A braking intention recognition method based on feature extraction and signal optimization

Abstract

The application discloses a brake intention recognition method based on feature extraction and signal optimization, relates to the field of electric vehicle brake intention recognition, and comprises the following steps: collecting multi-source sensor signals in a vehicle braking process, extracting brake feature signals, and constructing a basic brake feature dataset; key brake features are screened based on a maximum correlation minimum redundancy criterion; variation mode decomposition is performed on the key brake feature time sequence signals, each brake feature signal is decomposed into multiple mode components, effective mode components are selected based on sample entropy for signal reconstruction, and optimized brake feature time sequence signals are obtained; and a long short-term memory network model is input, and a brake intention recognition result is output. The method combines feature screening, signal optimization and a long short-term memory network, and improves the recognition accuracy of the driver's brake intention.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle braking intention recognition, and more specifically to a braking intention recognition method based on feature extraction and signal optimization. Background Technology

[0002] With the gradual application of brake-by-wire systems in electric vehicles, braking systems are no longer limited to meeting basic deceleration and stopping requirements. They also need to coordinate with regenerative braking systems to rationally allocate regenerative braking force and hydraulic braking force according to the driver's real-time braking intention, so as to achieve a balance between braking energy recovery efficiency and braking safety.

[0003] Most existing braking intent recognition methods rely on single signals or a small number of features, such as brake pedal opening, pedal force, or master cylinder pressure, to roughly classify braking behavior using preset thresholds. However, a single signal cannot fully reflect the driver's braking needs under complex conditions. For example, slight braking during low-speed crawling may overlap with regular braking during high-speed driving in terms of pedal displacement, and the rapid change in pedal speed during emergency braking needs to be distinguished from accidental pedal presses by the driver. Relying on a small number of signals for threshold classification is insufficient to cover the above complex scenarios. In addition, with the continuous enrichment of vehicle sensor configurations, the types of signals that can be acquired during braking have increased significantly. However, existing methods directly stack multi-source signals into the recognition model, ignoring the information redundancy between different braking features, which increases the computational burden on the model and makes it prone to overfitting and insufficient generalization ability with limited training samples.

[0004] Meanwhile, the raw signals acquired during braking have short-term transient characteristics. In the initial stages of different braking intentions, the signal curves such as pedal displacement and master cylinder pressure often overlap for a long time, making it difficult to effectively distinguish them in the early stages. Signals such as pedal speed are often accompanied by amplitude jitter, further masking the essential differences between different braking intentions. Existing methods usually use low-pass filtering or wavelet thresholding for preprocessing, but these methods are prone to filtering out transient change information related to braking intentions during the denoising process, and lack the ability to adaptively distinguish the boundary between effective and noise components in the signal. It is difficult to achieve a good balance between preserving effective signal features and suppressing noise interference, which restricts further improvement in recognition accuracy.

[0005] Furthermore, braking intent is essentially a behavioral process with continuous temporal evolution characteristics generated by the driver's operation. The braking intent at the current moment is often closely related to the braking operation sequence in the past few seconds. It is difficult to accurately characterize the driver's true intent by simply relying on the signal characteristics at the current moment. Existing methods based on static features or short-time window statistical features fail to make full use of the long temporal dependencies between multi-source braking signals and are difficult to capture the collaborative evolution of different feature parameters over time during the braking process. This leads to recognition lag or misjudgment in scenarios where the boundary between normal braking and slight braking is blurred and emergency braking is highly sudden.

[0006] Therefore, how to design a braking intention recognition method based on feature extraction and signal optimization, suppress non-stationary noise interference in the original signal, and improve the recognition accuracy of different braking intention types is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a braking intention recognition method based on feature extraction and signal optimization, which aims to solve the problems of high computational burden and overfitting, low recognition accuracy, and insufficient utilization of long-term braking feature association information in existing braking intention recognition methods, so as to achieve accurate recognition of the driver's light braking, normal braking and emergency braking intentions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A braking intent recognition method based on feature extraction and signal optimization includes the following steps: S1. Collect multi-source sensor signals during vehicle braking, extract braking feature signals, and construct a basic braking feature dataset; S2. Based on the maximum correlation and minimum redundancy criterion, each braking feature in the basic braking feature dataset is filtered to obtain the key braking features; S3. Perform variational mode decomposition on the key braking feature time-series signal, decompose each braking feature signal into multiple mode components, select the effective mode components based on sample entropy for signal reconstruction, and obtain the optimized braking feature time-series signal. S4. Input the optimized braking feature time-series signal into the long short-term memory network model and output the braking intention recognition result.

