Electric vehicle braking intention recognition method and braking control system

Abstract

The invention discloses an electric vehicle braking intention recognition method and a braking control system. The method comprises the steps that vehicle parameters are collected, and a braking intention recognition model for recognizing the braking strength is constructed; the braking intention recognition model is constructed based on a fuzzy neural network and comprises a front-end network and a rear-end network. And the front-end network extracts the standardized activation intensity of each fuzzy rule according to the vehicle parameters. And the rear-end network is used for obtaining the braking strength by combining the standardized activation strength corresponding to each fuzzy rule and the rear-end parameters. According to the method, the braking intention recognition model based on the fuzzy rule is constructed to recognize the braking intention contributing to the kinetic energy recovery efficiency. Meanwhile, in the parameter optimization process, the inferior solution acceptance probability is adjusted through the gradually-decreased acceptance, efficient parameter optimization is achieved, and meanwhile the model is prevented from falling into a local optimal trap.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle technology, specifically relating to a method for recognizing braking intentions in electric vehicles and a braking control system. Background Technology

[0002] Pure electric vehicles, with their advantages of zero emissions and low operating noise, have become a core direction for the upgrading of the automotive industry. The drive motors of pure electric vehicles possess the characteristic of "bidirectional electromechanical energy conversion," providing a technological foundation for the recovery and utilization of braking energy. However, traditional regenerative braking systems face challenges in energy recovery, comfort, and adaptability. Optimizing energy recovery can prevent battery overcharging or undercharging, making electric vehicles more sustainable and cost-effective. Braking intention recognition refers to identifying the driver's braking intention based on the vehicle's state during braking when the braking power demand is unknown. This directly determines whether the current braking intensity of the entire vehicle can be reasonably judged, thus affecting the distribution of motor power and mechanical braking force, ultimately impacting the regenerative braking energy recovery effect and the overall braking performance of the vehicle.

[0003] Existing technologies often employ fuzzy controllers based on fixed rules to recognize braking intentions. This relies heavily on subjective human factors, potentially leading to inaccuracies in vehicle models requiring high precision. While some studies have used neural networks to achieve a certain level of accuracy, precisely identifying conditions such as normal deceleration and emergency braking, they face challenges including high workload, poor data universality, and slow convergence. Insufficient accuracy or response delays in the recognition algorithm will reduce energy recovery efficiency and negatively impact the driver's braking feel. Summary of the Invention

[0004] The purpose of this invention is to provide a method and control system for recognizing braking intentions in electric vehicles, which takes vehicle speed, remaining battery power, and brake pedal displacement as inputs and braking intensity as outputs, thereby improving the problems of low energy recovery efficiency and affecting braking feel and smoothness.

[0005] This invention provides a method for recognizing braking intent in electric vehicles, comprising:

[0006] Vehicle parameters are collected; the interval of each vehicle parameter is divided into multiple fuzzy subsets.

[0007] A braking intent recognition model is constructed. This model is based on a fuzzy neural network and includes a front-end network and a back-end network. The front-end network extracts the standardized activation intensity of each fuzzy rule based on vehicle parameters. The back-end network combines the standardized activation intensity corresponding to each fuzzy rule with the back-end parameters to obtain the braking intensity.

[0008] Parameter optimization of the braking intention recognition model: (1) Initialize the parameter vector of the braking intention recognition model and set the initial acceptance level; (2) Randomly adjust the parameter vector to generate candidate vectors; (3) Use the objective function to judge the prediction performance to verify the candidate vectors. If the candidate vectors pass the verification, they are directly accepted as parameter vectors; if they fail the verification, they are accepted as parameter vectors according to the update probability. The update probability is positively correlated with the acceptance level; (4) Iterate through steps (2) and (3), and the acceptance level gradually decreases during the iteration process.

[0009] Vehicle parameters are input into a parameter-optimized braking intent recognition model to obtain braking intensity.

[0010] Preferably, the front-end network includes a front-end input layer, a front-end fuzzification layer, a rule layer, and a normalization layer. The front-end fuzzification layer is used to extract the membership degree of each vehicle parameter to each fuzzy subset. The rule layer is used to obtain the activation intensity of different fuzzy rules based on the membership degree of each vehicle parameter. The normalization layer is used to normalize the activation intensity of each fuzzy rule.

[0011] Preferably, in the front-end fuzzification layer, each fuzzy subset of each vehicle parameter has a membership function; the membership function is a Gaussian function. The width parameter and center position parameter of each membership function are added to the parameter vector.

