AGV relative position control method based on kernel function prediction

Abstract

The present application relates to the control field, specifically to a kind of AGV relative position control method based on kernel function prediction, method includes: AGV relative motion state feature is collected and constructs timing sequence and expected timing sequence;For each feature, based on the bandwidth adjustment coefficient of multiple cycle length analysis calculation, the feature adaptive bandwidth adjustment is carried out to Gaussian kernel function, and the adaptive kernel similarity between current and historical time is calculated;Based on the similarity and Gaussian process regression model, obtain the initial prediction value under each cycle length, and are fused into comprehensive prediction value by confidence weight;Finally, according to the deviation of comprehensive prediction value and expected value, the relative position of AGV is closed-loop controlled.The present application realizes feature adaptive kernel similarity calculation and multiple cycle prediction fusion, and improves the control precision and robustness under the disturbance such as tire slip, load change etc.

Description

Technical Field

[0001] This invention relates to the field of control. More specifically, this invention relates to an AGV relative position control method based on kernel function prediction. Background Technology

[0002] AGV (Automated Guided Vehicle) is a key piece of equipment for realizing automated material handling. It is often used in scenarios such as automated parking garages and ports, and its high-precision relative position control is crucial.

[0003] The Gaussian kernel function is a technique that maps input data to a high-dimensional feature space, handling nonlinear problems in the original space by constructing linear relationships in the high-dimensional space. In AGV relative position control, the kernel function can be used to process nonlinear features in sensor data, such as path curvature and obstacle shape, thereby more accurately describing the relationship between the AGV and its surrounding environment.

[0004] However, existing Gaussian kernel functions assume that each positional feature dimension (such as lateral and longitudinal deviations) contributes equally when calculating similarity. This ignores the different impacts of actual disturbances (such as sudden increases in longitudinal error due to inertia) on the reliability of different dimensions. This deficiency leads to inaccurate similarity measurements, causing biases in control predictions and potentially resulting in positional errors or even collision risks in densely packed multi-vehicle scenarios. Summary of the Invention

[0005] To address the aforementioned technical problem of insufficient AGV control precision, this invention provides the following technical solution.

[0006] An AGV relative position control method based on kernel function prediction includes:

[0007] The AGV that needs to be controlled is marked as the target AGV. Multiple state features of the relative motion between the target AGV and the reference AGV are collected. The feature vector at each time step is constructed, and the time series sequence and expected time series sequence corresponding to each feature are obtained.

[0008] For each feature, perform the following operation to obtain a combined predicted value for that feature:

[0009] Based on multiple preset period lengths, the bandwidth adjustment coefficient of the feature is calculated at any period length according to the difference between the time series sequence of the feature and the expected time series sequence. According to the bandwidth adjustment coefficient, the bandwidth of the Gaussian kernel function is adaptively adjusted, and the adaptive kernel similarity of the feature vector between the current time and the historical time is calculated. Based on the adaptive kernel similarity and the Gaussian process regression model, the initial predicted value of the feature at each period length is obtained, and the initial predicted values ​​are weighted and fused to obtain the comprehensive predicted value of the feature.

[0010] The relative position of the target AGV is controlled based on the deviation between the comprehensive predicted value and the expected value of each feature.

[0011] Optionally, the feature vector includes:

[0012] The relative position information, relative heading angle information, and respective linear velocity and angular velocity information between the target AGV and the reference AGV.

[0013] Optionally, obtaining the bandwidth adjustment coefficient includes:

[0014] The feature error sequence is obtained based on the time sequence of the feature and the expected time sequence. The mean of the absolute values ​​of the elements in the feature error sequence is used as the average error. The first degree of slip is calculated based on the ratio of the absolute value of each element to the average error.

[0015] To calculate the standard deviation of the feature error sequence, a sliding window is used to divide the feature error sequence based on a period length, and the standard deviation of all elements within the window is taken as the window standard deviation.

[0016] Calculate the ratio of the standard deviation of each window to the standard deviation of the feature error sequence, and sum all the ratios and take the average to obtain the second degree of slippage;

[0017] The bandwidth adjustment coefficient is obtained by adding the first slip degree and the second slip degree together and averaging them.