[0010] Preferably, S1 includes: Raw sensor signals during the braking process are collected by sensors installed on the vehicle, including a pedal force sensor, a linear displacement sensor, a pressure sensor, and a vehicle speed sensor. Braking feature parameters are extracted from the original sensor signals to construct a basic braking feature dataset. ;in, For pedal displacement, For pedal speed, For pedal force, For vehicle speed, For braking and deceleration, Main cylinder pressure.

[0011] Preferably, S2 includes: Based on the mutual information between the basic braking feature dataset S and the braking intention type c, the maximum correlation index max D and the minimum redundancy index min R are established. A joint operator for mutual information quotient is constructed based on the maximum correlation index max D and the minimum redundancy index min R. Select the joint operator Braking features exceeding a preset threshold are used as key braking feature time-series signals after screening.

[0012] Preferably, the mutual information is represented as:

[0013] in, Let S be the joint probability density function of S and c. Let S be the marginal probability density function. Let c be the marginal probability density function.

[0014] Preferably, the mutual information quotient joint operator Represented as:

[0015] Where |S| is the number of braking features in the basic braking feature dataset S, x i With x j Let I(x) be any braking feature in S. i c) represents braking characteristic x i The mutual information value between I(x) and braking intention type c i ,x j ) represents braking characteristic x i With braking feature x j The mutual information values ​​between them.

[0016] Preferably, in step S3, decomposing each braking characteristic signal into multiple modal components includes: S31. Obtain the single braking characteristic signal f(t) to be decomposed; S32. Set the total number of modal components K, and construct a variational mode decomposition constraint equation with the objective of minimizing the sum of the bandwidths of each modal component and the constraint that the signal to be decomposed is equal to the sum of the bandwidths of each modal component. S33, Introducing a secondary penalty factor and Lagrange multipliers The variational mode decomposition constraint equations are transformed into augmented Lagrangian functions; S34. Initialize each modal component and its center frequency, and solve the augmented Lagrangian function using the alternating direction multiplier method, and update each modal component and its center frequency in turn. S35. Determine whether the convergence condition is met. If it is met, terminate the iteration and output each modal component. Otherwise, continue iterating until convergence.

[0017] Preferably, the variational mode decomposition constraint equation is expressed as:

[0018] in, This represents the k-th modal component. k represents the center frequency of the k-th modal component. This represents the square of the gradient 2 norm. Here, δ(t) represents the time gradient operator, δ(t) represents the Dirac function, and j represents the imaginary unit. This represents the convolution operator, and 'e' represents the base of the natural logarithm.

[0019] Preferably, the updating of each modal component and the center frequency of each modal component are expressed as follows:

[0020]

[0021] in, This represents the Fourier transform form of the k-th modal component after the (N+1)-th iteration. This represents the center frequency of the k-th modal component after the (N+1)-th iteration, where N represents the current iteration number. Indicates the modal component index, This represents the Fourier transform form of the k-th modal component after the (N+1)-th iteration. This represents the Fourier transform of the signal f(t) to be decomposed. Represents the Fourier transform of the Lagrange multiplier λ(t). Indicates the first The Fourier transform of each modal component, where ω represents the frequency domain variable.

[0022] Preferably, in step S3, the optimized braking characteristic timing signal includes: The sample entropy of each of the K modal components obtained by decomposition is calculated. The modal component with the smallest sample entropy is selected as the effective modal component. The effective modal components are superimposed and summed to obtain the optimized braking feature time series signal. Calculate sample entropy Represented as:

[0023] in, For the embedding dimension, For similarity tolerance, This indicates that the embedding dimension is When the distance between template vectors is less than the tolerance. The matching probability, This indicates that the embedding dimension is expanded to The corresponding matching probability.

