A slope intelligent identification method combining kalman filtering and deep learning
Patent Information
- Application Number
- CN202611024948.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-10-09
AI Technical Summary
[0004]本发明的目的在于克服现有技术中单一卡尔曼滤波在动态工况下线性化假设与噪声统计特性的适应性不可靠，导致坡度识别精度不稳定的不足，提供一种卡尔曼滤波与深度学习融合的坡度智能识别方法，有利于提升复杂道路工况下坡度识别的精度
本发明的一种卡尔曼滤波与深度学习融合的坡度智能识别方法，通过构建分类模型，以判断运动学卡尔曼滤波误差较大的区段，在分类模型判断出卡尔曼滤波方法误差大时，使用构建的深度学习模型进行修正，得到最终的坡度预测值，提升复杂道路工况下坡度识别的精度。
Smart Images
Figure CN122885341A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of slope recognition methods, and more specifically, to a slope intelligent recognition method that integrates Kalman filtering and deep learning. Background Technology
[0002] Road gradient has a significant impact on vehicle driving safety. When a vehicle is driving on an incline, the component of gravity causes uneven force on the tires, directly affecting handling stability and driving safety. If the driver or autonomous driving system cannot accurately perceive the gradient and adjust the control strategy in time, the vehicle may experience dangerous situations such as skidding or rollover. Simultaneously, road gradient directly affects a vehicle's power, fuel economy, and safety: going uphill requires greater driving force to overcome the component of gravity, and insufficient power may lead to rolling backward; going downhill requires effective braking to prevent loss of speed control. Studies show that fuel consumption is significantly higher on uphill sections compared to flat roads and downhill sections. Accurate gradient information helps optimize vehicle control strategies, such as adjusting transmission shifting strategies, engine output torque, braking force distribution, and the energy recovery intensity of new energy vehicles. For high-level autonomous driving, gradient data is a key input for path planning and decision-making control; the system needs to adjust parameters such as vehicle speed, steering, and power distribution in real time based on the gradient.
[0003] In existing technologies, Sun Yong et al. proposed a slope dynamic recognition algorithm based on vehicle acceleration. This method first filters and transforms the sensed signal in the time and frequency domains to obtain reliable sensing information. Then, it fuses the data using relevant kinematic relationships and optimizes the slope data through discrete Kalman filtering, correction algorithms, and automatic adjustment. By fusing multiple data sources, this method improves the accuracy of slope estimation to some extent. However, this method mainly relies on a single Kalman filter for prediction and estimation. In actual driving conditions, when the vehicle experiences dynamic conditions such as excessively strong braking signals or drastic acceleration changes, the linearization assumption and noise statistics of a single Kalman filter cannot accurately describe the nonlinear time-varying characteristics of the system. This leads to significant deviations in the filtering estimation results, a significant decrease in slope recognition accuracy, and difficulty in maintaining stable and reliable recognition performance under complex and variable real-world road conditions. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies where the linearization assumption and noise statistical characteristics of a single Kalman filter are unreliable under dynamic conditions, leading to unstable slope recognition accuracy. This invention provides a slope intelligent recognition method that integrates Kalman filtering and deep learning, which is beneficial for improving the accuracy of slope recognition under complex road conditions.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A slope intelligent recognition method integrating Kalman filtering and deep learning is provided, including the following steps: Collect historical driving data and historical gradient data; The historical driving data and the historical slope data are input into the deep learning model, and the deep learning model is trained to obtain a trained deep learning model. Construct a Kalman filter model; input the historical driving data into the Kalman filter model, calculate the error data of the Kalman filter model by combining the output value of the Kalman filter model with the historical slope data; train the classification model based on the historical driving data and the error data to obtain the trained classification model; Collect current driving data; Based on current driving data and the Kalman filter model, the predicted slope value is obtained; An error threshold is set, and the slope error is obtained based on the current driving data and the trained classification model. When the slope error is not greater than the error threshold, the slope prediction value is output. When the slope error is greater than the error threshold, the current driving data and the slope prediction value are input into the trained deep learning model to obtain the corrected slope prediction value, and the corrected slope prediction value is output.
