A log data-based mine card simulation method and system

Abstract

The application discloses a mine card simulation method and system based on log data, and relates to unmanned driving: obtaining running log data of a mine card under multiple working conditions; extracting nonlinear response feature data according to the running log data; taking historical control instructions and historical vehicle states of continuous N periods as input features, taking a vehicle state of a current period as an output label, and constructing a training set; training an LSTM model by using the training set, inputting a current control instruction, a historical vehicle state and a current load value, and outputting a predicted vehicle response state; and inputting the predicted vehicle speed, acceleration, front wheel turning angle or hinged angle into a vehicle kinematics model to calculate the position and posture of the vehicle under the current load condition. The application improves the simulation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned driving, and particularly relates to a mine truck simulation method and system based on log data. BACKGROUND

[0002] With the rapid development of mining automation technology, unmanned mine trucks have become an important technical means to improve the production efficiency and operational safety of mines. The core of unmanned mine trucks lies in advanced path planning and motion control algorithms, which need to be fully tested and verified before being put into practical application. However, there are many problems in directly using real vehicles for algorithm testing on the mine site: first, the mine operation environment is dangerous, and the testing process may threaten personnel safety; second, mine truck equipment is expensive, and testing failure may cause significant economic losses; third, real vehicle testing is long and inefficient, which seriously restricts the iteration and optimization speed of algorithms. Therefore, it is an urgent need in the industry to establish an accurate mine truck simulation model to complete algorithm testing in a virtual environment.

[0003] Existing mine truck simulation technologies are mainly based on theoretical dynamic models. For example, CN111368424A discloses a vehicle simulation method based on tire dynamics, which simulates the dynamic response of the vehicle by establishing a tire mechanics model. However, this kind of simulation method based on theoretical models has fundamental defects when applied to mine trucks:

[0004] Firstly, the load of a mine truck varies greatly, from tens of tons when empty to hundreds of tons when fully loaded. This huge mass variation makes the dynamic characteristics of the vehicle highly nonlinear. Traditional simulation models usually use fixed dynamic parameters, which cannot accurately reflect the significant differences in vehicle response under different load conditions. For example, with the same throttle opening, an empty mine truck may accelerate to the target speed within 2 seconds, while a fully loaded mine truck may take more than 10 seconds. This load-dependent nonlinear response characteristic cannot be accurately described by a fixed parameter model.

[0005] Secondly, as a super-large inertial system, a mine truck has a significant time-varying delay from the issuance of control instructions to the actual change in vehicle state. This delay is not only related to the load but also affected by factors such as road slope, current vehicle state, etc. Existing simulation models either ignore this delay, assuming that the vehicle responds immediately to control instructions, or use fixed delay parameters, which cannot reflect the time-varying characteristics of the actual system. Especially in complex mine conditions, the transmission path of control instructions includes the communication system, vehicle-mounted controller, hydraulic actuator, and other links, and the delay of each link may vary with the system state, making the overall response delay exhibit complex time-varying characteristics.

[0006] In addition, the particularity of the mine environment further exacerbates the difficulty of simulation modeling. The mine truck needs to work on the transport road with a slope of up to 15%, and the road conditions vary greatly from hard rock to soft soil, which will significantly affect the dynamic response of the vehicle. The theoretical model is difficult to accurately consider the complex environment-vehicle interaction, resulting in a large deviation between the simulation results and the actual vehicle behavior.

[0007] Therefore, there is an urgent need for a new simulation method that can accurately model the nonlinear dynamic characteristics of the mine truck under different load conditions and effectively handle the time-varying delay between the control command and the actual response, thereby improving the accuracy and reliability of the mine truck simulation and providing technical support for the safe and efficient testing of the unmanned driving algorithm. SUMMARY

[0008] The fixed dynamic parameters used in the simulation of the mine truck cannot accurately simulate the nonlinear dynamic response caused by the large load change of the mine truck, which leads to the inability to handle the time delay between the control command and the actual response in the super-large inertia system. The present application provides a mine truck simulation method and system based on log data, which improves the simulation accuracy.