[0024] Preferably, in S4, the long short-term memory network model includes a forget gate, an input gate, and an output gate; The forget gate determines the proportion of information retained in the previous cell state based on the current input and the previous hidden state. The input gate determines the weight for writing the candidate unit state into the current unit state based on the current input and the hidden state of the previous time step; the candidate unit state is generated based on the current input and the hidden state of the previous time step; the previous unit state after being filtered by the forget gate is combined with the candidate unit state after being weighted by the input gate to update the current unit state. The output gate generates the hidden state output at the current moment based on the current input and the hidden state at the previous moment, combined with the current cell state.

[0025] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. This method uses the maximum correlation and minimum redundancy criterion to screen multi-source braking features, selecting key features that are highly correlated with braking intent and have low redundancy. This effectively reduces information redundancy in the data, lowers the computational burden of subsequent signal processing and intent recognition models, and avoids overfitting problems caused by feature redundancy.

[0026] 2. Variational mode decomposition is used to decompose key braking feature signals, and effective mode components are selected based on sample entropy for signal reconstruction. This can effectively suppress non-stationary fluctuations and noise interference in the original braking signal, enhance the distinguishability of signals between different braking intention types, and provide more stable and discriminative input data for subsequent time-series recognition models.

[0027] 3. The time-series braking signal, after feature screening and signal optimization, is input into the Long Short-Term Memory Network model. By utilizing its gating mechanism and memory units, the model jointly learns the historical and current braking information. This fully leverages the long-term temporal dependencies of multi-source features during braking, enabling accurate identification of the driver's multi-level braking intentions. This provides a reliable basis for intention discrimination for vehicle regenerative braking force distribution and composite braking coordination control. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0029] Figure 1 This is a flowchart of a braking intent recognition method based on feature extraction and signal optimization. Figure 2 This is a schematic diagram of the braking intent recognition process based on feature extraction and signal optimization. Figure 3 This is a schematic diagram of the initial data for the pedal displacement. Figure 4 This is a schematic diagram of the initial pedal speed data; Figure 5 A schematic diagram of the initial pressure data for the main cylinder; Figure 6 This is a schematic diagram of the pedal displacement reconstruction data; Figure 7 This is a schematic diagram of the pedal speed reconstruction data; Figure 8 A schematic diagram of the main cylinder pressure recombination data; Figure 9 A schematic diagram illustrating the accuracy verification results of the LSTM-based recognition method; Figure 10 A schematic diagram showing the accuracy verification results of the recognition method combining LSTM and variable mode decomposition; Figure 11 This is a schematic diagram illustrating the real-time verification results of slight braking intent. Figure 12 This is a schematic diagram illustrating the real-time verification results of conventional braking intent. Figure 13 This is a schematic diagram illustrating the real-time verification results of the emergency braking intent. Detailed Implementation

[0030] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] like Figure 1 and Figure 2 As shown, this embodiment provides a braking intent recognition method based on feature extraction and signal optimization, including the following steps: S1. Collect multi-source sensor signals during vehicle braking, extract braking feature signals, and construct a basic braking feature dataset; S2. Based on the maximum correlation and minimum redundancy criterion, each braking feature in the basic braking feature dataset is filtered to obtain the key braking features; S3. Perform variational mode decomposition on the key braking feature time-series signal, decompose each braking feature signal into multiple mode components, select the effective mode components based on sample entropy for signal reconstruction, and obtain the optimized braking feature time-series signal. S4. Input the optimized braking feature time-series signal into the long short-term memory network model and output the braking intention recognition result.

[0032] By performing maximum correlation and minimum redundancy screening on multi-source braking features, it effectively reduces the risk of feature redundancy and model overfitting. It also uses variational mode decomposition combined with sample entropy to reconstruct key signals, suppressing non-stationary fluctuations and noise interference in the original braking signals. On this basis, it uses the gating mechanism and memory unit of long short-term memory network to jointly learn historical and current braking information, making full use of the long temporal dependencies of multi-source features during braking, thereby improving the recognition accuracy of different braking intention types.