[0006] In one optional approach, the step of inputting the historical driving data and the historical slope data into a deep learning model and training the deep learning model to obtain a trained deep learning model specifically includes the following steps: The historical driving data and the historical slope data are preprocessed separately to construct a dataset; The dataset is input into the deep learning model for training; A well-trained deep learning model is obtained.
[0007] In one optional approach, the preprocessing of the historical driving data and the historical slope data to construct the dataset specifically includes the following steps: High-frequency noise in the historical slope data is filtered out using a low-pass filter method, and the filtered historical slope data is used as a label. The historical driving data is segmented using a sliding window with a window size of 3 and a step size of 1, and the time-series data within the window is used as the input feature. The labels and input features are combined to construct the dataset.
[0008] In one alternative approach, the input to the deep learning model is the driving data and the slope data of the previous time step, and the output of the deep learning model is the slope data of the next time step.
[0009] In one alternative approach, the historical driving data includes historical vehicle speed and historical longitudinal acceleration, wherein the historical vehicle speed is acquired via a vehicle controller local area network bus, and the historical longitudinal acceleration is acquired via a vehicle inertial measurement unit; the historical slope data is acquired via a high-precision vehicle inertial measurement unit.
[0010] In one alternative approach, the deep learning model employs a CNN-LSTM model.
[0011] In one alternative approach, the Kalman filter model is constructed based on a vehicle kinematics model.
[0012] In one alternative approach, the vehicle kinematics model specifically uses the following functions:
[0013] In the formula, This represents the longitudinal acceleration collected by the vehicle's inertial measurement unit; Indicates the vehicle speed at the current moment; Indicates the vehicle speed at the previous moment; Indicates time; Represents gravitational acceleration; Indicates the slope.
[0014] In one alternative approach, the classification model employs a random forest classifier.
[0015] In one alternative approach, when the slope error is greater than the error threshold, the corrected slope prediction value and the slope prediction value are fused using a probability-weighted fusion formula to output the fused slope prediction value.
[0016] In one alternative approach, before obtaining the slope error based on the current driving data and the trained classification model by setting the error threshold, the following pre-judgment step is further included: Collect the current braking signal strength and set the braking strength threshold; If the braking signal strength is greater than the braking strength threshold, the subsequent setting error threshold is executed, and the slope error process is obtained based on the current driving data and the trained classification model; if the braking signal strength is not greater than the braking strength threshold, the step of obtaining the slope error is skipped, and the slope prediction value is directly output.
[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a slope intelligent recognition method that integrates Kalman filtering and deep learning. By constructing a classification model to identify sections with large kinematic Kalman filtering errors, the constructed deep learning model is used to correct the Kalman filtering error when the classification model determines that the error is large, thus obtaining the final slope prediction value and improving the accuracy of slope recognition under complex road conditions.
[0018] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the slope intelligent recognition method that integrates Kalman filtering and deep learning according to the present invention. Figure 2 This is a flowchart illustrating the training and testing phases of the slope intelligent recognition method that integrates Kalman filtering and deep learning in this invention. Figure 3 This is a flowchart illustrating the Kalman filter model in the slope intelligent recognition method that integrates Kalman filtering and deep learning in this invention. Figure 4 This is the vehicle speed change curve in Example 3; Figure 5 This is the longitudinal acceleration variation curve in Example 3; Figure 6 This refers to the historical slope data in Example 3; Figure 7 This refers to the filtered historical slope data from Example 3; Figure 8 This is a schematic diagram of the random forest discrimination results in Example 3; Figure 9 This is a schematic diagram of the test results of this method in Example 3; Figure 10 This is a schematic diagram showing the comparison between the present method and the traditional Kalman method in Example 3; Figure 11 This is an error graph comparing the method described in Example 3 with the traditional Kalman method. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The present invention will be further described below with reference to specific embodiments.
[0021] Furthermore, if the embodiments of the present invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text is to include three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution that simultaneously satisfies A and B.
