A New Energy Battery Load Optimization System and Method Based on an Adaptive Algorithm

Abstract

The present invention discloses a new energy battery load optimization system and method based on an adaptive algorithm, which relates to the technical field of battery control, and includes a multi-source acquisition module, an analysis and decision-making module, and an optimization and adjustment module; the multi-source acquisition module acquires driving state data, battery state data, and road surface state data; the analysis and decision-making module obtains the remaining battery power by combining the battery state data through the Kalman filter algorithm, adopts a fusion three-dimensional reconstruction algorithm to reconstruct the road surface state data into a road surface three-dimensional model, and obtains the road surface slope through the sampling analysis method; the optimization and adjustment module obtains the driving state data, the remaining battery power, and the road surface slope, performs fuzzy reasoning based on the membership function and fuzzy rules and defuzzifies to determine the optimal battery output power, comprehensively evaluates the fuzzy reasoning benefit based on the reward function, and optimizes the membership function parameters through reinforcement learning to realize the intelligent optimization matching of the battery output power with the driving state and the driving road conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of battery control, and particularly to a new energy battery load optimization system and method based on an adaptive algorithm. Background Art

[0002] Currently, new energy batteries, as the core power sources of electric two-wheelers and three-wheelers, are facing many problems.

[0003] On the one hand, the limited battery capacity leads to insufficient cruising range. On the other hand, the driving conditions of the vehicle are complex and diverse, involving frequent uphill and downhill on urban roads and different road surface conditions. At the same time, users hope that the vehicle has good power performance to meet the needs of acceleration and climbing, and also expect the battery to have a long cruising range to reduce the use cost.

[0004] Traditional battery load optimization schemes only consider the battery capacity for output power regulation, ignoring the impact of the driving state and complex and changeable driving conditions on the battery output power, and it is difficult to achieve an intelligent matching of the battery output power with the driving state and road conditions. There is room for optimization in terms of battery life and riding experience. Summary of the Invention

[0005] In view of the deficiencies of the existing technology, the present invention proposes a new energy battery load optimization system and method based on an adaptive algorithm to achieve intelligent battery output power regulation considering the driving state, battery state, and road surface state comprehensively.

[0006] The technical solution adopted to achieve the object of the present invention is as follows:

[0007] A new energy battery load optimization system based on an adaptive algorithm includes a multi-source acquisition module, an analysis and decision-making module, and an optimization and adjustment module;

[0008] The multi-source acquisition module receives acquisition signals and real-time collects driving state data, battery state data, and road surface state data;

[0009] The analysis and decision-making module obtains the remaining battery power based on the battery system state model and the battery state data through the Kalman filter algorithm and, by using the fusion three-dimensional reconstruction algorithm, integrates the texture and color information of the left-eye image of the road surface and the right-eye image of the road surface in the road surface state data into the road surface laser point cloud to generate a road surface three-dimensional model, obtains the road surface slope and the road surface bumpiness through the sampling analysis method and determines the generation time of the next acquisition signal;

[0010] The optimization and adjustment module obtains the driving state data and the remaining battery power​ and road surface gradient , perform fuzzy inference and defuzzification based on the membership function and fuzzy rules to determine the optimal battery output power , evaluate the benefits of fuzzy inference from multiple aspects based on the reward function, and online optimize the parameters of the membership function through reinforcement learning to improve the performance of fuzzy inference

[0011] Furthermore, the multi-source acquisition module includes a driving perception unit, a battery acquisition unit, and a road surface monitoring unit;

[0012] The driving sensing unit receives the acquisition signal and measures the driving speed through a Hall sensor, an acceleration sensor, and a GPS locator , driving acceleration and single-trip distance ;

[0013] The battery acquisition unit receives the acquisition signal and measures the output current through a Hall current sensor and a voltmeter and terminal voltage ;

[0014] The road surface monitoring unit receives the acquisition signal and obtains the left-eye road surface image through a multi-functional camera composed of a dual-view camera and a laser device , right-eye road surface image and road surface laser point cloud .

[0015] Furthermore, the analysis and decision-making module includes a battery estimation unit, a road surface analysis unit, and an adjustment decision-making unit;

[0016] The battery estimation unit obtains the output current at the th moment and terminal voltage , and uses the Kalman filtering algorithm to predict the estimated remaining battery power at the th moment based on the remaining battery power at the th moment and the battery system state model, and corrects it according to the estimated error covariance at the th moment to obtain the remaining battery power at the th moment ; The road surface analysis unit uses the fusion three-dimensional reconstruction algorithm to obtain from the left-eye road surface image and right-eye road surface image

[0017] ;

[0017] The road surface analysis unit uses the fusion three-dimensional reconstruction algorithm to obtain from the left-eye road surface image and right-eye road surface image Match the same type of features and generate the road surface image point cloud in combination with the internal reference matrix , and the road surface laser point cloud Perform registration and fusion to obtain the complete road surface point cloud , use the improved Poisson reconstruction to redirect the normal vector of the complete road surface point cloud And model it as a Gaussian Poisson problem to reconstruct and generate the 3D road surface model;

[0018] Adjust the decision-making unit to sample the 3D road surface model through sampling analysis, find the approximate plane that best fits the sampling points based on the least squares method, and determine the road surface slope , select mathematical indicators to quantitatively define the road surface bumpiness , based on the road surface bumpiness Determine the generation time of the next acquisition signal, where the mathematical indicators include standard deviation, mean error, and Gaussian curvature.

[0019] Furthermore, the battery estimation unit uses the Kalman filter algorithm to obtain the time remaining battery power The specific steps are as follows:

[0020] Establish a battery system state model to determine the estimated remaining battery power at the time , and the difference between the estimated remaining battery power at the time and the remaining battery power at the time is equal to the charge and discharge amount from the time to the time plus the process noise at the time that follows a Gaussian distribution , where the charge and discharge amount has a proportional relationship with the output current ;

[0021] Establish an observation equation, and the transformation relationship between the terminal voltage at the time and the remaining battery power is established through an observation transformation function obtained by pre-fitting or based on an electrochemical model plus the observation noise that follows a Gaussian distribution ; Build;

[0022] Superimpose the estimated error covariance at the time with the and the Moment process noise Obtain the moment predicted estimated error covariance , and for the observation transformation function with respect to the remaining battery power take the derivative to obtain the Jacobian matrix ;

[0023] Through the Jacobian matrix map the predicted estimated error covariance at the moment to the observation space, and calculate the Kalman gain at the moment that can measure the correction contribution degree between the estimated remaining battery power and the terminal voltage ; moment ;

[0024] Based on the error between the terminal voltage and the mapped value after transformation of the estimated remaining battery power through the observation transformation function , combined with the Kalman gain correct to obtain the remaining battery power at the moment , and update through the Kalman gain and the predicted estimated error covariance to obtain the estimated error covariance at the moment

[0025] Furthermore, the road surface analysis unit adopts a fusion three-dimensional reconstruction algorithm to generate a road surface three-dimensional model, including the following specific steps:

[0026] Based on the left-eye image of the road surface through the stable feature acceleration algorithm and the right-eye image of the road surface establish a left-eye scale space and a right-eye scale space, find local maxima and eliminate weak response feature points to achieve feature point positioning, where the weak response feature points are the feature points with the determinant value of the Hessian matrix less than the response threshold;

[0027] Statistically calculate the gradient histogram in the neighborhood of a single feature point to determine the main direction of the single feature point, extract sixteen regional blocks around the single feature point, and statistically calculate the amplitude information of the eight gradient directions in each regional block to construct the feature description of the single feature point;

[0028] ​​​​​The distance metric method is used to compare the feature descriptions of the feature points in the left-eye scale space and the right-eye scale space to generate multiple groups of matching point pairs, and the fundamental matrix that maximizes the total number of matching point pairs that meet the geometric constraints is selected by the random sample consensus algorithm;

[0029] Calculate the internal parameter matrix, remove the influence of the internal parameter matrix on the fundamental matrix through bundle adjustment, obtain the external parameter matrix, and combine the principle of triangulation to perform projective transformation on the matching point pairs to generate the road surface image point cloud , and through the iterative closest point algorithm and the road surface laser point cloud Perform registration and fusion to obtain the complete road surface point cloud ;

[0030] Use improved Poisson reconstruction to redirect the normal vectors in the complete road surface point cloud and regard the 3D reconstruction as a Gaussian Poisson problem to solve and restore to generate the 3D model of the road surface.