[0033] The following provides a further detailed explanation of each step in the above method and its related technical features; In this embodiment, S1, the process involves collecting multi-source sensor signals during vehicle braking, extracting braking feature signals, and constructing a basic braking feature dataset; including: Raw sensor signals during the braking process are collected by sensors installed on the vehicle, including a pedal force sensor, a linear displacement sensor, a pressure sensor, and a vehicle speed sensor. Braking feature parameters are extracted from the original sensor signals to construct a basic braking feature dataset. ;in, For pedal displacement, For pedal speed, For pedal force, For vehicle speed, For braking and deceleration, Main cylinder pressure.

[0034] In this embodiment S2, based on the maximum correlation and minimum redundancy criterion, the braking features in the basic braking feature dataset are filtered to obtain key braking features; including: Based on the mutual information between the basic braking feature dataset S and the braking intention type c, the maximum correlation index max D and the minimum redundancy index min R are established. A joint operator for mutual information quotient is constructed based on the maximum correlation index max D and the minimum redundancy index min R. Select the joint operator Braking features exceeding a preset threshold are used as key braking feature time-series signals after screening.

[0035] Mutual information is represented as:

[0036] in, Let S be the joint probability density function of S and c. Let S be the marginal probability density function. Let c be the marginal probability density function.

[0037] Furthermore, the maximum correlation index max D and the minimum redundancy index min R are respectively expressed as:

[0038]

[0039] Where |S| is the number of braking features in the basic braking feature dataset S, x i With x j Let I(x) be any braking feature in S. i c) represents braking characteristic x i The mutual information value between I(x) and braking intention type c i ,x j ) represents braking characteristic x i With braking feature x j The mutual information values ​​between them.

[0040] Furthermore, the joint operators of mutual information providers Represented as:

[0041] Where |S| is the number of braking features in the basic braking feature dataset S, x i With xj Let I(x) be any braking feature in S. i c) represents braking characteristic x i The mutual information value between I(x) and braking intention type c i ,x j ) represents braking characteristic x i With braking feature x j The mutual information values ​​between them.

[0042] In this embodiment, each braking feature The values ​​are shown in Table 1 below: Table 1

[0043] Select the joint operator satisfy The braking characteristics are used as the key braking characteristic time-series signals after screening. That is, the final key braking characteristic time-series signals after screening include pedal displacement. Pedal speed and master cylinder pressure .

[0044] In this embodiment S3, variational mode decomposition is performed on the key braking feature time-series signal to decompose each braking feature signal into multiple mode components. Based on the sample entropy, the effective mode components are selected for signal reconstruction to obtain the optimized braking feature time-series signal. The decomposition of each braking characteristic signal into multiple modal components includes: S31. Obtain the single braking characteristic signal f(t) to be decomposed; S32. Set the total number of modal components K, and construct a variational mode decomposition constraint equation with the objective of minimizing the sum of the bandwidths of each modal component and the constraint that the signal to be decomposed is equal to the sum of the bandwidths of each modal component. S33, Introducing a secondary penalty factor and Lagrange multipliers The variational mode decomposition constraint equations are transformed into augmented Lagrangian functions; S34. Initialize each modal component and its center frequency, and solve the augmented Lagrangian function using the alternating direction multiplier method, and update each modal component and its center frequency in turn. S35. Determine whether the convergence condition is met. If it is met, terminate the iteration and output each modal component. Otherwise, continue iterating until convergence.