[0012] Preferably, the expression for the normalization layer is:

[0013]

[0014] in, , The activation intensities are before and after standardization, respectively.

[0015] Preferably, the backend network includes a backend input layer, a backend fuzzification layer, and a backend output layer connected in sequence. The backend fuzzification layer includes the same number of neural nodes as the number of fuzzy rules. Each neural node in the backend fuzzification layer calculates the backend parameters corresponding to different fuzzy rules. The backend output layer uses the standardized activation intensity corresponding to each fuzzy rule as a weight, and performs a weighted summation of the backend parameters corresponding to each fuzzy rule to obtain the braking intensity.

[0016] As a preferred option, candidate vectors in parameter optimization The expression is as follows:

[0017]

[0018]

[0019]

[0020] in, This is the current parameter vector; These are random numbers that follow a standard normal distribution. This is the step scaling value; For acceptability; ub is the current iteration number; lb is the upper bound of parameter adjustment; and lb is the lower bound of parameter adjustment. This is the initial value for acceptability.

[0021] Preferably, the candidate vector verification process is as follows: comparing the parameter vector s with the candidate vector. The function value obtained by inputting the objective function , .like If the check passes, the verification is successful. If the check fails, the probability is updated. .

[0022] As a preferred method, the collected vehicle parameters undergo noise reduction, outlier removal, and normalization.

[0023] Preferably, the vehicle parameters include vehicle speed, remaining battery power, and brake pedal displacement.

[0024] Secondly, the present invention provides an electric vehicle braking control system, which includes a parameter acquisition module, a braking intent recognition module, and a braking control module. The parameter acquisition module is used to acquire vehicle parameters in real time and input them into the braking intent recognition module. The braking intent recognition module is used to execute the aforementioned electric vehicle braking intent recognition method to obtain a braking intensity signal. The braking control module controls the electric vehicle to brake based on the braking intensity signal.

[0025] The present invention has the following beneficial effects.

[0026] This invention constructs a braking intention recognition model based on fuzzy rules, which can comprehensively identify braking intentions that contribute to kinetic energy recovery efficiency based on multiple vehicle parameters. Simultaneously, a random perturbation strategy is used to introduce appropriate randomness into the parameters of the braking intention recognition model, preventing the model from over-relying on local features in the training samples. Furthermore, during parameter optimization, this invention uses a gradually decreasing acceptability to adjust the acceptance probability of inferior solutions, achieving efficient parameter optimization while avoiding the model falling into local optima traps. Attached Figure Description

[0027] Figure 1 This is a flowchart of Embodiment 1 of the present invention;

[0028] Figure 2 This is a schematic diagram of the braking intent recognition model in Embodiment 1 of the present invention;

[0029] Figure 3 This is a comparison chart of the predicted and actual braking intention values ​​of the braking intention recognition model in Embodiment 1 of the present invention.

[0030] Figure 4 The above are the membership function and rule graph of the braking intent recognition model before optimization in Embodiment 1 of the present invention.

[0031] Figure 5 This refers to the optimized membership function and rule graph of the braking intent recognition model in Embodiment 1 of the present invention.

[0032] Figure 6 This is a typical operating condition diagram for the CLTC-P used in testing.

[0033] Figure 7 This is a comparison chart of the remaining power changes of Example 1 and the comparative example under CLTC-P conditions;

[0034] Figure 8 This is a comparison chart of the energy recovery of Example 1 and the comparative example under CLTC-P conditions. Detailed Implementation

[0035] Example 1

[0036] A method for recognizing braking intent in an electric vehicle includes the following steps:

[0037] Step 1. Collect vehicle data

[0038] 1-1. Vehicle parameters are collected using the vehicle control unit (VCU). These vehicle parameters are driver braking operation and vehicle response data, including vehicle speed (v), remaining battery charge (SOC), and brake pedal displacement (P_D).

[0039] 1-2 Noise reduction, outlier removal and normalization are performed on the collected raw vehicle parameter data to eliminate interference and unify the data range.

[0040] Step 2. Construct an initial braking intent recognition model.

[0041] like Figure 2 As shown, the braking intent recognition model is built on a fuzzy neural network (adaptive neurofuzzy inference system), which includes a front-end network and a back-end network.

[0042] The front-end network is used to match fuzzy rules and consists of a four-layer structure: a front-end input layer, a front-end fuzzification layer, a rule layer, and a normalization layer, which are connected in sequence.