[0031] Furthermore, the stable feature acceleration algorithm includes the following specific steps:

[0032] Calculate the determinant of the Hessian matrix of the road surface image at different scales to construct the corresponding scale space, and the road surface image includes the left-eye image of the road surface and the right-eye image of the road surface ;

[0033] In the scale space, compare the determinant values of the Hessian matrix of each pixel point and its neighborhood pixel points, find the pixel point with the largest determinant value of the Hessian matrix in the neighborhood and regard it as a potential feature point;

[0034] Eliminate all potential feature points with determinant values of the Hessian matrix less than the response threshold, and use sub-pixel analysis method to correct the pixel positions of the retained feature points by interpolation based on the pixel distribution around the feature points.

[0035] Furthermore, using improved Poisson reconstruction to restore the complete road surface point cloud to the 3D model of the road surface includes the following specific steps:

[0036] Smooth the complete road surface point cloud by moving least squares , and calculate the normal vector and point cloud curvature of each point in the complete road surface point cloud by principal component analysis, take the point with the minimum point cloud curvature as the origin and specify the positive direction;

[0037] Perform neighborhood search through the KD tree, successively perform the dot product of the several points found with the origin, reverse the normal vector direction of the points with dot product less than 0, and mark the points that have completed the dot product;

[0038] Take the last point in the neighborhood that has completed the dot product as the new origin, conduct a new round of neighborhood search through the KD tree, continue to perform dot products with unlabeled points to reverse or maintain the normal vector, and mark them;

[0039] Until all points in the complete road surface point cloud are marked, the normal vector redirection is completed. Based on the redirected complete road surface point cloud construct an octree topological relationship so that all points in the complete road surface point cloud fall on the leaf nodes;

[0040] Based on the three-dimensional box filter, construct the node function of the leaf nodes, linearly combine all node functions through cubic spline interpolation to construct a vector field, construct a Gaussian Poisson problem and transform it into a system of linear equations, solve it by the conjugate gradient method to obtain a three-dimensional scalar field, and use the marching cubes algorithm to extract the isosurface from the three-dimensional scalar field to generate a three-dimensional road surface model.

[0041] Furthermore, adjust the decision unit to obtain the road surface slope through sampling analysis and the road surface bumpiness and the generation time of the next acquisition signal, including the following specific steps:

[0042] Perform random sampling on the three-dimensional road surface model to obtain sampling points, where

[0043] Define an approximate plane equation, and construct an error function with the goal of minimizing the sum of the squares of the distances from sampling points to the approximate plane, where , and are the equation coefficients of the approximate plane respectively;

[0044] Solve the error function with respect to , and , rewrite it in matrix form and solve it through matrix operations to obtain the equation coefficients , and of the approximate plane;

[0045] Determine the normal vector based on the approximate plane equation, then the road surface slope is equal to the tangent value of the angle between the normal vector and the axis direction;

[0046] Define the road surface bumpiness by selecting a mathematical index that can reflect the undulation degree of each point on the three-dimensional road surface model to an approximate plane. , the mathematical index includes the standard deviation of the distances of sampling points relative to the approximate plane, the average error between the local slope of sampling points and the road surface slope

[0047] Compare the road surface bumpiness with the bumpiness threshold and adaptively determine the acquisition interval . The generation time of the next acquisition signal is equal to the generation time of the current acquisition signal plus the acquisition interval .

[0048] Furthermore, compare the road surface bumpiness with the bumpiness threshold and adaptively determine the acquisition interval , including the following specific steps:

[0049] Judge whether the road surface bumpiness is greater than or equal to the bumpiness threshold ;

[0050] If it is greater than or equal to the bumpiness threshold , calculate the difference between the road surface bumpiness and the bumpiness threshold , and calculate the acquisition interval using the bump acquisition relationship function . The dependent variable of the bump acquisition relationship function is the acquisition interval , and the independent variable is the difference between the road surface bumpiness and the bumpiness threshold , showing a monotonically decreasing trend and the value range is greater than or equal to the minimum acquisition interval ;

[0051] If the road surface bumpiness is less than the bumpiness threshold , then the acquisition interval is the value of the bump acquisition relationship function when the road surface bumpiness is equal to the bumpiness threshold .

[0052] Furthermore, the optimization and adjustment module includes a fuzzy inference unit and an online optimization unit;

[0053] The fuzzy inference unit selects an appropriate membership function according to the characteristics of fuzzy input and fuzzy output during the offline training phase, establishes fuzzy rules considering domain experience and determines the membership function parameters, and calculates the driving speed respectively according to the membership function 、 Driving acceleration 、 Remaining battery power and road surface gradient For the membership degree of the fuzzy set, perform linguistic fuzzy inference (Mamdani inference) based on fuzzy rules and defuzzify through the centroid method to generate the optimal battery output power ;

[0054] The online optimization unit obtains the driving speed 、 Driving acceleration 、 Single driving distance and remaining battery power , comprehensively consider battery energy consumption, driving safety and driving comfort to design a reward function to evaluate the benefits of fuzzy inference, and use reinforcement learning to online optimize the membership function parameters based on the reward function to improve the performance of fuzzy inference.

[0055] Furthermore, in the offline training stage, define the fuzzy sets of fuzzy inputs and fuzzy outputs and select the corresponding membership functions, including:

[0056] Define the speed fuzzy set of the driving speed including low speed 、 Medium speed and high speed , and the driving speed has a high degree and stability of belonging to a specific speed fuzzy set within a specific speed range, and select a trapezoidal membership function to describe the speed fuzzy set;

[0057] Define the acceleration fuzzy set of the driving acceleration including negative acceleration 、 Zero acceleration and positive acceleration , and the driving acceleration has a high and stable membership degree for negative acceleration or positive acceleration , and only when the driving acceleration is near 0, the membership degree for zero acceleration shows a trend of decreasing from the center of 0 to both sides. Select a trapezoidal membership function to describe negative acceleration and positive acceleration , and select a triangular membership function to describe zero acceleration ;

[0058] Define the power fuzzy set of the remaining battery power including low power 、 Medium power and high power , since the remaining battery power During driving, it shows a monotonically decreasing trend over time. A Sigmoid membership function with unilateral characteristics is selected to describe the power fuzzy set.

[0059] Define the road surface gradient The gradient fuzzy set of includes downhill flat road and uphill . Since the road surface gradients in plain areas are mostly close to flat roads, and there are corresponding standards for the road surface gradients in mountainous areas , the smaller the probability of occurrence of a larger road surface gradient

[0060] . A Gaussian membership function is selected to describe the gradient fuzzy set. Define the power fuzzy set of the battery output power includes low power medium power and high power . Since the battery output power

[0061] transitions smoothly between different power fuzzy sets and there are certain differences in the battery output power ranges of different vehicle models, a bell-shaped membership function with a clear center, adjustable width, and smooth curve is selected to describe the power fuzzy set. Specifically, the fuzzy rules are designed considering domain experience, and the mutual relationship between the four fuzzy inputs is combined to judge and make a decision on the battery output power

[0062] . A total of eighty-one fuzzy rules are established, mapping each of the eighty-one input combinations to one of the three fuzzy sets corresponding to the fuzzy output. Furthermore, the fuzzy inference unit performs Mamdani inference based on the fuzzy rules and defuzzifies through the centroid method to determine the optimal battery output power

[0063] including the following specific steps: Obtain the driving speed driving acceleration remaining battery power and road surface gradient

[0064] ; Traverse the eighty-one fuzzy rules. For a single fuzzy rule, determine the four fuzzy sets in the corresponding input combination, and calculate the membership degrees of the driving speed driving acceleration remaining battery power and road surface gradient respectively with the corresponding fuzzy sets, and take the minimum membership degree as the activation strength of the single fuzzy rule.