[0045] Specifically, firstly, the single braking signal in the initial data is decomposed using variable mode decomposition. Decomposed into Modal components The sum of the bandwidths of each component must be kept to a minimum. Therefore, the variable mode decomposition constraint equations are established:

[0046] In the formula, The center frequency of the modal component; The square of the gradient second normal form; The number of modes; Modal components; For time gradient; for Analyzed signals; Secondly, in order to obtain the optimal solution of the variable mode decomposition constraint equations, a quadratic penalty factor is introduced. and Lagrange multipliers Constructing the augmented Lagrangian function:

[0047] Furthermore, the augmented Lagrangian function is solved using the alternating direction method of multipliers, establishing its frequency domain form. , Iterative equations:

[0048]

[0049] In the formula, This represents the number of iterations. It is one of the modal components; , and They are respectively , and The Fourier transform form; The modal component with the minimum sampling entropy is selected to reconstruct the optimized signal. Modes with lower sample entropy typically exhibit stronger regularity and lower random fluctuations, thus better reflecting the effective evolutionary information in the braking signal. The sample entropy of each modal component is calculated... Wei and The negative logarithm of the ratio of matching probabilities of the template vectors is obtained as follows:

[0050] In the formula, Here is the expression for sample entropy; This indicates that the embedding dimension is When the distance between template vectors is less than the tolerance. The matching probability; This indicates that the embedding dimension is expanded to The corresponding matching probability; Pedal displacement obtained through braking test Pedal speed and master cylinder pressure Signals, and classify and label the initial signal data, such as Figure 3 , Figure 4 , Figure 5 As shown, it can be seen that in the initial stage of 0~0.2s, the braking intentions under the three conditions... and The curves basically overlap. As time increases, emergency braking can be distinguished after 0.6 seconds, but there are still some overlapping areas under light braking and normal braking. The curve exhibits significant jitter. Therefore, this embodiment designs a variable mode decomposition method to further optimize the braking signal based on the aforementioned problems. about , and Reconstructed data such as Figure 6 , Figure 7 , Figure 8 As shown, compared to the original signal, the signal after optimization... and In the initial stage, it demonstrates higher distinguishability of the three braking intentions, and the smoothness of the signal curve is improved. Braking intentions can be more clearly characterized at a braking time of approximately 0.4 seconds, reducing the overlap between light braking and regular braking. Furthermore, signal optimization enhances... The stability of the data reduced the amplitude of signal jitter, providing a good data foundation for subsequent braking intention recognition.

[0051] Example S4: Input the optimized braking feature time-series signal into the long short-term memory network model and output the braking intention recognition result; Considering the significant correlation between different braking intentions and historical time-series signals of braking characteristics, this embodiment selects the Long Short-Term Memory (LSTM) model, which has significant advantages in capturing temporal feature relationships, for the design of the braking intention recognition model; , and The reconstructed dataset is used as input to the LSTM, and the output is the braking intention. In the chain structure, each repeating module represents an LSTM cell; each LSTM cell mainly consists of three gating mechanisms: a forget gate, an input gate, and an output gate. The forget gate determines the proportion of information retained in the previous cell state based on the current input and the previous hidden state. The input gate determines the weight for writing the candidate unit state into the current unit state based on the current input and the hidden state of the previous time step; the candidate unit state is generated based on the current input and the hidden state of the previous time step; the previous unit state after being filtered by the forget gate is combined with the candidate unit state after being weighted by the input gate to update the current unit state. The output gate generates the hidden state output at the current moment based on the current input and the hidden state at the previous moment, combined with the current cell state.

[0052] Specifically, at time t, given the input at the current time... Hidden state from the previous moment The calculation methods for each gating mechanism are as follows: The forget gate uses the Sigmoid function. Handle the previous hidden state And the current input Filtering information from previous cells is specifically represented as follows:

[0053] In the formula, Input for the recognition model; The previous hidden state; Forget gate expression; Secondly, input the goalkeeper Useful information is stored in the current unit and updated for control purposes; among which, the following is adopted: The tanh function generates the potential state after determining the update information. :

[0054]

[0055] In the formula, This represents a potential cell state. For the input gate expression; Then, the current unit state Update to previous cell state and Combinations:

[0056] In the formula, This represents the current cell state. The previous cell state; Finally, output gate pair and conduct The tanh function transformation outputs the current hidden state for future cells. :

[0057]

[0058] In the formula, Currently hidden; This is the output gate expression; , , and These are the weight matrices corresponding to the gating mechanism; , , and These are the corresponding bias vectors.