[0043] The front-end input layer is used to pass input variables to the next layer (i.e., the fuzzification layer) without performing any calculations. The input variables are the standardized car speed, remaining battery power, and brake pedal displacement; the number of neurons matches the number of input variables, i.e., there are 3 neurons in this embodiment.

[0044]

[0045] in, This represents the feature vector passed from the input layer to the fuzzing layer; , , These represent the actual input values ​​for vehicle speed, remaining battery power, and brake pedal displacement, respectively.

[0046] The front-end fuzzification layer is used to convert the input feature vector into the membership degree of the fuzzy set through a membership function; in this embodiment, the membership function for fuzzification is a Gaussian function. The car speed, remaining battery power, and brake pedal displacement in the feature vector are each divided into three fuzzy subsets, so the fuzzification layer includes nine neurons. Each fuzzy subset corresponds to a linguistic variable. Among them, the range of car speed is [0, 120], and the set is defined as [L (low), M (medium), H (high)]; the range of remaining battery power is [0, 1], and the set is defined as [L (low), M (medium), H (high)]; the range of brake pedal displacement is [0, 1], and the set is defined as [L (low), M (medium), H (high)].

[0047] The membership function is constructed as follows:

[0048]

[0049] In the formula, i = 1, 2, 3; j = 1, 2, ..., m i ;m i It is a parameter The number of fuzzy segments; in this embodiment, m1, m2, and m3 are all 3; The membership degree is the output result of fuzzification. It is a parameter The A fuzzy subset This is the membership function.

[0050] The rule layer is used to match and activate fuzzy rules. Each node in the rule layer corresponds to a fuzzy control rule. In this embodiment, the rule layer has 27 nodes, corresponding to the activation strength of 27 fuzzy rules. The expression is as follows:

[0051]

[0052] In the formula, i1, i2, and i3 correspond to the linguistic variable numbers of the three input variables; i1=1,2,...,m1; i2=1,2,...,m2; i3=1,2,...,m3; j=1,2,...,27; It is the product of the membership degrees of the three vehicle parameters, corresponding to the activation intensity of the fuzzy rule.

[0053] The standardization layer is used to perform standardized computation, which measures the activation intensity of fuzzy language expressions. Transformed into the control quantity of the corresponding domain The number of nodes in this layer is the same as in the third layer, which is 27. The standardized calculation formula is as follows:

[0054]

[0055] The back-end network is used to process parameter standardization and includes a back-end input layer, a back-end fuzzification layer, and a back-end output layer connected in sequence. The back-end input layer takes into account the remaining battery power, vehicle speed, brake pedal displacement, and a constant 1. The constant is used to blur the constant terms appearing in the back-end network. It can participate in calculations.

[0056] Each node in the backend fuzzification layer corresponds to a fuzzy rule, established based on the fuzzy rules in the rule layer of the frontend network. In this embodiment, there are a total of 27 fuzzy rules, therefore, the backend fuzzification layer has a total of 27 neural nodes, which can be represented as follows:

[0057]

[0058] In the formula, =1, 2, 3....27 represents 27 fuzzy rules; This is a constant term used to ensure that the rule output meets the requirements of actual braking scenarios; Input variables In the The coefficients in the rule, Input variables In the The coefficients in the rule, Input variables In the The coefficients in the rule.

[0059] The output of the back-end output layer is the braking intensity Z; the braking intensity Z is obtained by weighted summation of the output parameters of the back-end fuzzification layer.

[0060]

[0061] in, The number of fuzzy rules is 27.

[0062] Step 3. Optimize the parameters of the braking intention recognition model.

[0063] 3-1. Parameter Initialization:

[0064] The braking intent recognition model includes front-end parameters and back-end parameters. The front-end parameters are the parameters in the membership function.

[0065] In this embodiment, the membership function adopts the form of a Gaussian function, and its expression is:

[0066]

[0067] In the formula, For membership degree, For the input vehicle parameters; This is the width parameter of the function; The center position parameter of the function;

[0068] Each fuzzy subset of each input variable corresponds to an independent Gaussian membership function, and each membership function has corresponding parameters. , There are nine Gaussian membership functions; 18 front-end parameters need to be set.

[0069] The fuzzy rule corresponds to four coefficients. The number of fuzzy rules is 27; the number of backend parameters that need to be set is 27 × 4 = 108.

[0070] Create a vector s containing all front-end and back-end parameters:

[0071]

[0072] Where i = 1, 2, ..., 9; j = 1, 2, ..., 27; For the front-end parameter corresponding to the i-th membership function; These are the backend parameters corresponding to the j-th fuzzy rule. There are 18 frontend parameters and 27 × 4 = 108 backend parameters.