[0065] Activation intensity based on a single fuzzy rule Modify the membership function of the power fuzzy set corresponding to the fuzzy rule, and perform a maximum composition on the modified power membership functions of 81 rules to generate a combined power membership function ;

[0066] Defuzzify the combined power membership function using the centroid method. Consider the combined power membership function and the battery output power The area enclosed by the independent variable coordinate axis where it is located is regarded as a planar object, and the abscissa value of the centroid of the planar object is the optimal battery output power , where the combined power membership function and the area of the abscissa is the integral of the combined power membership function with respect to the battery output power .

[0067] Furthermore, the design of the reward function comprehensively considers battery energy consumption, driving safety, and driving comfort, including the following specific steps:

[0068] Obtain the driving speed at time , driving acceleration , single-trip distance , and remaining battery charge , as well as the driving acceleration at time and the remaining battery charge ; ;

[0069] Consider the power consumption per unit distance from time to time to time to time . That is, the power consumption divided by the single-trip distance . The smaller the power consumption per unit distance, the higher the battery energy consumption reward. The battery energy consumption reward function at time is set in the exponential form of the reciprocal of the power consumption per unit distance;

[0070] Consider that the driving speed at time and the driving acceleration do not exceed the safe speed limit and the safe acceleration limit respectively. When the th Time travel speed and travel acceleration are respectively less than the upper limit of the safe speed and the upper limit of the safe acceleration When, the time travel safety reward function is set to the opposite form of the logarithmic function so that the travel safety reward value decreases as the travel speed and travel acceleration increase, and is more sensitive when approaching the upper limit of the safe speed and the upper limit of the safe acceleration When, a penalty factor time travel speed is greater than or equal to the upper limit of the safe speed and / or travel acceleration is greater than or equal to the upper limit of the safe acceleration When, a penalty factor is introduced and is set in combination with the transformation form of the Sigmoid function, so that the greater the degree of speeding, the faster the penalty increases;

[0071] Calculate the time to the time acceleration change value When, the time to the time acceleration change value exceeds the acceleration change upper limit When, the travel comfort reward decreases, and the comfort decreases smoothly as the acceleration change value gradually increases. The time travel comfort reward function is set to the hyperbolic tangent function form with the percentage error between the absolute acceleration change value and the acceleration change upper limit as the independent variable;

[0072] The time reward function is equal to the battery energy consumption reward function the travel safety reward function and the travel comfort reward function The linear weighted sum, where the linear weights determine the proportion of importance for battery energy consumption, driving safety, and driving comfort.

[0073] Specifically, reinforcement learning defines the state and the environmental state. The state includes the driving speed , driving acceleration , single-trip distance and remaining battery power , and the environmental state is the road surface gradient . The action space is defined as a set of adjustable battery output powers . The action is defined as adjusting all membership function parameters in fuzzy inference. The reward function is used as the Q value of a deep Q-network with an experience replay mechanism. The deep Q-network outputs the probability distribution of actions based on a set of environmental states and states. The proximal policy optimization algorithm selects different actions based on the probability distribution of actions, evaluates the impacts of different actions in parallel based on the Q value, and determines the optimal action through truncated importance sampling, so that the Q value continuously increases and finally converges.

[0074] A new energy battery load optimization method based on an adaptive algorithm, implemented based on a new energy battery load optimization system based on an adaptive algorithm, includes the following specific steps:

[0075] Adopt the Kalman filter algorithm to obtain the remaining battery power at the moment based on the battery system state model combined with the battery state data collected at the moment ; ;

[0076] Adopt the fusion three-dimensional reconstruction algorithm to integrate the texture information and color information of the left-eye road image and right-eye road image captured at the moment into the road surface laser point cloud in the road surface state data at the moment to generate the road surface three-dimensional model at the moment . Obtain the road surface gradient and road surface bumpiness at the moment through sampling analysis method and adjust the acquisition interval ; ;

[0077] Use reinforcement learning to comprehensively evaluate the moment The riding benefits brought by fuzzy inference are used to adjust the membership function parameters by the proximal policy optimization algorithm;

[0078] Based on the moment The collected driving state data, remaining battery power and road surface gradient , through the membership function and fuzzy rules, perform fuzzy inference and defuzzification to determine the moment optimal battery output power .

[0079] Compared with the prior art, the significant advantages of the present invention are that by analyzing the real-time collected driving state data, battery state data and road surface state data through the Kalman filter algorithm and the fusion three-dimensional reconstruction algorithm, an accurate remaining battery power and a three-dimensional road surface model are obtained, the three-dimensional road surface model is analyzed to adaptively determine the acquisition interval and infer the road surface gradient, and fuzzy inference and defuzzification are performed on the driving state data, remaining battery power and road surface gradient according to the membership function and fuzzy rules to determine the optimal battery output power suitable for the current road conditions and driving requirements, and based on the reward function, the benefits of fuzzy inference are evaluated from multiple aspects, and the membership function parameters are optimized online through reinforcement learning to improve the fuzzy inference performance, so that the electric two-wheeler and / or electric three-wheeler always balance the battery life and riding experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a schematic diagram of a new energy battery load optimization system based on an adaptive algorithm in the present invention;

[0081] Figure 2 It is a flow chart of an improved Poisson reconstruction in the present invention;

[0082] Figure 3 It is a fuzzy rule table established in the present invention;

[0083] Figure 4 It is a flow chart of a new energy battery load optimization method based on an adaptive algorithm in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0085] Embodiment 1

[0086] As Figure 1 shown, a specific embodiment of the present invention discloses a new energy battery load optimization system based on an adaptive algorithm, including a multi-source acquisition module, an analysis and decision-making module, and an optimization and adjustment module;

[0087] The multi-source acquisition module receives acquisition signals and collects driving state data, battery state data, and road surface state data in real time through intelligent sensors. The driving state data includes driving speed , driving acceleration and single-trip distance . The battery state data includes output current and terminal voltage . The road surface state data includes left-eye road surface image , right-eye road surface image and road surface laser point cloud . Among them, the single-trip distance refers to the driving distance of an electric two-wheeler or electric three-wheeler during the process from the previous acquisition to the current acquisition;

[0088] The analysis and decision-making module accurately obtains the remaining battery power through the Kalman filtering algorithm based on the battery system state model combined with the battery state data . It adopts a fusion three-dimensional reconstruction algorithm. Taking the road surface laser point cloud as a reference, it fuses the texture information and color information in the left-eye road surface image and the right-eye road surface image to generate a road surface three-dimensional model, and further obtains the road surface slope and road surface bumpiness through the sampling analysis method, and determines the generation moment of the next acquisition signal based on the road surface bumpiness ;

[0089] The optimization and adjustment module obtains the driving state data, remaining battery power and road surface slope , conducts fuzzy inference based on the membership function and fuzzy rules, determines the optimal battery output power through defuzzification and makes adjustments. It designs a reward function to evaluate the driving benefits brought by fuzzy inference, and conducts online learning through reinforcement learning to optimize the membership function parameters and improve the fuzzy inference performance.

[0090] Furthermore, the multi-source acquisition module includes a driving perception unit, a battery acquisition unit, and a road surface monitoring unit;

[0091] The driving sensing unit receives acquisition signals. The Hall sensor generates pulse signals based on the Hall effect when the wheel rotates, and obtains the driving speed based on the product of the pulse signal frequency and the wheel circumference . The acceleration sensor measures the driving acceleration by sensing the charge signal characteristics generated when the piezoelectric material is stressed and deformed . The GPS locator records the driving trajectory and adds the recording point when receiving the acquisition signal, and obtains the trajectory distance between the recording point and the previous recording point as the single-trip distance ;

[0092] The battery acquisition unit receives the acquisition signal and measures the output current through a Hall current sensor and a voltmeter and the terminal voltage ;

[0093] The road surface monitoring unit receives the acquisition signal and obtains the left-eye image of the road surface through a multi-functional camera , the right-eye image of the road surface and the road surface laser point cloud , where the multi-functional camera consists of a dual-view camera and a laser device. The dual-view camera is used to capture the road surface ahead to obtain the left-eye image of the road surface and the right-eye image of the road surface . The laser device emits a laser beam and receives the reflected laser signal to measure the distance between the laser beam and the road surface, thereby generating the road surface laser point cloud .