[0022] Example 1 This embodiment is the first embodiment of a slope intelligent recognition method that integrates Kalman filtering and deep learning. Figure 1 As shown, it includes the following steps: Training and testing phases: Historical driving data and historical gradient data are collected. Historical driving data includes historical vehicle speed and historical longitudinal acceleration. Historical vehicle speed is collected through the vehicle controller area network (CAN) bus, and historical longitudinal acceleration is collected through the vehicle inertial measurement unit. Historical gradient data is collected through a high-precision vehicle inertial measurement unit (IMU).
[0023] Historical driving data and historical gradient data are input into a deep learning model to train it, resulting in a trained deep learning model. The specific steps include: Historical driving data and historical gradient data are preprocessed separately to construct datasets. Specifically, this includes the following steps: A low-pass filtering method is used to filter high-frequency noise in historical slope data, and the filtered historical slope data is used as a label. Historical driving data is segmented using a sliding window with a window size of 3 and a step size of 1, and the time series data within the window is used as the input feature. The dataset is constructed by merging labels and input features.
[0024] The dataset is input into the deep learning model for training; the input to the deep learning model is the driving data and the slope data of the previous time step, and the output of the deep learning model is the slope data of the next time step.
[0025] A trained deep learning model is obtained. In some embodiments, the deep learning model is an offline model, which is trained and then embedded in the vehicle.
[0026] A Kalman filter model is constructed, with driving data as input and slope value as output. Historical driving data is input into the Kalman filter model, and the output value of the Kalman filter model is calculated with the historical slope data to obtain the error data of the Kalman filter model. Based on the historical driving data and the error data, a classification model is trained to obtain a trained classification model. Specifically, when training the classification model, historical driving data is divided into training and test sets. The training set data is input into the Kalman filter model to obtain the error data of the Kalman filter model corresponding to the historical driving data. The historical driving data from the training set and the obtained error data are then input into the classification model for training. After training, the model enters the testing phase. The historical driving data from the test set is input into the Kalman filter model, and the classification model is used to determine whether the slope prediction value output by the Kalman filter model needs correction by the deep learning model. If the determination result is yes, the slope prediction value is corrected using the deep learning model and then output; otherwise, the slope prediction value is output directly. The accuracy of the overall method is verified using the final output slope prediction value until a well-trained classification model that meets the accuracy requirements is obtained.
[0027] The overall training and testing flowchart in this embodiment is as follows: Figure 2 As shown.
[0028] Prediction phase: Collect current driving data; Based on current driving data and the Kalman filter model, the predicted slope value is obtained; Set an error threshold, obtain the slope error based on the current driving data and the trained classification model; when the slope error is not greater than the error threshold, output the slope prediction value; when the slope error is greater than the error threshold, input the current driving data and the slope prediction value into the trained deep learning model to obtain the corrected slope prediction value, and output the corrected slope prediction value.
[0029] In some embodiments, before setting an error threshold and obtaining the slope error based on current driving data and a trained classification model, the following pre-judgment steps are included: Collect the current braking signal strength and set the braking strength threshold; If the braking signal strength is greater than the braking strength threshold, the subsequent setting error threshold is executed, and the slope error process is obtained based on the current driving data and the trained classification model; if the braking signal strength is not greater than the braking strength threshold, the step of obtaining the slope error is skipped, and the slope prediction value is directly output.
[0030] In some embodiments, when the slope error is greater than the error threshold, the corrected slope prediction value is fused with the slope prediction value using a probability-weighted fusion formula, supplemented by a sliding window history filling and start-up grace period mechanism, and the fused slope prediction value is output to ensure the temporal smoothness of the slope estimation and the robustness of the system.
[0031] This embodiment presents a slope intelligent recognition method that integrates Kalman filtering and deep learning. By constructing a classification model, it identifies sections with large kinematic Kalman filtering errors. When the classification model determines that the Kalman filtering method has a large error, the constructed deep learning model is used to correct it, thereby obtaining the final slope prediction value and improving the accuracy of slope recognition under complex road conditions.