[0018] Optionally, obtaining the adaptive kernel similarity includes:

[0019] Get a preceding window containing the current time, and select any time other than the current time within the preceding window as the target time;

[0020] Based on the statistical information of the feature error sequence of each feature within the preceding window, a comprehensive weight is assigned to each feature; the comprehensive weight is the average of the sum of the calculated first similarity weight and the second similarity weight.

[0021] Based on the combined weight of each feature, the reciprocal of the bandwidth adjustment coefficient, and the initial bandwidth value of the Gaussian kernel function, the adaptive kernel similarity between the feature vectors at the current time and the target time is calculated.

[0022] Optionally, obtaining the adaptive kernel similarity includes:

[0023] Get a preceding window containing the current time, and select any time other than the current time within the preceding window as the target time;

[0024] The initial bandwidth of the Gaussian kernel function is adjusted based on the reciprocal of the bandwidth adjustment coefficient corresponding to each feature, resulting in the adjusted bandwidth of the Gaussian kernel function.

[0025] Calculate the distance components of each feature between the current time and the target time based on the adjusted Gaussian kernel function bandwidth;

[0026] Calculate the negative exponential function value of the sum of the distance components of all features to obtain the adaptive kernel similarity between the feature vectors at the current time and the target time.

[0027] Optionally, obtaining the weights in the weighted fusion includes:

[0028] For each period length, a preceding window containing the current time is obtained. The mean of all adaptive kernel similarities within the preceding window at the current time is calculated. The ratio of the mean of the adaptive kernel similarities under a single period length to the sum of the mean of the adaptive kernel similarities under all period lengths is used as the weight of the prediction confidence corresponding to that period length.

[0029] Optionally, obtaining the first similarity weight includes:

[0030] Calculate the standard deviation of the feature error sequence of each feature within the preceding window, and use the ratio of the standard deviation of the feature error sequence of each feature to the average of the standard deviations of the feature error sequences of all features as the first similarity weight of that feature.

[0031] Optionally, obtaining the second similarity weight includes:

[0032] Calculate the cumulative absolute value of the feature error sequence of each feature within the preceding window, and use the ratio of the cumulative value of each feature to the mean of the cumulative values ​​of all features as the second similarity weight of that feature.

[0033] Optionally, the adjusted Gaussian kernel function bandwidth is the product of the reciprocal of the bandwidth adjustment coefficient and the initial bandwidth.

[0034] The beneficial effects of this invention are:

[0035] This invention introduces a bandwidth adjustment coefficient to adaptively adjust the bandwidth of the Gaussian kernel function. Combined with a multi-cycle length prediction mechanism, this effectively improves the accuracy and adaptability of AGV relative position control. Specifically, for each feature, the bandwidth adjustment coefficient is dynamically calculated based on the difference between its time series and the expected time series, enabling the Gaussian kernel function to more accurately reflect the local characteristics of feature changes. Using a Gaussian process regression model based on adaptive kernel similarity, initial predicted values ​​for different cycle lengths are obtained, and a weighted fusion is used to obtain a comprehensive predicted value, thereby enhancing the robustness of the prediction and its ability to characterize complex motion patterns. Finally, control decisions are made based on the deviation between the comprehensive predicted value and the expected value for each feature, achieving more accurate and stable tracking and control of the AGV's relative position, reducing control errors, and improving the system's response performance and reliability in dynamic environments. Attached Figure Description

[0036] Figure 1 This is a flowchart of steps S1-S5 in an AGV relative position control method based on kernel function prediction according to an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] The application scenarios of this invention are: relative position control environments of AGVs with strong nonlinear disturbances (such as tire slippage, slippery ground, load inertia, etc.) such as automated parking garages and heavy-duty AGVs in ports, and include at least two AGVs.

[0039] Reference Figure 1 A relative position control method for AGVs based on kernel function prediction includes steps S1-S5, as detailed below:

[0040] S1: Mark the AGV that needs to be controlled as the target AGV, collect multiple state features of the relative motion between the target AGV and the reference AGV, construct the feature vector at each time step, and obtain the time series sequence and expected time series sequence corresponding to each feature.