[0094] Furthermore, the analysis and decision-making module includes a battery estimation unit, a road surface analysis unit, and an adjustment decision-making unit;

[0095] The battery estimation unit obtains the output current at the moment and the terminal voltage and performs analysis using the Kalman filter algorithm. Based on the remaining battery power at the moment and the battery system state model, predicts the estimated remaining battery power at the moment . According to the estimated error covariance at the moment and the observation noise , it is corrected to obtain the remaining battery power at the moment ;

[0096] For the problem of holes and texture loss in the 3D reconstruction based on the road surface laser point cloud , the road surface analysis unit adopts a fusion 3D reconstruction algorithm to match the same type of features from the left-eye image of the road surface and the right-eye image of the road surface , calculates the internal parameter matrix and generates the road surface image point cloud , and registers and fuses it with the road surface laser point cloud to obtain the complete road surface point cloud . Using the improved Poisson reconstruction, the complete road surface point cloud The three-dimensional reconstruction is regarded as a Gaussian-Poisson problem to reconstruct and generate a three-dimensional model of the road surface;

[0097] The adjustment decision unit randomly samples the three-dimensional model of the road surface through the sampling analysis method, and finds the approximate plane with the minimum sum of the squares of the distances from the sampling points through the least squares method and determines the road surface slope and selects appropriate mathematical indicators to quantitatively define the road surface bumpiness Based on the road surface bumpiness and the bumpiness threshold to determine the generation moment of the next acquisition signal, where the mathematical indicators include standard deviation, mean error, and Gaussian curvature.

[0098] Furthermore, the battery estimation unit uses the Kalman filter algorithm to obtain the moment remaining battery power including the following specific steps:

[0099] Establish a battery system state model to describe the state change of the remaining battery power Considering the linear relationship between the remaining battery power and the output current during the charging and discharging process of the battery, therefore, the estimated remaining battery power at the moment and the remaining battery power at the moment The difference is approximately equal to the charge and discharge amount from the moment to the moment Since the power output of the power supply is not adjusted from the moment to the moment Therefore, the output current is default unchanged and always equal to the output current at the moment Then the charge and discharge amount from the moment to the moment is proportional to the output current at the moment The specific formula is as follows:

[0100] ;

[0101] Where and are the rated battery capacity and the charge and discharge efficiency, respectively, is the time to the time acquisition interval, is the time process noise, which describes the uncertainty of the battery charge and discharge process, obeys a Gaussian distribution with a mean of 0 and a variance of ;

[0102] An observation equation is established to describe the transformation relationship between the terminal voltage at time that can be directly observed and the remaining battery capacity , and the specific formula is , where is the observation transformation function with respect to the remaining battery capacity , which can be obtained by fitting the pre-observed data or determined based on the electrochemical model of the battery, is the observation noise at time which describes the error in the measurement process, and obeys a Gaussian distribution with a mean of 0 and a variance of ; The estimated error covariance

[0103] at time is superimposed with the process noise at time as the predicted estimated error covariance at time , and the Jacobian matrix is obtained by taking the derivative of the observation transformation function with respect to the remaining battery capacity , and the Jacobian matrix reflects the influence degree of the remaining battery capacity on the terminal voltage ; The predicted estimated error covariance at time is mapped to the observation space through the Jacobian matrix to calculate the Kalman gain

[0104] at time , and the Kalman gain ​​​​​​Used to measure the moment estimated remaining battery power and the contribution degree to the correction of the terminal voltage When the moment predicted estimated error covariance is greater than the observation noise , it indicates that the uncertainty of the estimation is greater than the uncertainty of the observation, and the Kalman gain increases. At the moment the terminal voltage is more accurate and has a higher contribution degree to the correction. When the moment predicted estimated error covariance is less than the observation noise , it indicates that the uncertainty of the estimation is less than the uncertainty of the observation, and the Kalman gain decreases. At the moment the estimated remaining battery power has a higher contribution degree to the correction; is the transpose of the Jacobian matrix ;

[0105] Combined with the Kalman gain at the moment and the terminal voltage to correct the estimated remaining battery power , the remaining battery power at the moment is obtained. The estimated remaining battery power is substituted into the observation transformation function , and the mapping value after substitution is used. The estimated remaining battery power is corrected by the error between the terminal voltage and the mapping value to make the error of the remaining battery power smaller;

[0106] Based on the Kalman gain at the moment and the predicted estimated error covariance , the estimated error covariance at the moment is updated and obtained, where represents the identity matrix. When the Kalman gain is smaller, it indicates that at the moment The estimated error covariance is smaller, that is, the uncertainty of the estimation is gradually decreasing.

[0107] Furthermore, the road surface analysis unit adopts a fusion three-dimensional reconstruction algorithm to generate a road surface three-dimensional model, which includes the following specific steps:

[0108] Based on the left-eye image of the road surface through the stable feature acceleration algorithm and the right-eye image of the road surface respectively establish the corresponding left-eye scale space and right-eye scale space, find the local maximum value of the scale space and eliminate the feature points with weak responses to achieve sub-pixel level feature point positioning;

[0109] For a single feature point in the left-eye scale space or the right-eye scale space, the gradient histogram is statistically calculated within the neighborhood of the single feature point, and the direction exceeding 0.8 times the maximum value of the gradient histogram is used as the main direction of the single feature point. The determination of the main direction makes the single feature point have rotational invariance, that is, no matter the left-eye image of the road surface or the right-eye image of the road surface how to rotate, the feature description in the main direction always remains consistent;

[0110] Extract around the single feature point A total of sixteen regional blocks, and the amplitude information of eight gradient directions within each regional block is statistically calculated to construct a -dimensional vector as the feature description of the single feature point, and the feature description comprehensively reflects the gradient distribution information of the single feature point;

[0111] The distance measurement method is used to compare and match the feature descriptions of the feature points in the left-eye scale space and the right-eye scale space. That is, for a single feature point in the left-eye scale space, find the feature point with the closest distance and the feature point with the second-closest distance in the right-eye scale space and calculate the distance ratio. If the distance ratio is less than the matching threshold, then the feature point with the closest distance in the right-eye scale space is matched with the single feature point in the left-eye scale space to generate a matching point pair. If the distance ratio is greater than or equal to the matching threshold, it is regarded as a false match;

[0112] Use the random sample consensus algorithm to extract at least eight groups of matching point pairs and calculate the fundamental matrix, and use the fundamental matrix to verify the geometric constraints of the remaining matching point pairs. Statistically calculate the total number of matching point pairs that meet the geometric constraints. Repeat this step multiple times, and select the fundamental matrix that makes the total number of matching point pairs that meet the geometric constraints the largest. Among them, the fundamental matrix is used to describe the projection transformation relationship from the right-eye view to the left-eye view of the binocular camera, and is used to unify the left-eye scale space and the right-eye scale space;

[0113] Calculate the intrinsic matrix of the left-eye camera based on the focal length of the left-eye camera of the dual-view camera. Through bundle adjustment, strip the influence of the intrinsic matrix on the fundamental matrix, and perform the best estimation of the extrinsic matrix of the dual-view camera. The extrinsic matrix describes the pose change of the right-eye camera relative to the left-eye camera. Bundle adjustment is an existing algorithm;

[0114] Based on the extrinsic matrix, use the principle of triangulation to perform projective transformation on the matching point pairs, calculate the three-dimensional coordinates of the matching point pairs in the world coordinate system, and obtain the road surface image point cloud , and further through the iterative closest point algorithm and the road surface laser point cloud Perform registration and fusion to obtain the complete road surface point cloud ;

[0115] Use improved Poisson reconstruction to correct the ambiguity problem of the normal vector in the complete road surface point cloud Select the point with the minimum point cloud curvature as the origin, search for neighboring points through the KD tree and redefine the normal vector direction, and propagate with the neighboring points as the origin. Consider the three-dimensional reconstruction of the complete road surface point cloud with the normal vector redirected as a Gaussian Poisson problem, solve and restore to generate the road surface three-dimensional model.

