 inputs the driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors.

[0100] During the driving process of the vehicle, the collected driving data is input into the GA-CNN-BiGRU hybrid model to obtain the classification results of driving behavior. For example, when the driving behavior shows obvious acceleration abnormality or direction angle abnormality, the model can accurately classify it into the relevant abnormal category. The key advantage of this detection method lies in its multi-dimensional modeling ability of driving behavior and its sensitivity to abnormal behavior.

[0101] Embodiment 2: The present invention also provides a dangerous driving behavior recognition system based on a hybrid deep neural network, comprising: A model building module, used to build a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm; A training module, used for constructing a data set to train the GA-CNN-BiGRU hybrid model; The data acquisition module is used to collect driving data of the vehicle during driving, including speed, acceleration and steering angle; and convert the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve; The recognition module is used to input driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors; The GA-CNN-BiGRU hybrid model includes: Convolutional layer, used to extract spatial features from driving data and output feature maps; The pooling layer is used to reduce the dimension of the feature map output by the convolutional layer; A bidirectional gated recurrent component is used to obtain hidden states in feature maps from the forward and reverse directions of the time series and fuse them into time series features; The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features; Softmax function, used for classification prediction based on spatiotemporal features.

[0102] The steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model by the model building module include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration; S102 builds the corresponding GA-CNN-BiGRU hybrid model based on chromosomes; S103 evaluates the fitness of the GA-CNN-BiGRU hybrid model; S104 screens chromosomes in the population based on fitness; S105 performs crossover mutation on the screened chromosomes to form a new population; S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

[0103] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0104] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a computer terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0105] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A dangerous driving behavior recognition method based on a hybrid deep neural network, characterized in that include: S1 builds a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm; S2 constructs a data set to train the GA-CNN-BiGRU hybrid model; S3 collects driving data of the vehicle during driving, including speed, acceleration and steering angle; and converts the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve; S4 inputs the driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors; The GA-CNN-BiGRU hybrid model includes: Convolutional layer, used to extract spatial features from driving data and output feature maps; A pooling layer, used for reducing the dimension of the feature map output by the convolutional layer; A bidirectional gated loop component, used for obtaining hidden states in the feature map from the forward and reverse directions of the time series and fusing them into time series features; The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features; The Softmax function is used to perform classification prediction based on the spatiotemporal features.

2. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1 is characterized in that The steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model using a genetic algorithm include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration; S102 builds the corresponding GA-CNN-BiGRU hybrid model based on chromosomes; S103 evaluates the fitness of the GA-CNN-BiGRU hybrid model; S104 screens chromosomes in the population based on fitness; S105 performs crossover mutation on the screened chromosomes to form a new population; S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

3. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 2 is characterized in that The hyperparameters of the GA-CNN-BiGRU hybrid model optimized using the genetic algorithm include: the number of convolutional layers, kernel size, number of channels, and activation function; the type and number of pooling layers; the number of layers, number of hidden units, and activation function of the bidirectional gated recurrent component.

4. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1 is characterized in that The step of extracting spatial features from driving data by the convolutional layer includes: The convolution kernel S201 slides on the velocity time history curve data, the acceleration time history curve data and the direction angle time history curve according to the set step size, and obtains the local eigenvalue by performing convolution operation with the local data points; S202 concatenates the obtained local eigenvalues ​​to obtain a feature map.

5. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1 is characterized in that The step of obtaining the timing characteristics of the bidirectional gated loop component comprises: S301 forward gated recurrent unit combines the hidden state at the previous moment with the current information to output the positive hidden state at the current moment; S302 The backward gated recurrent unit combines the hidden state at the next moment with the current information to output the reverse hidden state at the current moment; S303 merges the current forward hidden state and the reverse hidden state into the current hidden state.

6. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1, characterized in that: The number of convolutional layers is 1 to 30, the size of the convolution kernel in each convolutional layer is any one of 3, 5, 7, and 9, the number of channels in each convolutional layer is any one of 32, 64, 128, and 256, and the activation function is any one of ReLU, LeakyReLU, and ELU; the type of the pooling layer is max or average, and the number of pooling layers is 2 or 3; the number of layers of the bidirectional gated recurrent component is 1 to 5, the number of hidden units is any one of 64, 128, 256, and 512, and the activation function is any one of ReLU, LeakyReLU, and ELU.

7. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1, characterized in that: The GA-CNN-BiGRU hybrid model is trained using the speed time history curve, acceleration time history curve and direction angle time history curve of labeled driving behaviors as samples; the driving behaviors include normal driving, sudden acceleration, sudden deceleration, abnormal left lane change and abnormal right lane change.

8. The method for identifying dangerous driving behaviors based on a hybrid deep neural network according to claim 1 is characterized in that The GA-CNN-BiGRU hybrid model also includes: The input layer before the convolutional layer is used to preprocess the input data; The output layer after the Softmax function is used to output the results of the Softmax function according to the corresponding driving behavior.

9. A dangerous driving behavior recognition system based on a hybrid deep neural network, characterized in that include: A model building module, used to build a GA-CNN-BiGRU hybrid model based on a convolutional neural network and a bidirectional gated recurrent unit; the hyperparameters of the GA-CNN-BiGRU hybrid model are optimized using a genetic algorithm; A training module, used for constructing a data set to train the GA-CNN-BiGRU hybrid model; The data acquisition module is used to collect driving data of the vehicle during driving, including speed, acceleration and steering angle; and convert the driving data into speed time history curve data, acceleration time history curve data and direction angle time history curve; The recognition module is used to input driving data into the trained GA-CNN-BiGRU hybrid model to identify dangerous driving behaviors; The GA-CNN-BiGRU hybrid model includes: Convolutional layer, used to extract spatial features from driving data and output feature maps; The pooling layer is used to reduce the dimension of the feature map output by the convolutional layer; A bidirectional gated recurrent component is used to obtain hidden states in feature maps from the forward and reverse directions of the time series and fuse them into time series features; The fully connected layer is used to integrate the spatial features extracted by the convolutional layer and the temporal features extracted by the BiGRU component to obtain spatiotemporal features; Softmax function, used for classification prediction based on spatiotemporal features.

10. The dangerous driving behavior recognition system based on hybrid deep neural network according to claim 9 is characterized in that The steps of optimizing the hyperparameters of the GA-CNN-BiGRU hybrid model by the model building module include: S101 randomly generates several chromosomes in the specified search space, each chromosome corresponds to a feasible convolutional neural network and bidirectional gated recurrent unit configuration; S102 builds the corresponding GA-CNN-BiGRU hybrid model based on chromosomes; S103 evaluates the fitness of the GA-CNN-BiGRU hybrid model; S104 screens chromosomes in the population based on fitness; S105 performs crossover mutation on the screened chromosomes to form a new population; S106 executes steps S102 to S105 in a loop until a preset indicator is reached.

Images