[0009] One aspect of the present application provides a mine truck simulation method based on log data, comprising: S1, obtaining the running log data of the mine truck under multiple working conditions; wherein the working conditions include: empty load, 1 / 4 load, 1 / 2 load, 3 / 4 load, full load; the running log data includes: control command, vehicle state data and load data; S2, extracting nonlinear response feature data according to the running log data: extracting the actual acceleration value of the vehicle generated after the same throttle or brake control command input under different load conditions as the acceleration response data; extracting the actual change value of the front wheel steering angle or the articulated angle after the same steering control command input under different load conditions as the steering angle change data;

[0010] S3, aligning the control command and the nonlinear response feature data according to the time stamp; taking the historical control command and the historical vehicle state of the continuous N periods as the input features, and taking the vehicle state of the current period as the output label to construct the training set; S4, training the LSTM model using the training set, the input of the LSTM model includes: historical control command sequence, historical vehicle state sequence, current load value and response delay time, and the output is the predicted current vehicle state; S5, using the trained LSTM model, inputting the current control command, historical vehicle state and current load value, and outputting the predicted vehicle response state; the vehicle response state includes vehicle speed, acceleration, front wheel steering angle or articulated angle; inputting the predicted vehicle speed, acceleration, front wheel steering angle or articulated angle into the vehicle kinematics model to calculate the position and attitude of the vehicle under the current load condition.

[0011] In particular, the fundamental difficulty of mining truck simulation lies in the highly nonlinear and strong environment dependence of its dynamic behavior. Traditional simulation methods based on physical equations require a large number of simplifications for vehicle-ground interaction, hydraulic system response, engine characteristics, etc. These simplifications will produce unacceptable errors when faced with the huge load variation of mining trucks (from tens of tons to hundreds of tons) and the complex mine road conditions. Therefore, the present application directly learns the real dynamic behavior pattern from the operation log.

[0012] Log data contains the combined effects of complex factors that cannot be directly reflected by traditional simulation methods, such as when the mining truck is fully loaded on a gravel road with a 15% slope, its actual acceleration response has integrated the effects of tire skidding, engine power attenuation, hydraulic system pressure fluctuations, etc. By extracting the actual response of the same control command under different working conditions, the present application bypasses the modeling between complex physical variables and directly extracts the input-output mapping relationship from the operation log.

[0013] On this basis, an LSTM model is used. The super-large inertia of the mining truck determines that its state change has a strong historical accumulation effect - to predict the speed at the next moment, not only the current throttle opening needs to be known, but also the acceleration history in the past few seconds needs to be understood. The gating mechanism of LSTM can selectively remember key historical information and automatically learn the dependency relationship at different time scales, which is highly consistent with the physical characteristics that the delay from command issuance to response generation varies with the working conditions.

[0014] However, a purely data-driven method may produce prediction results that violate physical laws. Therefore, the present scheme decouples the entire simulation task into two levels: LSTM is responsible for learning the complex dynamic response law, and outputs the speed, acceleration and steering angle that conform to the historical data distribution; the kinematic model calculates the vehicle trajectory based on these dynamic quantities through strict geometric constraints. This dual-layer architecture of data-driven dynamics + constraint-ensured kinematics not only ensures accurate modeling of real complex dynamic behavior, but also ensures the physical reasonableness of the simulation trajectory, realizing the organic integration of data-driven methods and physical constraints.

[0015] Further, the vehicle state data includes: vehicle position coordinates, heading angle, speed, acceleration, front wheel steering angle or articulation angle, gear position, throttle opening, brake pressure.

[0016] Further, the control command includes a longitudinal control command and a lateral control command; the longitudinal control command includes: throttle opening, brake pressure, desired acceleration and gear position; for Ackerman vehicles, the lateral control command includes front wheel steering angle or steering wheel steering angle; for articulated vehicles, the lateral control command includes left and right cylinder opening or articulation angle.

[0017] Further, before step S2, further comprising: removing invalid data that does not affect the change of the vehicle state, the invalid data including: the accelerator or brake operation log in the neutral state; removing abnormal data introduced due to log recording errors, the abnormal data including: speed values, acceleration values or steering angle values exceeding the physical limit range;

[0018] Further, S3, constructing a training set, comprising: time stamp matching control instruction data and vehicle state data; obtaining the control instruction issuing time t1 and the actual vehicle state change response time t2; wherein the actual vehicle state change refers to the change of speed, acceleration, front wheel steering angle or hinge angle relative to the state value before the control instruction is issued; calculating the vehicle state response delay time Δt according to t2 and t1; wherein the response delay time includes: acceleration / braking response delay time under different load conditions, and steering system response delay time; for articulated vehicles, it also includes hinge mechanism response delay time; resampling data with different sampling frequencies, the resampling frequency being T; using a sliding window method to construct training samples, the window size being N cycles and the sliding step being 1 cycle; wherein each training sample includes: input part: control instruction sequence, vehicle state sequence, current load value and response delay time sequence at t-N+1 to t time; output part: vehicle state at t+1 time;

[0019] In particular, the traditional simulation method usually assumes that there is a fixed time relationship between the control instruction and the vehicle response, or lets the model learn this time sequence relationship from the data. However, as a super-inertia system, the response delay of the mine truck has a high degree of time-varying and nonlinearity - the acceleration response delay when full load climbing may be as long as several seconds, while the response delay of the empty truck on the flat road may be only a few tenths of a second.