[0073] 3-2. Set the initial acceptance level The rate of decline in acceptance is 100. The value is 0.95, the maximum number of iterations is 100, the upper boundary of parameter adjustment ub is 1, and the lower boundary of parameter adjustment lb is 0.

[0074] 3-3. Perturb the neighborhood of vector s:

[0075] Randomly fine-tuned parameter vector Generate candidate vectors as follows:

[0076]

[0077]

[0078]

[0079] in, These are random numbers that follow a standard normal distribution. This is the step scaling value; For acceptability; This is the current iteration number, initially set to 0.

[0080] In the early stages of iteration, the acceptance rate is relatively high, and a certain probability of accepting a "worse solution" is acceptable in order to escape the local optimum. As the number of iterations decreases, the probability of accepting a worse solution gradually decreases.

[0081] 3-4. Candidate vector acceptance judgment:

[0082] To better verify the reliability and accuracy of the model, an objective function for prediction performance is constructed. as follows:

[0083]

[0084] Where N is the number of samples, For the actual braking intention of the i-th sample, This is the predicted output of the braking intent recognition model.

[0085] objective function The closer to 0, the better the model's predictive performance.

[0086] like If so, then accept the candidate vector as the updated parameter vector; otherwise, use probability. Accept candidate vectors as updated parameter vectors.

[0087]

[0088] 3-5. Termination Judgment:

[0089] Increase the iteration count by 1 and re-execute steps 3-3 and 3-4 until the current iteration count k is greater than or equal to the maximum iteration count.

[0090] Step 4. During vehicle operation, three vehicle parameters are collected in real time and input into the braking intent recognition model optimized in Step 3. The braking intent recognition model outputs a braking intensity signal in real time.

[0091] The accuracy evaluation of the braking intent recognition model optimized by parameters in this embodiment is as follows: Figure 3 As shown. The calculated root mean square error (RMSE) of the prediction results in this embodiment is 0.030115.

[0092] The membership functions corresponding to the fuzzy subsets of vehicle parameters before and after step 3 iteration are as follows: Figure 4 , 5 As shown.

[0093] Based on the electric vehicle braking intent recognition method provided in this embodiment, regenerative braking control of electric vehicles can be performed. Specifically, the braking intent recognition model is used to output a braking intensity signal in real time to control the vehicle's kinetic energy recovery system, enabling the vehicle's drive motor to recover kinetic energy and generate electricity.

[0094] This embodiment takes the typical braking conditions of electric vehicles as the research object and analyzes them in conjunction with actual vehicle operating parameters. Typical driving conditions of CLTC-P are selected as follows: Figure 6 As shown, a regenerative braking control system model based on a braking intention recognition model was built in the MATLAB / Simulink platform. The actual vehicle speed (v), remaining battery power (SOC), brake pedal opening (P_D) and other parameters were input, and the energy recovery efficiency of the control method was quantitatively analyzed and compared in the CRUISE platform.

[0095] The CLTC-P test cycle selected in this embodiment is the Chinese passenger car test cycle, which includes three speed ranges: low speed, medium speed, and high speed. The test cycle lasts for 1800 seconds, with the low-speed range accounting for 37.4% of the time, the medium-speed range accounting for 38.5%, and the high-speed range accounting for 24.1%. The average vehicle speed is 29.0 km / h, the maximum vehicle speed is 114.0 km / h, and the idle speed ratio is 22.1%. The curve shape shows some borrowing from the WLTP approach. However, compared to WLTP, the ultra-high-speed driving portion has been removed, the proportion of medium- and high-speed driving is slightly reduced, and the maximum speed and frequency of change in the low- and medium-speed ranges are both lower. The idle speed ratio is also significantly higher, making it more consistent with Chinese road driving conditions.

[0096] Comparative Example 1

[0097] A method for recognizing braking intent in an electric vehicle. The difference between this comparative example and Example 1 is that step three is omitted; instead, the braking intensity Z is directly calculated using a braking intent recognition model without parameter optimization. The unoptimized parameters are applied to the Sugeno braking intent recognition model to generate the corresponding braking intent, such as... Figure 4 Membership degree and rule graph for braking intent recognition model.