[0116] Furthermore, the stable feature acceleration algorithm includes the following specific steps:

[0117] Calculate the determinant of the Hessian matrix of the road surface image at different scales to construct the scale space. The Hessian matrix is used to describe the second-order derivative information of the road surface image in different directions. The road surface image includes the left-eye road surface image and the right-eye road surface image , and the scale space includes the left-eye scale space and the right-eye scale space;

[0118] In the scale space, compare the determinant values of the Hessian matrix of each pixel point of the road surface image with those of the neighboring pixel points, and find the pixel point with the largest determinant value of the Hessian matrix in the neighborhood and regard it as a potential feature point. Among them, the neighborhood refers to the set of pixel points that are adjacent in scale and pixel position to the pixel point;

[0119] For potential feature points, further eliminate all potential feature points whose determinant values of the Hessian matrix are less than the response threshold. These potential feature points with weak responses may be caused by noise or interference and have a negative impact on feature point matching and three-dimensional reconstruction;

[0120] Use sub-pixel analysis method based on the pixel distribution around the feature point, and use interpolation to correct the pixel position of the retained feature point to achieve feature point localization.

[0121] As Figure 2 shown, furthermore, use improved Poisson reconstruction to process the complete road surface point cloud Restoring to a three-dimensional road surface model includes the following specific steps:

[0122] Smoothing the complete road surface point cloud by moving least squares method, and calculating the normal vector and point cloud curvature of each point in the complete road surface point cloud by principal component analysis method;

[0123] Taking the point with the minimum point cloud curvature as the origin, and specifying the normal vector direction of the origin as the positive direction;

[0124] Performing neighborhood search through KD tree, successively taking the dot product of the several searched points with the origin. If the dot product is less than 0, reversing the normal vector direction of the point; if the dot product is greater than or equal to 0, keeping the original normal vector direction, and marking the points that have completed the dot product;

[0125] Taking the last point that has completed the dot product in the neighborhood as the new origin, performing a new round of neighborhood search through KD tree and excluding the points with existing marks, continuing to perform the dot product to reverse or maintain the normal vector direction, and making marks;

[0126] Until all points in the complete road surface point cloud are marked, the normal vector redirection is completed. Further, constructing an octree topological relationship for the complete road surface point cloud with completed redirection through Poisson distribution, so that all points in the complete road surface point cloud fall on the leaf nodes at an appropriate depth in the octree;

[0127] Using a three-dimensional box filter as the basis function to construct the node function of the leaf node, and linearly combining all node functions through cubic spline interpolation to construct a vector field, making the vector field match the normal vector of each point. The three-dimensional box filter is an existing filter;

[0128] Reformulating the three-dimensional reconstruction as a Gaussian Poisson problem, further transforming it into a system of linear equations by finite difference or finite element method, and solving it by conjugate gradient method to obtain a three-dimensional scalar field. Further, using the marching cubes algorithm to extract the isosurface from the three-dimensional scalar field to form a three-dimensional road surface model. This step is a commonly used technical means in traditional Poisson reconstruction.

[0129] Furthermore, adjusting the decision unit to obtain the road surface slope

[0130] ​​​​​​​​

[0131] Define the approximate plane equation , with the goal of finding an approximate plane that best fits the overall trend of the road surface three-dimensional model, that is the sum of the squares of the distances from the sampling points to the approximate plane is minimized, and an error function is constructed , and are the equation coefficients of the approximate plane respectively, are the -th sampling point's axis, axis and axis values respectively;

[0132] Solve the partial derivatives of the error function with respect to , and and set them equal to 0, and simplify and rewrite them in matrix form as follows:

[0133] ;

[0134] Obtain the equation coefficients , and of the approximate plane through matrix operations, where matrix operations include Gaussian elimination method, LU decomposition method and QR decomposition method;

[0135] The normal vector of the approximate plane, then the tangent value of the angle between the normal vector of the approximate plane and the axis direction can be regarded as the road surface slope of the road surface three-dimensional model. Based on the trigonometric function relationship, the road surface slope can be deduced;

[0136] Define the road surface roughness as any mathematical index that can reflect the undulation degree of each point on the road surface three-dimensional model to the approximate plane. The larger the mathematical index, the greater the undulation degree, and the more bumpy the road surface three-dimensional model is relative to the smooth approximate plane. The mathematical index includes the standard deviation of the distances of the sampling points to the approximate plane, the average error between the local slopes of the sampling points and the road surface slope

[0137] Compare the road surface roughness and the roughness threshold and adaptively determine the acquisition interval , then the generation time of the next acquisition signal is equal to the generation time of the current acquisition signal plus the acquisition interval. , where the bumpiness threshold is determined according to the mathematical index selected when defining the road surface bumpiness . Taking the generation time of the current acquisition signal as the moment as an example, according to the moment of the road surface slope , the acquisition interval can be determined. Then the generation time of the next acquisition signal is the moment .

[0138] Further, compare the road surface bumpiness with the bumpiness threshold and adaptively determine the acquisition interval , including the following specific steps:

[0139] Judge whether the road surface bumpiness is greater than or equal to the bumpiness threshold ;

[0140] If the road surface bumpiness is greater than or equal to the bumpiness threshold , calculate the difference between the road surface bumpiness and the bumpiness threshold , and calculate the acquisition interval using the bump acquisition relationship function. The bump acquisition relationship function requires that the acquisition interval shows a monotonically decreasing trend with respect to the difference between the road surface bumpiness and the bumpiness threshold , that is, the more bumpy the road surface, the smaller the acquisition interval, but it cannot be less than the minimum acquisition interval . The bump relationship function can be selected in various forms. In this embodiment, an inverse proportional function is selected, and the specific formula is as follows:

[0141] ;

[0142] Where and are predetermined constants, which are taken as 2 and 0.4 respectively in this embodiment;

[0143] If the road surface bumpiness is less than the bumpiness threshold , it is considered that the road surface three-dimensional model is flat, then the acquisition interval is directly selected as when the road surface bumpiness is equal to the bumpiness threshold The function value of the bump collection relationship at that time, that is, the collection interval 。

[0144] Furthermore, the optimization and adjustment module includes a fuzzy inference unit and an online optimization unit;

[0145] The fuzzy inference unit selects corresponding membership functions respectively according to the characteristics of fuzzy input and fuzzy output during the offline training stage, establishes fuzzy rules from fuzzy input to fuzzy output considering domain experience and determines membership function parameters, and obtains the driving speed 、driving acceleration 、remaining battery power and road surface slope , calculates the membership degrees of the driving speed 、driving acceleration 、remaining battery power and road surface slope for the fuzzy sets respectively, conducts Mamdani inference based on the fuzzy rules, and determines the optimal battery output power through defuzzification by the centroid method ;

[0146] The online optimization unit obtains the driving speed 、driving acceleration 、single driving distance and remaining battery power , designs a reward function considering battery energy consumption, driving safety and driving comfort to evaluate the driving benefits brought by fuzzy inference, adaptively optimizes the membership function parameters determined in the offline training stage based on the reward function using reinforcement learning, and continuously improves the fuzzy inference benefits through online learning to approach the benefit upper limit.

[0147] Even further, during the offline training stage, when the fuzzy input is the driving speed , it is defined that the speed fuzzy set of the driving speed includes low speed 、medium speed and high speed , and the degree and stability of the driving speed of the electric two-wheeler and / or electric three-wheeler belonging to the specific speed fuzzy set within a specific speed range are relatively high. The speed membership function of the driving speed selects a trapezoidal membership function, and the specific formula is as follows:

[0148] ;

[0149] Among them, and are the low speed 、medium speed and high speed The corresponding speed membership function and are the boundary values of the speed fuzzy set respectively. In this embodiment, considering the actual driving speed range of electric two-wheelers and / or electric three-wheelers, and take 15, 20, 30, 35 respectively, and the unit is .

[0150] Furthermore, in the offline training stage, when the fuzzy input is the driving acceleration , the acceleration fuzzy set of the driving acceleration is defined to include negative acceleration , zero acceleration and positive acceleration . And the driving acceleration of electric two-wheelers and / or electric three-wheelers has a high degree and stability in belonging to negative acceleration or positive acceleration . When and only when the driving acceleration belongs to zero acceleration in a small range near 0, and the membership degree decreases from the center of 0 to both sides. Therefore, the acceleration membership functions of negative acceleration and positive acceleration of the driving acceleration are selected as trapezoidal membership functions, while the acceleration membership function of zero acceleration is selected as a triangular membership function. The specific formulas are as follows:

[0151] ;

[0152] Among them, and are the acceleration membership functions corresponding to negative acceleration , zero acceleration and positive acceleration respectively. and are the boundary values of the acceleration fuzzy set respectively. In this embodiment, considering the actual driving acceleration range of electric two-wheelers and / or electric three-wheelers, and take -0.5, -1, 0.5 and 1 respectively, and the unit is .