[0041] In AGV relative position control, relying solely on data from a single sensor or at a single moment is insufficient to accurately describe the system's dynamic behavior. Traditional methods directly use fixed features for similarity calculations, ignoring the actual differences in the contributions of different features during motion. Therefore, it is necessary to collect multi-source time-series data and combine it with the system's desired motion trajectory to construct a complete feature vector and expected sequence.

[0042] Specifically, the AGV that needs to be controlled is marked as the target AGV, and another AGV is used as the reference AGV. The relative motion state characteristics between the two are collected, including: relative lateral position, relative longitudinal position, relative heading angle, target AGV linear velocity, reference AGV linear velocity, target AGV angular velocity, and reference AGV angular velocity.

[0043] The feature vector for each of the seven feature values ​​at each time point is constructed by performing conventional preprocessing operations on the seven feature values ​​mentioned above.

[0044] In addition, from system startup to the current moment, all historical values ​​of each feature constitute the time series sequence of that feature. At the same time, the expected time series sequence of each feature from system startup to the current moment is obtained from the control system. This sequence comes from the upper-level path planning and motion control system.

[0045] S2: Based on multiple preset period lengths, calculate the bandwidth adjustment coefficient of each feature at any period length according to the difference between the time series sequence of each feature and the expected time series sequence; based on the bandwidth adjustment coefficient, adjust the bandwidth of the Gaussian kernel function adaptively.

[0046] Traditional Gaussian kernel functions use a fixed bandwidth, assuming all features contribute equally to similarity. However, in real-world motion, different features are affected by disturbances to varying degrees. For example, lateral position errors are larger during turns, while longitudinal position fluctuates significantly during starts and stops. Therefore, the bandwidth needs to be dynamically adjusted based on the magnitude of the error and the degree of fluctuation for each feature.

[0047] Furthermore, AGVs experience different motion patterns at different time scales during operation, such as short-term start-stop jitter, medium-term turning adjustments, and long-term trajectory tracking. To capture these dynamic characteristics across multiple time scales, this invention presets various cycle lengths (such as 5, 7, and 10 time intervals) and performs analysis based on different cycles to achieve comprehensive modeling of motion details and trends.

[0048] Specifically, for any given feature, perform the following operations:

[0049] Calculate the feature error sequence of the feature, which is the difference between the corresponding elements of the time series sequence of the feature and the expected time series sequence.

[0050] Calculate the mean of the absolute values ​​of all elements in the above feature error sequence as the average error; calculate the ratio of the absolute value of each element to the average error, sum all the ratios and take the average to obtain the first degree of slippage of the feature.

[0051] Calculate the standard deviation of the feature error sequence. Select any of the above period lengths, such as 5 time points, and divide the feature error sequence into multiple sliding windows. Use the standard deviation of all elements in the window as the window standard deviation. Calculate the ratio of the standard deviation of each window to the standard deviation of the feature error sequence, and sum all the ratios and take the average to obtain the second degree of slip.

[0052] The bandwidth adjustment coefficient of this feature under this period length is obtained by adding the first slip degree and the second slip degree together and averaging them.

[0053] This allows us to obtain the bandwidth adjustment coefficients for all features at each cycle length.

[0054] The greater the first slip, the higher the overall error level of the feature, and the bandwidth should be reduced to enhance the distinguishability of the feature; the greater the second slip, the stronger the local fluctuation of the feature, and the bandwidth should be reduced to suppress noise and transient disturbances.

[0055] Combining the first and second slip degrees, the larger the bandwidth adjustment coefficient, the more it indicates that the feature has both overall deviation and strong fluctuation. Its kernel bandwidth should be significantly reduced to make the Gaussian kernel function more sensitive to the feature and improve its adaptability to nonlinear disturbances.

[0056] S3: Calculate the adaptive kernel similarity of feature vectors between the current time and historical time based on the adjusted Gaussian kernel function bandwidth.

[0057] In existing technology, the formula for calculating the standard Gaussian kernel function is:

[0058]

[0059] However, in order to address the problem that AGVs exhibit significant differences in errors of different characteristics under strong disturbance scenarios, the present invention makes an adaptive improvement to the above formula.