[0032] In this training process, the deep learning model uses the driving data and gradient data from the previous moment, and outputs the gradient data for the next moment. This requires relatively accurate real-time gradient values. If the deep learning model's prediction from the previous moment is used as the gradient input for the next moment, the model's final output will tend to converge to a fixed value due to the inherent errors of the deep learning model itself. Therefore, this embodiment utilizes a Kalman filter model to provide real-time updated gradient data to the deep learning model, enabling the deep learning model to help correct prediction biases under braking conditions.
[0033] Example 2 This embodiment is the second embodiment of the slope intelligent recognition method that integrates Kalman filtering and deep learning. This embodiment is similar to the first embodiment, except that... Deep learning models can employ the CNN-LSTM model. The CNN-LSTM model is a hybrid deep learning model that combines a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network. The CNN module automatically extracts local features and short-term patterns from the input time-series data through convolution and pooling operations, while the LSTM module utilizes its unique gating mechanism and memory units to further capture long-term temporal dependencies in the data based on the features extracted by the CNN. The two modules are connected sequentially, allowing the model to focus on both local changes within a window and the overall evolution of slope over time. This enables comprehensive and effective modeling of slope information in vehicle driving data, improving the accuracy of slope prediction.
[0034] like Figure 3 As shown, the Kalman filter model is constructed based on the vehicle kinematics model.
[0035] The vehicle kinematics model specifically uses the following functions:
[0036] In the formula, This represents the longitudinal acceleration collected by the vehicle's inertial measurement unit; Indicates the vehicle speed at the current moment; Indicates the vehicle speed at the previous moment; Indicates time; Represents gravitational acceleration; Indicates the slope.
[0037] Considering that the vehicle pitch angle and road slope angle are very small, the vehicle kinematics model can be transformed into the following function:
[0038] In the formula, This represents the longitudinal acceleration collected by the vehicle's inertial measurement unit; This represents the vehicle's actual longitudinal acceleration; φ Indicates the vehicle's pitch angle; Represents gravitational acceleration; Indicates the longitudinal slope angle.
[0039] Based on the transformed function described above, a continuous-time differential equation is established with vehicle speed and longitudinal slope angle as state variables; the continuous-time differential equation specifically uses the following function:
[0040] In the formula, Indicates the rate of change of vehicle speed; This represents the rate of change of the longitudinal slope angle.
[0041] The continuous-time differential equation is discretized, and a Kalman filter model is constructed. The Kalman filter model specifically uses the following function:
[0042] in, , express Prior state estimation at time step; , express The control input vector at each time step; ,express The measured value at time 1, which is the value output by the PINN model; , represents the state transition matrix; , represents the control input matrix; , representing the observation matrix.
[0043] That is, the Kalman filter model is:
[0044] In the formula, express The vehicle speed at any given moment; express The longitudinal slope angle at any given time; Indicates time; express The vehicle speed at any given moment; express The longitudinal slope angle at any given time; express The longitudinal acceleration collected by the vehicle's inertial measurement unit at a given time; express Error in calculating the time step; express The estimated slope at time; express The measurement error of time.
[0045] The classification model employs a random forest classifier. Random forest is an ensemble learning algorithm based on decision trees. In this embodiment, the real-time estimation error of the Kalman filter is used as the supervision signal. A three-dimensional input feature vector is constructed by combining vehicle kinematic features (vehicle speed, absolute acceleration) and dynamic response features (rate of change of acceleration). 0 / 1 labels are automatically generated using the upper quantile of the Kalman filter error as an adaptive threshold to train the random forest model. The trained random forest model outputs continuous probability values for new samples and compares them with a dynamic threshold automatically optimized based on the maximum F1 score criterion on the validation set. This results in a "soft decision," outputting an activation signal indicating whether deep learning intervention is needed.
[0046] Example 3 This embodiment is the third embodiment of the slope intelligent recognition method that integrates Kalman filtering and deep learning. Based on the first and second embodiments, this embodiment verifies the prediction results of this method.