[0020] The present application calculates the delay time Δt between the control instruction issuing time t1 and the actual vehicle state change time t2, and adds it as an independent input feature sequence to the LSTM model, which realizes the explicit modeling of the time-varying delay characteristic. The model no longer needs to implicitly infer the delay rule from complex time sequence data, but directly obtains the delay time as an input, which greatly reduces the complexity of model learning. More importantly, the response delay time sequence as input enables the trained model to predict the future response delay according to the current system state.

[0021] Further, S4, training the LSTM model using the training set, comprising: preprocessing the constructed training set; constructing an LSTM network model: the input layer receives four types of features: historical control instruction sequence, historical vehicle state sequence, current load value and response delay time sequence, the input dimension being N*D, wherein N is the length of the historical sequence and D is the feature dimension;

[0022] The hidden layer adopts multi-layer LSTM units to obtain the time sequence dependency under different load conditions; wherein, the time sequence dependency represents the time delay law between the control instruction and the vehicle state response; the specific setting is: the first layer LSTM: 128 hidden units are set, the control instruction information at the historical time is selectively retained through the forgetting gate in the gating mechanism, the vehicle state change at the current time is fused through the input gate, and the modeling of the time sequence dependency is realized; the second layer LSTM: 64 hidden units are set, the current load value and the response delay time are taken as additional condition inputs, which are input after being spliced with the output of the first layer LSTM, so that the same control sequence activates different neuron combinations under different load conditions, and the modeling of the nonlinear dynamic characteristics is realized; a Dropout layer is added after each LSTM layer, and the dropout rate is set to 0.2 to prevent overfitting; the output layer adopts a fully connected layer to output the predicted current vehicle state; a loss function and an optimizer are set; the preprocessed training set is used to train the constructed LSTM network model;

[0023] Further, the constructed training set is preprocessed, including: the load data is classified and coded according to: empty load, 1 / 4 load, 1 / 2 load, 3 / 4 load, full load, and coded as discrete values of 0, 1, 2, 3 and 4; the control instruction sequence, the vehicle state sequence and the response delay time sequence are mapped to the interval of 0 to 1; interpolation is performed on the data between adjacent load levels to generate virtual samples of intermediate load states, so as to enhance the adaptability of the model to continuous load changes;

[0024] Further, the loss function is: ; wherein, and are the predicted and actual vehicle speeds, and are the predicted and actual accelerations, and are the predicted and actual front wheel turning angles or hinged angles; represents the speed change rate calculated based on the predicted speed; represents the speed change rate calculated based on the actual speed; is a weight coefficient, which is set according to the importance and dimension difference of different state variables;

[0025] In particular, the present application introduces a speed change rate error term , which directly models and optimizes the time-varying response delay characteristics of the super-large inertia system of the mine truck. The traditional loss function only focuses on the state prediction accuracy at a certain time, and ignores the dynamic process from the issuance of the control instruction to the actual change of the state.

[0026] In the mine truck simulation, due to the huge change of vehicle mass from dozens of tons of empty load to hundreds of tons of full load, the same control instruction produces different speed change rates under different load conditions - slow acceleration when full load, and rapid response when empty load. The speed change rate Δv / Δt exactly reflects this dynamic response characteristic, which not only contains the final speed value, but more importantly, depicts the speed of reaching the speed.

[0027] Further, S5, the position and attitude of the vehicle under the current load condition are calculated, including: for the Ackerman car, the kinematic model adopts the Ackerman steering geometry model; for the articulated vehicle, the kinematic model adopts the articulated vehicle kinematic model.

[0028] Another aspect of the present application also provides a mine truck simulation system based on log data, comprising: a data acquisition module for acquiring running log data of a mine truck under multiple working conditions; wherein the working conditions include: empty load, 1 / 4 load, 1 / 2 load, 3 / 4 load, full load; the running log data includes: control instruction, vehicle state data and load data;

[0029] A feature extraction module is configured to extract nonlinear response feature data from the running log data: extract the actual acceleration value generated by the vehicle under the same throttle or brake control instruction input under different load conditions as the acceleration response data; extract the actual change value of the front wheel steering angle or the articulated angle after the same steering control instruction input under different load conditions as the steering angle change data;