[0153] Furthermore, in the offline training stage, when the fuzzy input is the remaining battery power , the power fuzzy set of the remaining battery power is defined to include low power , medium power and high battery level , since the remaining battery level shows a monotonically decreasing trend over time during the operation of electric two-wheelers and / or electric three-wheelers, the membership function of the remaining battery level selects a sigmoid membership function with one-sidedness. The specific formula is as follows:

[0154] ;

[0155] ;

[0156] ;

[0157] where, and are the membership functions of the low battery level , medium battery level and high battery level respectively. and are the slope factors respectively, which determine the steepness of the membership function of the battery level. The coefficients and are the central values of the battery level ranges of the low battery level and high battery level respectively. and are the central value and half of the width of the medium battery level range respectively. In this embodiment, and both take the same value of 0.1, indicating that the attenuation trend of the remaining battery level is the same in different battery level fuzzy sets and take and respectively as the rated battery capacity.

[0158] Furthermore, in the offline training stage, when the fuzzy input is the road slope , the slope fuzzy set of the road slope is defined to include downhill , flat road and uphill . Based on the clear design standards and specification requirements in road design, the road slope in plain areas is mostly close to flat road, and there are also corresponding standards for the road slope in mountainous areas. The larger the road slope , the smaller the occurrence probability. Therefore, the membership function of the road slope selects a Gaussian membership function. The specific formula is as follows:

[0159] ;

[0160] Among them, and are the slope membership functions corresponding to downhill , flat road and uphill respectively, and are the typical slopes in downhill , flat road and uphill respectively, and are the slope variances in downhill , flat road and uphill respectively, which determine the slope range of the slope fuzzy set. In this embodiment, and are taken as -0.05, 0 and 0.05 respectively, and are taken as 0.02, 0.01 and 0.02 respectively.

[0161] Furthermore, in the offline training stage, the fuzzy output is the battery output power . Define that the power fuzzy set of the battery output power includes low power , medium power and high power . Since the transition of the battery output power between different power fuzzy sets is smooth, the power membership function of the battery output power selects a bell-shaped membership function with a clear center, adjustable width and smooth curve. The specific formula is as follows:

[0162] ;

[0163] Among them, and are the power membership functions corresponding to low power , medium power and high power respectively, and are the center output powers of low power , medium power and high power respectively, and are the center output powers of low power , medium power and high power The width of the medium power membership function determines the power range of the power fuzzy set. and are the shape coefficients of the low power , medium power and high power medium power membership functions respectively. In this embodiment, and are taken as 120 W, 350 W and 650 W respectively, and are taken as 30, 80 and 120 respectively, and are taken as 1.5, 1.2 and 1.8 respectively.

[0164] Specifically, in the offline training stage, the fuzzy rules are designed considering domain experience to simulate human decision-making thinking, and the mutual relationship between four fuzzy inputs is combined to judge and decide the battery output power , and the decision-making thinking mode is transformed into rules that can be processed by a computer. Since there are four fuzzy inputs in total, and each fuzzy input is divided into three different fuzzy sets respectively, therefore, the input combinations of fuzzy inference total types, and there are also eighty-one corresponding fuzzy rules, which map the eighty-one input combinations to one of the three fuzzy sets corresponding to the fuzzy output respectively. Figure 3 shows the fuzzy rule table, where the first fuzzy rule is the fuzzy rule in the upper left corner of the fuzzy rule table , and it is numbered in the order of columns first and then rows. Then the eighty-first fuzzy rule is the fuzzy rule in the lower right corner . Taking the first fuzzy rule as an example, it shows that when the driving speed belongs to low speed , the driving acceleration belongs to negative acceleration , the remaining battery power belongs to low power and the road surface gradient belongs to downhill , the corresponding battery output power belongs to low power .

[0165] Furthermore, the fuzzy inference unit performs Mamdani inference based on the fuzzy rules and defuzzifies through the centroid method to determine the optimal battery output power , which includes the following specific steps:

[0166] Obtain the driving speed , the driving acceleration , and the remaining battery power and road surface gradient ;

[0167] Traverse the eighty-one fuzzy rules determined in the offline training stage. For a single fuzzy rule, determine the four fuzzy sets in the input combination corresponding to the fuzzy rule, and use the driving speed , driving acceleration , remaining battery power and road surface gradient to substitute into the membership functions corresponding to their respective fuzzy sets to obtain 4 membership degrees and take the minimum membership degree as the activation strength of the single fuzzy rule . Taking the first fuzzy rule as an example, the activation strength corresponding to the first fuzzy rule is specifically as follows:

[0168] ;

[0169] According to the activation strength of the single fuzzy rule, modify the membership function of the power fuzzy set corresponding to the fuzzy rule. Taking the first fuzzy rule as an example, the activation strength of the first fuzzy rule and the power membership function corresponding to low power are respectively and . Then the power membership function corresponding to the modified low power ;

[0170] Perform the maximum composition on the power membership functions modified by the eighty-one rules to generate the combined power membership function ;

[0171] Use the centroid method to defuzzify the combined power membership function. The principle of the centroid method is to regard the area enclosed by the combined power membership function and the independent variable coordinate axis where the battery output power is located as a plane object with a certain mass distribution. The abscissa value corresponding to the centroid of the plane object is the optimal battery output power obtained after defuzzification . Since the battery output power is a continuous value within a specific range, the area between the combined power membership function and the abscissa is the integral of the combined power membership function with respect to the battery output power within the specific range. Then the specific formula for the optimal battery output power is as follows:

[0172] .

[0173] Furthermore, the design of the reward function comprehensively considers battery energy consumption, driving safety and driving comfort, including the following specific steps:

[0174] For time Driving speed , driving acceleration , Single driving distance and remaining battery charge , get the time Driving acceleration and remaining battery charge ;

[0175] Set a battery energy reward, considering the time To time The power consumption per unit distance is equal to the power consumption Divide by the distance traveled per trip , the smaller the power consumption per unit distance, the time Optimal battery output power determined by fuzzy reasoning The more it helps to reduce energy consumption, the higher the corresponding battery energy consumption reward. time The battery energy consumption reward function Set it to the exponential form of the inverse of the power consumption per unit distance to further enhance the impact of battery energy consumption. The specific formula is as follows:

[0176] ;

[0177] in, is the energy consumption influencing factor;

[0178] Set up driving safety rewards, consider time Driving speed and driving acceleration Do not exceed the upper limit of safe speed and the upper limit of safe acceleration , therefore, when time Driving speed and driving acceleration Less than the upper limit of safety speed and the upper limit of safe acceleration When the driving safety reward value increases with the driving speed and driving acceleration decreases with the increase of and the driving speed and / or driving acceleration gets closer to the upper limit of safe acceleration , the decreasing speed of the driving safety reward value is faster. When at the moment the driving speed is greater than or equal to the upper limit of safe speed and / or driving acceleration is greater than or equal to the upper limit of safe acceleration , the driving safety reward will be a very large negative value as a punishment. And at the moment the driving speed and / or driving acceleration the greater the degree of speeding, the faster the punishment intensity increases, so as to drive reinforcement learning to adjust the membership function parameters in fuzzy inference and avoid speeding. Therefore, at the moment the driving safety reward function at the moment the driving speed and driving acceleration are respectively less than the upper limit of safe speed and the upper limit of safe acceleration , it is set in the form of the opposite of a logarithmic function so that the driving safety reward value decreases with the increase of driving speed and driving acceleration and is more sensitive when approaching the upper limit of safe speed and the upper limit of safe acceleration moment the driving speed is greater than or equal to the upper limit of safe speed and / or driving acceleration is greater than or equal to the upper limit of safe acceleration , a large penalty factor is introduced and designed in combination with the reciprocal form of the Sigmoid function, so that the greater the degree of speeding, the faster the punishment intensity increases, specifically as follows:

[0179] ;

[0180] Among them, and respectively represent the "and" relationship and the "and / or" relationship;

[0181] Set a driving comfort reward, and calculate the moment to the moment acceleration change value , when the moment to the moment acceleration change value exceeds the acceleration change upper limit , it indicates that the electric two-wheeler and / or electric three-wheeler has performed a sudden acceleration or deceleration, and the driving comfort reward decreases, and the comfort decreases smoothly as the acceleration change value gradually increases. Therefore, the driving comfort reward function at the moment is designed as a hyperbolic tangent function form with the percentage error between the absolute acceleration change value and the acceleration change upper limit as the independent variable. The hyperbolic tangent function has a smooth curve, specifically as follows:

[0182] ;

[0183] Then the moment reward function is equal to the linear weighted sum of the battery energy consumption reward function , the driving safety reward function and the driving comfort reward function . The linear weights determine the proportion of the importance of battery energy consumption, driving safety, and driving comfort during the reinforcement learning optimization. In this embodiment, battery energy consumption and driving safety are given priority, and the linear weights of battery energy consumption, driving safety, and driving comfort are set to 0.4, 0.4, and 0.2 respectively, that is .