[0060] In one embodiment, a bandwidth adjustment coefficient is introduced to achieve feature adaptation:

[0061]

[0062] In the formula, The adaptive kernel similarity between the feature vectors at the current time and the target time. For the number of features, For the current moment The value of each feature, For the target time The value of each feature, For the initial bandwidth, For the first The reciprocal of the bandwidth adjustment coefficient for each feature. It is an exponential function with the natural number e as its base. It is the sum of the distance components of all features.

[0063] Specifically, a preceding window containing the current time is obtained, meaning the endpoint of the preceding window is the current time, and any time within the preceding window other than the current time is selected as the target time.

[0064] pass The bandwidth of each feature is adjusted independently; the larger the error, the more... The smaller the value, the smaller the denominator, and the greater the contribution of this feature to the similarity. This allows the Gaussian kernel function to adaptively focus on high-error features when calculating similarity, thereby improving its ability to identify key perturbations.

[0065] Although the above operations have achieved bandwidth adaptation, the difference terms of each feature are still assumed by default. The contribution to overall similarity is adjusted solely by bandwidth. In actual motion, fluctuations or accumulated errors in certain features (such as lateral position) have a greater impact on safety and control and should be given higher weight.

[0066] In another embodiment, a preceding window containing the current time is obtained, and any time other than the current time within the preceding window is selected as the target time.

[0067] Calculate the standard deviation of the feature error sequence of each feature within the preceding window, and use the ratio of the standard deviation of the feature error sequence of each feature to the average of the standard deviations of the feature error sequences of all features as the first similarity weight of that feature; calculate the cumulative value of the absolute error of the feature error sequence of each feature within the preceding window, and use the ratio of the calculated cumulative value of each feature to the average of the cumulative values ​​of all features as the second similarity weight of that feature.

[0068] Furthermore, the average of the sum of the first similarity weight and the second similarity weight is used as the comprehensive weight of this feature.

[0069] Based on the combined weights of each feature, the reciprocal of the bandwidth adjustment coefficient, and the initial bandwidth value of the Gaussian kernel function, the adaptive kernel similarity between the feature vectors at the current time and the target time is calculated. That is, it satisfies the following relationship:

[0070]

[0071] In the formula, The adaptive kernel similarity between the feature vectors at the current time and the target time. For the number of features, For the first The combined weight of each feature, For the current moment The value of each feature, For the target time The value of each feature, For the initial bandwidth, For the first The reciprocal of the bandwidth adjustment coefficient for each feature. It is an exponential function with the natural number e as its base.

[0072] A higher first similarity weight indicates greater volatility of the feature; a higher second similarity weight indicates greater cumulative error of the feature. The combined weight integrates both, further emphasizing the focus on highly volatile and error-prone features during kernel similarity calculation. Increasing the combined weight enhances the feature's contribution to similarity calculation, thus more accurately reflecting key state changes in prediction.

[0073] It should be noted that S3 above is performed with a certain period length. By repeating the above operation for each period length, the adaptive kernel similarity of the feature vectors between the current time and the historical time can be obtained for each period length.

[0074] S4: Based on the adaptive kernel similarity and Gaussian process regression model, the initial predicted values ​​of each feature at each period length are obtained, and the initial predicted values ​​are weighted and fused to obtain the comprehensive predicted value of the feature.

[0075] In S2 above, multiple period lengths were introduced to capture motion characteristics at different time scales, and in S3 above, adaptive kernel similarity calculation based on each period length was completed. However, if prediction is based solely on a single period length, multi-scale information cannot be fully utilized, and inappropriate period selection may affect prediction accuracy. For example:

[0076] If only short periods are used, although they can respond to instantaneous disturbances, they are easily affected by noise and the prediction results fluctuate greatly; if only long periods are used, although the trend is stable, they cannot capture sudden changes in the state in time.

[0077] Therefore, in order to achieve comprehensive and robust prediction of AGV motion status, a multi-cycle prediction fusion mechanism is proposed: that is, predictions are made separately at each cycle length, and then weighted and fused according to the confidence of the prediction results at each cycle length, thereby combining the detail sensitivity of short cycles with the trend stability of long cycles.

[0078] Specifically, for each period length, taking the aforementioned preceding window as an example, the feature vectors of each time step within the preceding window, the feature vector of the current time step, and the adaptive kernel similarity between the current time step and each time step within the preceding window step are input into the trained Gaussian process regression model. The model outputs the initial predicted values ​​of each feature at the next time step of the current time step.