[0030] A training set construction module is configured to align the control instruction and the nonlinear response feature data according to the timestamp; the historical control instruction and the historical vehicle state of the continuous N periods are taken as the input features, and the vehicle state of the current period is taken as the output label to construct the training set; the module specifically includes: a time alignment submodule for timestamp matching control instruction data and vehicle state data; a delay calculation submodule for obtaining the control instruction issuing time t1 and the actual vehicle state change response time t2, and calculating the vehicle state response delay time Δt according to t2 and t1; a sample construction submodule for constructing training samples using a sliding window method, with a window size of N periods and a sliding step of 1 period;

[0031] A model training module is configured to train an LSTM model using the training set, wherein the input of the LSTM model includes: a historical control instruction sequence, a historical vehicle state sequence, a current load value and a response delay time, and the output is a predicted current vehicle state; the loss function configured by the module is: ;

[0032] The simulation prediction module is configured to input the current control instruction, the historical vehicle state and the current load value into the trained LSTM model to output a predicted vehicle response state; and input the predicted vehicle speed, acceleration, front wheel steering angle or articulation angle into a vehicle kinematics model to calculate the position and attitude of the vehicle under the current load condition.

[0033] Compared with the prior art, the application has the following advantages:

[0034] In view of the problem in the prior art that the simulation model of the mine truck uses fixed dynamic parameters, which cannot accurately simulate the nonlinear dynamic response of the mine truck in a huge load change range from empty load (tens of tons of self weight) to full load (hundreds of tons of total weight), and cannot process the time-varying delay characteristics between the control instruction and the actual response in the super-large inertia system, resulting in a large deviation between the simulation result and the actual vehicle behavior, the application provides a mine truck simulation method based on log data, which extracts actual response data corresponding to the same control instruction under different load conditions from actual operation logs, calculates the response delay time from the issuance of the control instruction to the actual change of the vehicle state, and takes the historical control instruction sequence, the historical vehicle state sequence, the current load value and the response delay time sequence as inputs of an LSTM model, thereby establishing a dynamic simulation model considering load change and response delay, and realizing accurate modeling of the nonlinear dynamic characteristics and time-varying response delay of the mine truck under different load conditions, so that the simulation system can dynamically adjust the vehicle response parameters according to the actual load and system state, and the accuracy and authenticity of the mine truck simulation are improved. BRIEF DESCRIPTION OF DRAWINGS

[0035] The application will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same reference numbers represent the same structures, wherein:

[0036] Figure 1 is an exemplary flowchart of a mine truck simulation method based on log data according to some embodiments of the application;

[0037] Figure 2 is a kinematics model of an Ackerman vehicle according to the application;

[0038] Figure 3 is a kinematics model of an articulated vehicle according to the application;

[0039] Figure 4 is a structural schematic diagram of an articulated vehicle according to the application. DETAILED DESCRIPTION

[0040] The method and system provided by the embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0041] AsFigure 1 The running log data of the mine card under multiple working conditions is obtained, as shown in the drawings; the working conditions include: no load, 1 / 4 load, 1 / 2 load, 3 / 4 load, full load; the running log data includes: control instructions, vehicle state data and load data; nonlinear response feature data is extracted according to the running log data: under different load conditions, the actual acceleration value generated by the vehicle after the same throttle or brake control instruction input is extracted as the acceleration response data; under different load conditions, the actual change value of the front wheel steering angle or the hinged angle after the same steering control instruction input is extracted as the steering angle change data; the control instructions and the nonlinear response feature data are aligned according to the time stamp; the historical control instructions and the historical vehicle state of the continuous N periods are taken as the input features, and the vehicle state of the current period is taken as the output label to construct a training set; the LSTM model is trained by using the training set, the input of the LSTM model includes: the historical control instruction sequence, the historical vehicle state sequence, the current load value and the response delay time, and the output is the predicted current vehicle state; the trained LSTM model is used, the current control instruction, the historical vehicle state and the current load value are input, and the predicted vehicle response state is output; the vehicle response state includes vehicle speed, acceleration, front wheel steering angle or hinged angle; the predicted vehicle speed, acceleration, front wheel steering angle or hinged angle is input into the vehicle kinematics model to calculate the position and attitude of the vehicle under the current load condition.

[0042] S1, the unmanned program periodically issues control instructions to the vehicle chassis. Common longitudinal control instructions include: throttle, brake, desired acceleration, gear, etc. For Ackerman vehicles, common lateral control instructions include: front wheel steering angle, steering wheel angle, etc. For articulated vehicles, common control instructions include: left and right cylinder opening or hinged angle, etc. The vehicle chassis will periodically report its various state information, including vehicle current position and attitude information, current front wheel steering angle, current hinged angle, current gear, current speed, current throttle, current brake, etc.