[0098] Comparison of SOC changes between Example 1 and Comparative Example 1 under CLTC-P cycle conditions, and comparison of energy recovery. Figure 7 , Figure 8 As shown. In a single CLTC-P operating cycle, the initial SOC value is set at 95%. Regarding power consumption, in Example 1, the SOC decreased from 95% to 91.82%, resulting in a power consumption of 3.18%, achieving a power saving of 0.21% compared to Comparative Example 1. Regarding energy recovery, the control method provided in Example 1 recovered 1113.62 kJ of energy, which is 103.25 kJ more than in Comparative Example 1.

[0099] Analysis results show that the electric vehicle braking intention recognition method provided in Example 1 exhibits superior control performance under typical complex operating conditions of electric vehicles, achieves a higher energy recovery rate, and effectively extends the vehicle's driving range. In addition, the neural network module has online learning and dynamic adjustment capabilities, enabling the system to adaptively optimize control strategies based on different driving behaviors and environmental changes, further enhancing the intelligence and robustness of the control system.

Claims

1. A method for recognizing braking intent in an electric vehicle, characterized in that: include Collect vehicle parameters; the interval of each vehicle parameter is divided into multiple fuzzy subsets; A braking intent recognition model is constructed; the braking intent recognition model is based on a fuzzy neural network and includes a front-end network and a back-end network; the front-end network extracts the standardized activation intensity of each fuzzy rule according to vehicle parameters; the back-end network is used to combine the standardized activation intensity corresponding to each fuzzy rule with the back-end parameters to obtain the braking intensity; Parameter optimization of the braking intent recognition model: (1) Initialize the parameter vector of the braking intent recognition model and set the initial acceptance level; (2) Randomly adjust the parameter vector to generate candidate vectors; (3) Use the objective function to judge the prediction performance to verify the candidate vectors. If the candidate vector passes the verification, it is directly accepted as the parameter vector; if it fails the verification, it is accepted as the parameter vector according to the update probability. The update probability is positively correlated with the acceptance level; (4) Iteratively execute steps (2) and (3), and the acceptance level gradually decreases during the iteration process. Vehicle parameters are input into a parameter-optimized braking intent recognition model to obtain braking intensity.

2. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: The front-end network includes a front-end input layer, a front-end fuzzification layer, a rule layer, and a standardization layer. The front-end fuzzification layer is used to extract the membership degree of each vehicle parameter to the fuzzy subset. The rule layer is used to obtain the activation intensity of different fuzzy rules based on the membership degree of each vehicle parameter. The standardization layer is used to standardize the activation intensity of each fuzzy rule.

3. The method for recognizing braking intent in an electric vehicle according to claim 2, characterized in that: In the front-end fuzzification layer, each vehicle parameter has a membership function for each fuzzy subset; the membership function is a Gaussian function; the width parameter and center position parameter of each membership function are added to the parameter vector.

4. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: The expression for the normalization layer is: in, , The activation intensities are before and after standardization, respectively.

5. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: The backend network includes a backend input layer, a backend fuzzification layer, and a backend output layer connected in sequence; the backend fuzzification layer includes the same number of neural nodes as the number of fuzzy rules; each neural node in the backend fuzzification layer calculates the backend parameters corresponding to different fuzzy rules; the backend output layer uses the standardized activation intensity corresponding to each fuzzy rule as the weight, and performs a weighted summation of the backend parameters corresponding to each fuzzy rule as the braking intensity.

6. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: Candidate vectors in parameter optimization The expression is as follows: in, This is the current parameter vector; These are random numbers that follow a standard normal distribution. This is the step scaling value; For acceptability; ub is the current iteration number; lb is the upper bound of parameter adjustment; and lb is the lower bound of parameter adjustment. This is the initial value for acceptability.

7. The method for recognizing braking intent in an electric vehicle according to claim 6, characterized in that: The candidate vector verification process is as follows: comparing the parameter vector s with the candidate vector. The function value obtained by inputting the objective function , ;like If the verification is successful, then the probability is updated; otherwise, the probability is updated. .

8. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: The collected vehicle parameters have undergone noise reduction, outlier removal, and normalization.

9. The method for recognizing braking intent in an electric vehicle according to claim 1, characterized in that: The vehicle parameters include vehicle speed, remaining battery power, and brake pedal displacement.

10. A braking control system for an electric vehicle, characterized in that: It includes a parameter acquisition module, a braking intent recognition module, and a braking control module; the parameter acquisition module is used to acquire vehicle parameters in real time and input them into the braking intent recognition module; the braking intent recognition module is used to execute the electric vehicle braking intent recognition method as described in claim 1 to obtain a braking intensity signal; the braking control module controls the electric vehicle to brake according to the braking intensity signal.

Images