[0184] Specifically, reinforcement learning defines the state and the environmental state. The state includes the driving speed , the driving acceleration , the single-trip distance and the remaining battery power , and the environmental state is the road surface slope . The action space is defined as the battery output power A set is defined, and the action is to adjust all membership function parameters in fuzzy inference. The reward function is used as the Q value of the deep Q network. The deep Q network adopts an experience replay mechanism, storing the environmental state, state, action, and output function value in each decision-making process in the experience pool. When a set of environmental states and states are input into the deep Q network, the probability distribution of the output action is obtained. The proximal policy optimization algorithm is used as the policy gradient algorithm of the deep Q network. Different actions are selected based on the probability distribution of the action, and the influence of different actions on the Q value is evaluated in parallel based on the Q value. The optimal action is determined through truncated importance sampling, so that the Q value continuously increases and finally converges.

[0185] Embodiment 2

[0186] As Figure 4 shown, a new energy battery load optimization method based on an adaptive algorithm is implemented based on a new energy battery load optimization system based on an adaptive algorithm, and includes the following specific steps:

[0187] At the moment collect and obtain driving state data, battery state data, and road surface state data;

[0188] Adopt the Kalman filtering algorithm, and accurately obtain the remaining battery power at the moment at the moment based on the battery system state model combined with the battery state data at the ;

[0189] Adopt the fusion three-dimensional reconstruction algorithm, and integrate the texture information and color information of the left-eye road surface image and the right-eye road surface image taken at the moment into the road surface laser point cloud in the road surface state data at the at the moment to generate the road surface three-dimensional model at the moment at the moment . Obtain the road surface slope and road surface bumpiness at the moment by sampling analysis method and adjust the acquisition interval ;

[0190] Use reinforcement learning to evaluate at the moment The riding benefits brought by fuzzy inference are used to adjust the membership function parameters by the proximal policy optimization algorithm;

[0191] Based on the moment travel state data, remaining battery power and road surface gradient , through the membership function and fuzzy rules, perform fuzzy inference and defuzzification to determine the moment optimal battery output power .

[0192] The present invention discloses a new energy battery load optimization system and method based on an adaptive algorithm, including a multi-source acquisition module, an analysis and decision-making module, and an optimization and adjustment module; the multi-source acquisition module acquires travel state data, battery state data, and road surface state data; the analysis and decision-making module obtains the remaining battery power by combining the battery state data through the Kalman filter algorithm, uses the fusion three-dimensional reconstruction algorithm to reconstruct the road surface state data into a road surface three-dimensional model, obtains the road surface gradient and road surface bumpiness through the sampling analysis method, and adaptively adjusts the generation moment of the next acquisition signal; the optimization and adjustment module obtains the travel state data, the remaining battery power, and the road surface gradient, performs fuzzy inference and defuzzification based on the membership function and fuzzy rules to determine the optimal battery output power, comprehensively evaluates the fuzzy inference benefits based on the reward function, and optimizes the membership function parameters through reinforcement learning to achieve the intelligent optimization matching of the battery output power with the travel state and travel road conditions.

[0193] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A new energy battery load optimization system based on an adaptive algorithm, characterized in that: Including analysis and decision-making module and optimization and adjustment module; The analysis and decision module uses the Kalman filter algorithm, combined with the battery system status model and battery status data to obtain the remaining battery power, adopts the fusion 3D reconstruction algorithm, integrates the texture and color information of the left-eye image and the right-eye image of the road surface into the road surface laser point cloud to generate a 3D model of the road surface, obtains the road surface slope and road surface bumpiness through the sampling analysis method and determines the next acquisition time; The optimization and adjustment module obtains driving status data, remaining battery power and road slope, performs fuzzy reasoning based on membership functions and fuzzy rules, and determines the optimal battery output power through centroid method defuzzification, evaluates fuzzy reasoning benefits through reward functions, and guides reinforcement learning to optimize membership function parameters; The battery system state model determines that the difference between the estimated battery remaining capacity at a single moment and the battery remaining capacity at a previous moment is equal to the charge and discharge capacity from the previous moment to a single moment plus the process noise at a single moment; The fusion 3D reconstruction algorithm establishes scale spaces based on the left-eye image and the right-eye image of the road surface and finds local maximum values ​​to locate feature points, extracts the amplitude information of the gradient direction in the area block around the feature point as feature description, matches similar features through the distance measurement method, generates a road surface image point cloud in combination with the internal parameter matrix, and after alignment with the laser point cloud, performs normal vector redirection through improved Poisson reconstruction to generate a 3D model of the road surface; The sampling analysis method samples the three-dimensional model of the road surface and determines the approximate plane based on the least square method and calculates the road surface slope and road surface bumpiness; The optimization and adjustment module calculates the membership of the driving speed, driving acceleration, battery remaining power and road slope in the driving state data to the fuzzy set according to the membership function, and performs Mamdani reasoning based on the fuzzy rules; The reward function is designed with linear weights to coordinately consider the power consumption per unit distance, the driving speed and the driving acceleration, the relationship between the upper limit of the safe speed and the upper limit of the safe acceleration, and the acceleration change value.

2. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 1, characterized in that: The optimization and adjustment module includes a fuzzy reasoning unit and an online optimization unit; The fuzzy reasoning unit preselects a membership function according to the characteristics of the fuzzy input and the fuzzy output, establishes fuzzy rules and determines the membership function parameters, calculates the membership of the driving speed, driving acceleration, battery remaining power and road slope in the driving state data to the fuzzy set according to the membership function, performs Mamdani reasoning based on the fuzzy rules and generates the optimal battery output power by defuzzification through the centroid method; The online optimization unit obtains the driving speed, driving acceleration, single driving distance and remaining battery power in the driving status data, comprehensively considers battery energy consumption, driving safety and driving comfort to design a reward function to evaluate the fuzzy reasoning benefit, and guides reinforcement learning to optimize the membership function parameters online to improve the fuzzy reasoning performance.

3. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 2, characterized in that: The membership function is pre-selected according to the characteristics of the fuzzy input and the fuzzy output, including: The speed fuzzy set is divided into low speed, medium speed and high speed according to the predefined speed fuzzy set boundary value, and a trapezoidal membership function is selected to describe the speed fuzzy set; According to the predefined acceleration fuzzy set boundary value, the acceleration fuzzy set is divided into negative acceleration, zero acceleration and positive acceleration, and the trapezoidal membership function is selected to describe the negative acceleration and positive acceleration, and the triangle membership function is selected to describe the zero acceleration; The power fuzzy set is defined to include low power, medium power and high power, and the unilateral Sigmoid membership function is selected to describe the power fuzzy set. Define the slope fuzzy set including downhill, flat road and uphill, and choose Gaussian membership function to describe the slope fuzzy set; The power fuzzy set is defined as low power, medium power and high power, and a bell-shaped membership function is selected to describe the power fuzzy set.

4. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 2, characterized in that: The method of performing Mamdani reasoning based on fuzzy rules and defuzzifying the optimal battery output power by the centroid method includes the following specific steps: Traverse all fuzzy rules, and for a single fuzzy rule, determine the specific speed fuzzy set, acceleration fuzzy set, power fuzzy set and slope fuzzy set in the corresponding input combination, calculate the membership of the driving speed, driving acceleration, battery remaining power and road slope with the corresponding fuzzy set respectively, and take the minimum membership as the activation strength of the single fuzzy rule; According to the activation intensity of a single fuzzy rule, the membership function of the power fuzzy set corresponding to the fuzzy rule is modified, and the power membership functions after all fuzzy rules are modified are synthesized to generate a synthetic power membership function. The center of gravity method is used to defuzzify the synthetic power membership function, and the area enclosed by the synthetic power membership function and the corresponding independent variable coordinate axis is taken as a plane object. The horizontal coordinate value of the center of gravity of the plane object is the optimal battery output power.

5. The new energy battery load optimization system based on adaptive algorithm as claimed in claim 1, characterized in that: The reward function is obtained according to battery energy consumption, driving safety and driving comfort, and includes the following specific steps: Get the power consumption per unit distance. According to the principle that the smaller the power consumption per unit distance, the higher the battery energy consumption reward, the battery energy consumption reward function is set to the exponential form of the inverse of the power consumption per unit distance. According to the requirement that both the driving speed and the driving acceleration cannot exceed the upper limit of the safe speed and the upper limit of the safe acceleration, when the driving speed and the driving acceleration are respectively less than the upper limit of the safe speed and the upper limit of the safe acceleration, the driving safety reward function is set to the inverse form of the logarithmic function, so that the driving safety reward value decreases with the increase of the driving speed and the driving acceleration. When the driving speed is greater than or equal to the upper limit of the safe speed and / or the driving acceleration is greater than or equal to the upper limit of the safe acceleration, a penalty factor is introduced to set the driving safety reward function in combination with the transformation form of the Sigmoid function; The acceleration change value is calculated, and the driving comfort decreases as the acceleration change value increases. The driving comfort reward function is set to a hyperbolic tangent function with the percentage error between the absolute acceleration change value and the upper limit of the acceleration change as the independent variable; The reward function is equal to the linear weighted sum of the battery energy consumption reward function, the driving safety reward function, and the driving comfort reward function.

6. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 1, characterized in that: The analysis and decision module includes a battery estimation unit, a road surface analysis unit and an adjustment decision unit; The battery estimation unit obtains the output current and terminal voltage in the battery status data at a single moment, adopts the Kalman filter algorithm, predicts the estimated battery remaining capacity at a single moment based on the battery remaining capacity at the previous moment and the battery system status model, and makes corrections based on the estimation error covariance and observation noise at a single moment to obtain the battery remaining capacity at a single moment; The pavement analysis unit adopts a fusion 3D reconstruction algorithm to match similar features from the left-eye image and the right-eye image of the pavement and generate a pavement image point cloud in combination with an internal reference matrix, and performs registration and fusion with the pavement laser point cloud to obtain a complete pavement point cloud. The improved Poisson reconstruction is used to redirect the normal vector of the complete pavement point cloud and model it as a Gaussian Poisson problem to reconstruct and generate a 3D pavement model. The adjustment decision unit samples the three-dimensional road model through the sampling analysis method, determines the approximate plane based on the least squares method and calculates the road slope and road bumpiness, and determines the generation time of the next acquisition signal based on the road bumpiness.

7. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 6, characterized in that: The method of obtaining the remaining battery power at a single moment includes the following specific steps: Establish a battery system status model; The observation equation is established. The transformation relationship between the terminal voltage at a single moment and the remaining battery capacity is established by adding the observation transformation function to the observation noise at a single moment. The estimation error covariance of the previous moment is superimposed with the process noise of a single moment to obtain the prediction estimation error covariance of a single moment, and the Jacobian matrix is ​​obtained by taking the derivative of the observation transformation function with respect to the remaining battery power at a single moment; The prediction estimation error covariance at a single moment is mapped to the observation space through the Jacobian matrix, and the Kalman gain at a single moment is calculated; The remaining battery capacity at a single moment is obtained by combining the error between the terminal voltage at a single moment and the mapping value after the estimated remaining battery capacity is transformed by the observation transformation function with the Kalman gain correction at a single moment, and the estimated error covariance at a single moment is obtained based on the prediction estimation error covariance update.

8. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 6, characterized in that: The fusion 3D reconstruction algorithm includes the following specific steps: The left eye scale space and the right eye scale space are established based on the left eye image and the right eye image of the road surface by using a stable feature acceleration algorithm, local maximum values ​​are found and weak response feature points are eliminated to achieve feature point positioning, wherein the weak response feature point is a feature point whose Hessian matrix determinant value is less than the response threshold; The gradient histogram in the neighborhood of a single feature point is counted to determine the main direction of the single feature point, the area block around the single feature point is extracted and the amplitude information of the gradient direction in the area block is counted to construct the feature description of the single feature point; The distance measurement method is used to compare and match the feature descriptions of the feature points in the left eye scale space and the right eye scale space to generate multiple sets of matching point pairs, and the basic matrix with the largest number of matching point pairs that meet the geometric constraints is selected through a random sampling consistency algorithm; Calculate the internal parameter matrix, remove the influence of the internal parameter matrix on the basic matrix through bundle adjustment, obtain the external parameter matrix, and combine the triangulation principle to perform projection transformation on the matching point pairs to generate the road surface image point cloud, and obtain the complete road surface point cloud by registering and fusing it with the road surface laser point cloud through the iterative closest point algorithm; Improved Poisson reconstruction is used to redirect the normal vectors in the complete point cloud of the road surface, and the three-dimensional reconstruction is regarded as a Gaussian Poisson problem to solve and restore the three-dimensional model of the road surface.

9. A new energy battery load optimization system based on an adaptive algorithm as claimed in claim 8, characterized in that: The improved Poisson reconstruction comprises the following specific steps: The complete point cloud of the road surface is smoothed by the moving least square method, and the normal vector and point cloud curvature of each point in the complete point cloud of the road surface are calculated by the principal component analysis method. The point with the smallest curvature of the point cloud is taken as the origin and the positive direction is specified. Perform neighborhood search through KD tree, perform scalar product of multiple searched points with the origin in turn, reverse the direction of the normal vector of the point whose scalar product is less than 0, and mark the point that completes the scalar product; The last point in the neighborhood that completes the scalar product is used as the new origin, and a new round of neighborhood search is performed through the KD tree. The scalar product is continued with the unmarked point to reverse or maintain the normal vector, and the normal vector is marked. After all points in the complete road point cloud are marked, an octree topology is constructed so that all points in the complete road point cloud fall on leaf nodes. The node functions of leaf nodes are constructed based on three-dimensional box filtering. All node functions are linearly combined through cubic spline interpolation to construct a vector field. The Gaussian Poisson problem is constructed and converted into a system of linear equations. The three-dimensional scalar field is obtained by solving it through the conjugate gradient method. The marching cubes algorithm is used to extract the isosurface from the three-dimensional scalar field to generate a three-dimensional road model.

10. A new energy battery load optimization method based on an adaptive algorithm, executed based on the new energy battery load optimization system based on an adaptive algorithm as claimed in any one of claims 1 to 9, characterized in that: The specific steps include: The Kalman filter algorithm is used to obtain the remaining battery power at a single moment based on the battery system state model and the battery state data collected at a single moment; A fusion 3D reconstruction algorithm is used to integrate the texture information and color information of the left and right images of the road surface taken at a single moment into the road surface laser point cloud collected at a single moment to generate a 3D model of the road surface. The road surface slope and road surface bumpiness at a single moment are obtained through sampling analysis and the collection interval is adjusted. Reinforcement learning is used to evaluate the riding benefits brought by the fuzzy reasoning at the previous moment in multiple aspects based on the reward function, and the proximal strategy optimization algorithm is used to adjust the membership function parameters; Based on the driving status data collected at a single moment, the remaining battery power at a single moment, and the road slope at a single moment, fuzzy reasoning and defuzzification are performed through membership functions and fuzzy rules to determine the optimal battery output power at a single moment.

Images