[0059] In this embodiment, the specific braking intention types include light braking, normal braking, and emergency braking. After outputting the braking intention recognition result, the category with the highest probability among the braking intention types is used as the final braking intention determination result.

[0060] In this embodiment, to verify the effectiveness of the braking intent recognition method, accuracy and real-time performance experiments were conducted. A comparison was performed between the basic recognition method using only a single LSTM and the recognition method combining LSTM and variable mode decomposition in this embodiment. 294 sets of sample data containing three types of braking intent were selected for the experiment and randomly divided into two parts at a 2:1 ratio, resulting in 196 training samples and 98 test samples. Here, 1, 2, and 3 represent braking intents such as light braking, normal braking, and emergency braking, respectively. The experiment was set to a no-braking state at both the beginning and end. Experiment 1 verified the accuracy of braking intention recognition. The experimental results are as follows: Figure 9 , Figure 10 As shown, Figure 9 , Figure 10 The accuracy results of applying the basic identification method and the identification method of this embodiment are shown respectively. The percentage on the diagonal represents the correct identification rate, and the others represent the false identification rate of misidentification as other intentions. It can be seen that the accuracy rate of the basic identification method alone is 95.6%, with false identifications distributed among all braking intentions. However, the accuracy rate of the identification method of this embodiment can reach 98.2%, with false identifications mainly concentrated between light braking and normal braking. Emergency braking can be completely and accurately identified. Experiment 2 uses the collected master cylinder pressure signal as an example to perform online recognition of braking intent. The real-time verification results are as follows: Figure 11 , Figure 12 , Figure 13 As shown, where, Figure 11As shown in the results of the slight braking intention recognition, it can be seen that after the driver applied braking pressure at 3.2s, the basic recognition method accurately identified the braking intention as slight braking at 3.578s, taking 0.378s, while the recognition method in this embodiment only identified it as slight braking at 3.560s, taking 0.360s. Figure 12 As a result of conventional braking intention recognition, braking pressure was applied at 1.7s. The recognition method in this embodiment accurately identified it as conventional braking after 0.451s; however, the basic recognition method incorrectly identified the braking intention as slight braking after 0.468s. Figure 13 Regarding the emergency braking intent recognition results, the recognition method in this embodiment still performs outstandingly. When braking pressure is applied at 1.5s, the emergency braking intent is accurately recognized after 0.485s, while the basic recognition method recognizes emergency braking within 0.505s. Therefore, the recognition method in this embodiment has good real-time performance and accuracy under all three braking intents. The recognition speed of the basic recognition method is similar to that of the recognition method in this embodiment, but its accuracy is lower due to the high complexity of the original signal.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0062] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A braking intent recognition method based on feature extraction and signal optimization, characterized in that, Includes the following steps: S1. Collect multi-source sensor signals during vehicle braking, extract braking feature signals, and construct a basic braking feature dataset; S2. Based on the maximum correlation and minimum redundancy criterion, each braking feature in the basic braking feature dataset is filtered to obtain the key braking features; S3. Perform variational mode decomposition on the key braking feature time-series signal, decompose each braking feature signal into multiple mode components, select the effective mode components based on sample entropy for signal reconstruction, and obtain the optimized braking feature time-series signal. S4. Input the optimized braking feature time-series signal into the long short-term memory network model and output the braking intention recognition result.

2. The braking intent recognition method based on feature extraction and signal optimization according to claim 1, characterized in that, S1 includes: Raw sensor signals during the braking process are collected by sensors installed on the vehicle, including a pedal force sensor, a linear displacement sensor, a pressure sensor, and a vehicle speed sensor. Braking feature parameters are extracted from the original sensor signals to construct a basic braking feature dataset. ;in, For pedal displacement, For pedal speed, For pedal force, For vehicle speed, For braking and deceleration, Main cylinder pressure.

3. The braking intent recognition method based on feature extraction and signal optimization according to claim 1, characterized in that, S2 includes: Based on the mutual information between the basic braking feature dataset S and the braking intention type c, the maximum correlation index maxD and the minimum redundancy index minR are established. A joint operator for mutual information quotient is constructed based on the maximum correlation index max D and the minimum redundancy index min R. Select the joint operator Braking features exceeding a preset threshold are used as key braking feature time-series signals after screening.