[0043] S2, on the real vehicle that needs to be simulated, start the unmanned program, and confirm that the corresponding control instruction information and vehicle state information will be recorded in the log. Then drive in the real working environment of the real vehicle to collect relevant logs. The log needs to contain the following information:

[0044] Control instructions. Mainly various control instructions that affect the driving of the vehicle chassis, including longitudinal control instructions (acceleration and deceleration control, gear, etc.) and lateral control instructions (steering instructions).

[0045] Vehicle state information. Includes the following:

[0046] a) Vehicle position and attitude information. Position information is the coordinate information of the real-time position of the vehicle. Attitude information is the heading angle of the vehicle. For articulated vehicles, it also includes the heading angle information of the front and rear bodies.

[0047] b) Chassis state information. Including gear information, current throttle / brake state, current engine / motor speed, current front wheel angle, vehicle weight, etc.

[0048] After collecting logs, data cleaning is needed. Mainly to remove invalid data and abnormal data. Invalid data refers to data that does not affect the change of vehicle state, such as throttle or brake operation in neutral state. Abnormal data refers to obviously unreasonable data, which may be introduced due to log recording errors, etc.

[0049] S3, time stamp matching of control command data and vehicle state data from different data sources. Since the control system and state acquisition system may have different sampling frequencies and clock biases, a unified time reference needs to be established. Convert the time stamp of all data to a unified millisecond time stamp and arrange them in chronological order. For data points with time stamps not completely aligned, use the nearest neighbor matching principle, that is, associate the control command with the vehicle state data closest to its time stamp.

[0050] Determine the delay time by detecting the actual change of vehicle state after the control command is issued:

[0051] For longitudinal control (throttle / brake), the system monitors the change of speed or acceleration relative to the reference value. The reference value is defined as the state value of the previous sampling period before the control command is issued. When the speed change exceeds the threshold value (such as 0.1 m / s) or the acceleration change exceeds the threshold value (such as 0.05 m / s²), it is considered that the response has started, and the time t2 is recorded.

[0052] For lateral control (steering), the system monitors the change of front wheel angle or hinge angle. When the angle change exceeds the threshold value (such as 0.5 degrees), it is considered that the steering response has started. The delay time Δt = t2 - t1, where t1 is the time when the control command is issued, and t2 is the time when the response starts. The system will record the acceleration delay, braking delay and steering delay according to the current load condition and control type.

[0053] Since the sampling frequencies of different sensors and systems may be different (such as control command may be 20Hz, while vehicle state may be 50Hz), resampling is needed to obtain time-aligned data sequences. The system uses linear interpolation method to resample all data to a unified frequency T (such as 50Hz). For discrete data (such as gear), use forward padding; for continuous data (such as speed, angle), use linear interpolation to calculate the value at the intermediate time.

[0054] The training samples are extracted from the time series data using a sliding window method. The window size is set to N periods (e.g., N = 100, corresponding to 2 seconds of historical data), ensuring that sufficient historical information is included to reflect the inertial characteristics of the mine truck. The sliding step is 1 period, ensuring full utilization of the data and diversity of the samples.

[0055] The input part of each training sample includes four components: a historical control command sequence, including the throttle, brake, and steering control values from time t-N+1 to t; a historical vehicle state sequence, including the speed, acceleration, and steering angle values at the corresponding times; the current load value, indicating the current load level (discrete values from 0 to 4); and the response delay time sequence, including the calculated delay times at the historical times. The output part is the vehicle state at time t+1, including the predicted speed, acceleration, front wheel steering angle, or articulation angle.

[0056] By accurately calculating and recording the response delay, the time-varying delay characteristics of the mine truck are explicitly expressed. The sliding window retains sufficient historical information, allowing the model to learn the dynamic response rules of the mine truck. Through uniform resampling processing, the time consistency of different data sources is ensured.

[0057] S4, training the LSTM model using the training set, the input of the LSTM model includes: historical control command sequence, historical vehicle state sequence, current load value and response delay time, and the output is the predicted current vehicle state, including:

[0058] The constructed training set is preprocessed; the load data is classified and coded according to: empty, 1 / 4 load, 1 / 2 load, 3 / 4 load, and full load, coded as discrete values 0, 1, 2, 3, and 4. The continuous load value is mapped to 5 discrete levels (0-4), which is based on the fact that the load in the mine usually presents a stepwise distribution. Discretization allows the model to better learn the typical dynamics patterns under different load levels, avoiding noise interference caused by continuous values.