[0047] like Figure 4 The figure shows the vehicle speed variation curve, as shown below. Figure 5 The figure shows the longitudinal acceleration variation curve. Figure 6 The image shows historical slope data, such as... Figure 7 The image shows the filtered historical slope data.
[0048] like Figure 8 The results of the random forest discrimination are shown below. Figure 9 The test results for this method are shown below.
[0049] like Figure 10 The results show a comparison between this method and the traditional Kalman method. Figure 11 The figure shows the error between this method and the traditional Kalman method.
[0050] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0051] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A slope intelligent recognition method integrating Kalman filtering and deep learning, characterized in that, Includes the following steps: Collect historical driving data and historical gradient data; The historical driving data and the historical slope data are input into the deep learning model, and the deep learning model is trained to obtain a trained deep learning model. Construct a Kalman filter model; input the historical driving data into the Kalman filter model, calculate the error data of the Kalman filter model by combining the output value of the Kalman filter model with the historical slope data; The classification model is trained based on the historical driving data and the error data to obtain a trained classification model; Collect current driving data; Based on current driving data and the Kalman filter model, the predicted slope value is obtained; Set an error threshold, and obtain the slope error based on the current driving data and the trained classification model; when the slope error is not greater than the error threshold, output the slope prediction value; When the slope error is greater than the error threshold, the current driving data and the slope prediction value are input into the trained deep learning model to obtain the corrected slope prediction value, and the corrected slope prediction value is output.
2. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 1, characterized in that, The process of inputting the historical driving data and the historical slope data into a deep learning model and training the deep learning model to obtain a trained deep learning model specifically includes the following steps: The historical driving data and the historical slope data are preprocessed separately to construct a dataset; The dataset is input into the deep learning model for training; A well-trained deep learning model is obtained.
3. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 2, characterized in that, The preprocessing of the historical driving data and the historical slope data to construct the dataset includes the following steps: High-frequency noise in the historical slope data is filtered out using a low-pass filter method, and the filtered historical slope data is used as a label. The historical driving data is segmented using a sliding window with a window size of 3 and a step size of 1, and the time-series data within the window is used as the input feature. The labels and input features are combined to construct the dataset.
4. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 2, characterized in that, The input to the deep learning model is the driving data and the slope data of the previous moment, and the output of the deep learning model is the slope data of the next moment.
5. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 2, characterized in that, The historical driving data includes historical vehicle speed and historical longitudinal acceleration. The historical vehicle speed is collected through the vehicle controller local area network bus, and the historical longitudinal acceleration is collected through the vehicle inertial measurement unit. The historical slope data is collected through a high-precision vehicle inertial measurement unit.
6. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 1, characterized in that, The deep learning model used is the CNN-LSTM model.
7. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 1, characterized in that, The Kalman filter model is constructed based on a vehicle kinematics model; the vehicle kinematics model specifically uses the following functions: In the formula, This represents the longitudinal acceleration collected by the vehicle's inertial measurement unit; Indicates the vehicle speed at the current moment; Indicates the vehicle speed at the previous moment; Indicates time; Represents gravitational acceleration; Indicates the slope.
8. The slope intelligent recognition method integrating Kalman filtering and deep learning according to claim 1, characterized in that, The classification model uses a random forest classifier.
9. The slope intelligent recognition method integrating Kalman filtering and deep learning according to any one of claims 1 to 8, characterized in that, When the slope error is greater than the error threshold, the corrected slope prediction value and the slope prediction value are fused using a probability weighted fusion formula, and the fused slope prediction value is output.
10. The slope intelligent recognition method fusion of Kalman filtering and deep learning according to any one of claims 1 to 8, characterized in that, Before obtaining the slope error based on the current driving data and the trained classification model after setting the error threshold, the following preliminary judgment steps are also included: Collect the current braking signal strength and set the braking strength threshold; When the braking signal strength is greater than the braking strength threshold, the subsequent set error threshold is executed, and the slope error process is obtained based on the current driving data and the trained classification model. If the braking signal strength is not greater than the braking strength threshold, the step of obtaining the slope error is skipped, and the slope prediction value is directly output.