4. The braking intent recognition method based on feature extraction and signal optimization according to claim 3, characterized in that, The mutual information is represented as follows: in, Let S be the joint probability density function of S and c. Let S be the marginal probability density function. Let c be the marginal probability density function.

5. The braking intent recognition method based on feature extraction and signal optimization according to claim 3, characterized in that, The mutual information quotient joint operator Represented as: Where |S| is the number of braking features in the basic braking feature dataset S, x i With x j Let I(x) be any braking feature in S. i c) represents braking characteristic x i The mutual information value between I(x) and braking intention type c i ,x j (x) represents braking characteristics. i With braking feature x j The mutual information values ​​between them.

6. The braking intent recognition method based on feature extraction and signal optimization according to claim 1, characterized in that, In step S3, decomposing each braking characteristic signal into multiple modal components includes: S31. Obtain the single braking characteristic signal f(t) to be decomposed; S32. Set the total number of modal components K, and construct a variational mode decomposition constraint equation with the objective of minimizing the sum of the bandwidths of each modal component and the constraint that the signal to be decomposed is equal to the sum of the bandwidths of each modal component. S33, Introducing a secondary penalty factor and Lagrange multipliers The variational mode decomposition constraint equations are transformed into augmented Lagrangian functions; S34. Initialize each modal component and its center frequency, and solve the augmented Lagrangian function using the alternating direction multiplier method, and update each modal component and its center frequency in turn. S35. Determine whether the convergence condition is met. If it is met, terminate the iteration and output each modal component. Otherwise, continue iterating until convergence.

7. The braking intent recognition method based on feature extraction and signal optimization according to claim 6, characterized in that, The variational mode decomposition constraint equation is expressed as: in, This represents the k-th modal component. k represents the center frequency of the k-th modal component. This represents the square of the gradient 2 norm. Here, δ(t) represents the time gradient operator, δ(t) represents the Dirac function, and j represents the imaginary unit. This represents the convolution operator, and 'e' represents the base of the natural logarithm.

8. The braking intent recognition method based on feature extraction and signal optimization according to claim 6, characterized in that, The updated modal components and their center frequencies are expressed as follows: in, This represents the Fourier transform form of the k-th modal component after the (N+1)-th iteration. This represents the center frequency of the k-th modal component after the (N+1)-th iteration, where N represents the current iteration number. Indicates the modal component index, This represents the Fourier transform form of the k-th modal component after the (N+1)-th iteration. This represents the Fourier transform of the signal f(t) to be decomposed. Represents the Fourier transform of the Lagrange multiplier λ(t). Indicates the first The Fourier transform of each modal component, where ω represents the frequency domain variable.

9. The braking intent recognition method based on feature extraction and signal optimization according to claim 1, characterized in that, In step S3, the optimized braking characteristic timing signal includes: The sample entropy of each of the K modal components obtained by decomposition is calculated. The modal component with the smallest sample entropy is selected as the effective modal component. The effective modal components are superimposed and summed to obtain the optimized braking feature time series signal. Calculate sample entropy Represented as: in, For the embedding dimension, For similarity tolerance, This indicates that the embedding dimension is When the distance between template vectors is less than the tolerance. The matching probability, This indicates that the embedding dimension is expanded to The corresponding matching probability.

10. The braking intent recognition method based on feature extraction and signal optimization according to claim 1, characterized in that, In S4, the long short-term memory network model includes a forget gate, an input gate, and an output gate; The forget gate determines the proportion of information retained in the previous cell state based on the current input and the previous hidden state. The input gate determines the weight for writing the candidate unit state into the current unit state based on the current input and the hidden state of the previous time step; the candidate unit state is generated based on the current input and the hidden state of the previous time step; the previous unit state after being filtered by the forget gate is combined with the candidate unit state after being weighted by the input gate to update the current unit state. The output gate generates the hidden state output for the current moment based on the current input and the hidden state of the previous moment, combined with the current cell state.

Images