[0059] The control command sequence (throttle 0% to 100%, brake 0-100 bar, steering angle -30° to +30°), vehicle state sequence (speed 0-30 m / s, acceleration -3 to +2 m / s²), and response delay time (0-5 seconds) are mapped to the [0, 1] interval. This normalization process eliminates the dimensional differences of different physical quantities, preventing certain features with larger numerical ranges from dominating the model training process. The minimum-maximum normalization method is used, which preserves the relative distribution characteristics of the data.

[0060] Linear interpolation between adjacent load levels to generate virtual samples for intermediate load conditions. For example, based on real data for empty and 1 / 4 load, virtual samples for 1 / 8 load are generated. This data augmentation strategy enables the model to learn the gradual dynamics of varying load, improving its generalization ability under arbitrary load conditions.

[0061] Building the LSTM network model: The input layer receives four types of features: historical control command sequence, historical vehicle state sequence, current load value, and response delay time sequence, with an input dimension of N*D, where N is the length of the historical sequence and D is the feature dimension.

[0062] The hidden layer uses multiple LSTM units to capture the time-dependent relationship under different load conditions. The time-dependent relationship represents the time delay pattern between control commands and vehicle state responses. The specific settings are as follows:

[0063] First layer LSTM (128 hidden units): Mainly responsible for extracting the time-dependent pattern of control-response. This layer selectively retains important historical control information through the forget gate mechanism, such as continuous acceleration or continuous braking control patterns. The input gate is responsible for integrating the state change information at the current time, establishing a mapping relationship between the past N periods of control sequence and the current state. The setting of 128 hidden units ensures sufficient expression ability to capture complex time patterns.

[0064] Second layer LSTM (64 hidden units): Specifically handles nonlinear modeling under load and delay conditions. This layer inputs the load value and response delay time as conditional information after concatenating with the output of the first layer. This design enables the network to activate different neuron combinations according to different load conditions, achieving modeling of different responses under the same control sequence under different loads. The setting of 64 units balances the model complexity and computational efficiency.

[0065] Dropout regularization: Add 20% Dropout after each LSTM layer to randomly deactivate part of the neuron connections. This technique prevents the model from over-relying on specific neuron paths, improving the model's generalization ability, especially in handling load condition changes.

[0066] Setting the loss function and optimizer:

[0067] where, represents the predicted vehicle speed at t+1 time output by the LSTM model, unit: m / s; represents the real vehicle speed at t+1 time obtained from the log data, unit: m / s; represents the predicted vehicle acceleration at time t+1 output by the LSTM model, with the unit of m / s²; represents the real vehicle acceleration at time t+1 obtained from the log data, with the unit of m / s²; represents the predicted front wheel steering angle (Ackerman vehicle) or articulation angle (articulated vehicle) at time t+1 output by the LSTM model, with the unit of rad; represents the real front wheel steering angle (Ackerman vehicle) or articulation angle (articulated vehicle) at time t+1 obtained from the log data, with the unit of rad.

[0068] represents the rate of change of speed calculated based on the predicted speed, with the unit of m / s²; represents the rate of change of speed calculated based on the real speed, with the unit of m / s²; T represents the data resampling period, with the unit of s; MAPE represents the mean absolute error, and MSE represents the mean square error;

[0069] represents the speed prediction error weight coefficient, with the value range of [0, 1], used to adjust the contribution of the speed prediction error in the total loss; represents the acceleration prediction error weight coefficient, with the value range of [0, 1], used to adjust the contribution of the acceleration prediction error in the total loss; represents the steering angle prediction error weight coefficient, with the value range of [0, 1], used to adjust the contribution of the steering angle prediction error in the total loss; represents the rate of change of speed prediction error weight coefficient, with the value range of [0, 1], used to adjust the contribution of the dynamic response prediction error in the total loss. .

[0070] The training process adopts the batch gradient descent method, and the batch size is set to a value that considers the balance of samples of different load levels (such as including samples of each load level in each batch). The learning rate adopts an adaptive adjustment strategy, with a large initial learning rate to quickly converge, and gradually reduced in the later period to fine-tune. The validation set loss is monitored in real time during the training process, and the early stopping strategy is adopted to prevent overfitting.

[0071] The LSTM network model is trained and constructed using the preprocessed training set;

[0072] S5, calculating the position and attitude of the vehicle under the current load condition, including:

[0073] Obtain the control instruction sequence at the current time and the previous N-1 times, including longitudinal control instructions and lateral control instructions; obtain the vehicle state sequence at the previous N times, including vehicle position coordinates, heading angle, speed, acceleration, front wheel steering angle or articulation angle; obtain the current load value and perform category encoding; obtain the calculated response delay time sequence;

[0074] The control instruction sequence, vehicle state sequence and response delay time sequence are mapped to the interval [0, 1] using the same normalization parameters as in the training phase; the preprocessed data is organized into an input tensor with dimensions NxD;

[0075] The preprocessed input tensor is input into the trained LSTM model; the model outputs the predicted current vehicle response state, including vehicle speed , acceleration , front wheel steering angle or articulation angle ; the output result is denormalized to restore the actual physical dimension;

[0076] The predicted vehicle speed , front wheel steering angle or articulation angle is input into the vehicle kinematics model; according to the vehicle type, the corresponding kinematics equation is selected: as shown in Figures 2 to 4 , the Ackerman steering geometry model is used for Ackerman vehicles, and the articulated vehicle kinematics model is used for articulated vehicles; combined with the time step T, the vehicle position coordinates (x, y) and heading angle at the next time are calculated;

[0077] Specifically, for Ackerman vehicles, combined with Figure 2 , the following method is used to calculate:

[0078] The yaw rate of the Ackerman steering vehicle is calculated by the following formula: ; where represents the front wheel steering angle at the current time, represents the wheelbase of the vehicle, represents the predicted vehicle speed at the current time.

[0079] The vehicle heading angle of the Ackerman vehicle at the next time is: ;

[0080] The position coordinates at the next time are:

[0081] ;

[0082] ;

[0083] For articulated vehicles, combined with Figure 3 and Figure 4 , the following method is used to calculate the position coordinates at the next time:

[0084] ;

[0085] ;

[0086] That is,

[0087] ;

[0088] ;

[0089] ;

[0090] wherein, represents the coordinate at the previous time; represents the coordinate at the next time; represents the heading angle at the previous time; represents the heading angle at the next time; represents the vehicle speed at the previous time; represents the displacement at the previous time; represents the displacement at the next time;

[0091] The newly calculated position, heading angle, and LSTM-predicted state quantities such as speed, acceleration, and turning angle are added to the historical state sequence. At the same time, the earliest historical data is removed, and a sliding window of fixed length N is maintained. Through the above data processing process, the system realizes a complete simulation link from the control command to the vehicle trajectory, which not only retains the complex dynamic characteristics learned by the LSTM, but also guarantees the physical rationality of the trajectory through the kinematic model.

[0092] The above description of the application and its embodiments is illustrative and not restrictive, and the application can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it, without departing from the spirit or essential characteristics of the application, similar structural forms and embodiments can be designed without creative design, which should belong to the protection scope of the application. In addition, the word "comprising" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The words "first", "second" and the like are used to indicate names, and do not mean any particular order.

Claims

1. A mining card simulation method based on log data, characterized in that, include: S1, acquire the mining truck's operation log data under multiple operating conditions; the operating conditions include: no load, 1 / 4 load, 1 / 2 load, 3 / 4 load, and full load; the operation log data includes: control commands, vehicle status data, and load data; S2, extract nonlinear response feature data based on the operation log data: Extract the actual acceleration values ​​generated by the vehicle after inputting the same throttle or brake control command under different load conditions, and use them as acceleration response data. Extract the actual changes in the front wheel angle or articulation angle under the same steering control command input under different load conditions, and use them as steering angle change data; S3, align control commands and nonlinear response characteristic data according to timestamps; The training set is constructed by using historical control commands and historical vehicle states from N consecutive cycles as input features and the vehicle state of the current cycle as the output label. S4. Train the LSTM model using the training set. The input of the LSTM model includes: historical control command sequence, historical vehicle state sequence, current load value and response delay time. The output is the predicted current vehicle state. S5 uses a trained LSTM model, takes the current control command, historical vehicle state and current load value as input, and outputs the predicted vehicle response state; the vehicle response state includes vehicle speed, acceleration, front wheel steering angle or articulation angle; Input the predicted vehicle speed, acceleration, front wheel steering angle or articulation angle into the vehicle kinematics model to calculate the vehicle's position and attitude under the current load conditions; S4, training an LSTM model using the training set, includes: Preprocess the constructed training set; Constructing an LSTM network model: The input layer receives four types of features: historical control command sequence, historical vehicle status sequence, current load value, and response delay time sequence. The input dimension is N*D, where N is the length of the historical sequence and D is the feature dimension. The hidden layer uses multi-layer LSTM units to obtain the temporal dependencies under different load conditions; where the temporal dependencies represent the time delay pattern between control commands and vehicle state responses. The output layer outputs the predicted current vehicle state; Set the loss function and optimizer; The constructed LSTM network model is trained using the preprocessed training set; The constructed training set is preprocessed, including: The load data is categorized into: unloaded, 1 / 4 load, 1 / 2 load, 3 / 4 load, and fully loaded, and coded as discrete values ​​of 0, 1, 2, 3, and 4. Map the control command sequence, vehicle status sequence, and response delay time sequence to the interval between 0 and 1; Interpolate the data between adjacent load levels to generate virtual samples of intermediate load states to enhance the model's adaptability to continuous load changes.

2. The mining card simulation method based on log data according to claim 1, characterized in that: Vehicle status data includes: vehicle position coordinates, heading angle, speed, acceleration, front wheel steering angle or articulation angle, gear, throttle opening, and brake pressure.

3. The mining card simulation method based on log data according to claim 2, characterized in that: Control commands include longitudinal control commands and lateral control commands; Longitudinal control commands include: throttle opening, brake pressure, desired acceleration, and gear; For the Ackermann vehicle, lateral control commands include front wheel angle or steering wheel angle; For articulated vehicles, lateral control commands include the opening degree of the left and right hydraulic cylinders or the articulation angle.

4. The mining card simulation method based on log data according to claim 3, characterized in that: Before step S2, the following is also included: Remove invalid data that does not affect changes in vehicle status from control commands. Invalid data includes: accelerator or brake operation logs in neutral. Remove abnormal data introduced by log recording errors. Abnormal data includes: speed values, acceleration values, or steering angle values ​​that exceed physical limits.

5. The mining card simulation method based on log data according to any one of claims 1 to 4, characterized in that: S3, construct the training set, including: Timestamp matching is performed on control command data and vehicle status data; The control command is issued at time t1, and the response time t2 is when the vehicle state actually changes. The actual change in vehicle state refers to a change in speed, acceleration, front wheel angle, or articulation angle relative to the state value before the control command was issued. Based on t2 and t1, calculate the vehicle state response delay time Δt; where the response delay time includes: acceleration / braking response delay time under different load conditions, and steering system response delay time; for articulated vehicles, it also includes the articulation mechanism response delay time; Data with different sampling frequencies are resampled, and the resampling frequency is T. Training samples are constructed using a sliding window method, with a window size of N periods and a sliding step of 1 period; each training sample includes: Input: Control command sequence from time t-N+1 to time t, vehicle status sequence, current load value, and response delay time sequence; Output: Vehicle state at time t+1.

6. The mining card simulation method based on log data according to claim 1, characterized in that: Loss function: ; in, and These are predicted and actual vehicle speeds, respectively. and These are the predicted and actual accelerations, respectively. and These are the predicted and actual front wheel steering angles or hinge angles, respectively. This represents the rate of change of velocity calculated based on the predicted velocity. This represents the rate of change of velocity calculated based on the actual velocity. These are weighting coefficients, set according to the importance and dimensional differences of different state variables.

7. The mining card simulation method based on log data according to claim 6, characterized in that: S5 calculates the vehicle's position and attitude under the current load conditions, including: For the Ackermann vehicle, the kinematic model adopts the Ackermann steering geometry model; For articulated vehicles, the kinematic model is adopted.

8. A mining card simulation system based on log data, used to implement the method according to any one of claims 1 to 7, characterized in that, include: The data acquisition module acquires the mining truck's operating log data under multiple operating conditions. These operating conditions include: no load, 1 / 4 load, 1 / 2 load, 3 / 4 load, and full load. The operating log data includes: control commands, vehicle status data, and load data. The feature extraction module is used to extract nonlinear response feature data based on the operation log data: extracting the actual acceleration value generated by the vehicle after the same throttle or brake control command input under different load conditions, as acceleration response data; extracting the actual change value of the front wheel steering angle or articulation angle after the same steering control command input under different load conditions, as steering angle change data. The training set construction module is used to align control commands and nonlinear response feature data according to timestamps; it takes historical control commands and historical vehicle states of N consecutive cycles as input features and the vehicle state of the current cycle as output labels to construct the training set. The model training module is used to train the LSTM model using the training set. The input of the LSTM model includes: historical control command sequence, historical vehicle state sequence, current load value and response delay time, and the output is the predicted current vehicle state. The simulation prediction module is used to take the current control command, historical vehicle state and current load value as input from the trained LSTM model, and output the predicted vehicle response state; the predicted vehicle speed, acceleration, front wheel steering angle or articulation angle are input into the vehicle kinematics model to calculate the vehicle's position and attitude under the current load conditions.

Images