[0079] Furthermore, simply averaging the prediction results for each cycle length would ignore the reliability differences between different cycles at different times. Therefore, it is necessary to calculate the prediction confidence weight for each cycle length to assess the reliability of the prediction.

[0080] For each period length, calculate the mean of all adaptive kernel similarities within the preceding window at the current time. Use the ratio of the mean of adaptive kernel similarities for a single period length to the sum of the mean of adaptive kernel similarities for all period lengths as the prediction confidence weight for that period length.

[0081] A higher prediction confidence weight indicates that the historical characteristics are more consistent with the current state over a given period, and the prediction result is more reliable.

[0082] Furthermore, the initial predicted values ​​for each period length are weighted and summed according to their prediction confidence weights to obtain the comprehensive predicted value of the feature at the next time step.

[0083] S5: Control the relative position of the target AGV based on the deviation between the comprehensive predicted value and the expected value of each feature.

[0084] After obtaining the comprehensive predicted values ​​of various features, they need to be compared with the expected values, and control commands are generated based on the deviation to achieve precise relative position tracking of the AGV. Traditional methods often rely on feedback control of the current error, resulting in response lag. This invention introduces feedforward prediction information to compensate for expected deviations in advance, thereby improving the dynamic performance of the system.

[0085] Specifically, the comprehensive predicted value and expected value of each feature are input into the position loop and speed loop controllers. The controllers generate control signals based on the deviation, driving the target AGV to perform corresponding position and speed adjustments.

[0086] The controller described above can be one or a combination of common closed-loop controllers such as PID controllers, fuzzy controllers, and model predictive controllers. This invention does not specifically limit the type of controller.

[0087] This invention systematically solves the problems of large errors and easy collisions in traditional AGV relative position control under strong nonlinear disturbances through a triple mechanism of feature adaptive bandwidth adjustment, multi-cycle similarity fusion, and feedforward-feedback composite control. Each step is interconnected, forming a fully adaptive closed loop from data acquisition, feature analysis, predictive modeling to control execution, improving the system's reliability, accuracy, and adaptability in complex scenarios.

[0088] It should be noted that the Gaussian kernel function used in this invention has a basic form and calculation method that are well-known technologies. For specific implementation, please refer to relevant classic literature or open-source tool libraries.

[0089] The training process of a Gaussian process regression model includes steps such as kernel function selection, hyperparameter optimization (e.g., bandwidth, noise variance), and model fitting based on historical data, all of which are standard procedures for Gaussian process regression.

[0090] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An AGV relative position control method based on kernel function prediction, characterized in that, The method comprises the following steps: Mark the AGV to be controlled as a target AGV, collect a plurality of state features of the relative motion between the target AGV and a reference AGV, construct a feature vector at each time, and obtain a time sequence corresponding to each feature and an expected time sequence; For each feature, the following operations are performed to obtain a comprehensive prediction value of the feature: Based on a plurality of preset cycle lengths, a bandwidth adjustment coefficient of the feature under any cycle length is calculated according to the difference between the time sequence and the expected time sequence of the feature; the bandwidth of the Gaussian kernel function is adaptively adjusted according to the bandwidth adjustment coefficient, and an adaptive kernel similarity between the feature vectors between the current time and the historical time is calculated; the initial prediction value of the feature under each cycle length is obtained based on the adaptive kernel similarity and the Gaussian process regression model, and each initial prediction value is weighted and fused to obtain the comprehensive prediction value of the feature; The bandwidth adjustment coefficient is obtained by: According to the time sequence and the expected time sequence of the feature, a feature error sequence of the feature is obtained, the mean of the absolute values of the elements in the feature error sequence is taken as the average error, and the first slip degree is calculated based on the ratio of the absolute value of each element to the average error; The standard deviation of the feature error sequence is calculated, the feature error sequence is divided into sliding windows based on a cycle length, and the standard deviation of all elements in the window is taken as the window standard deviation; The ratio of each window standard deviation to the standard deviation of the feature error sequence is calculated, and the sum of all ratios is averaged to obtain the second slip degree; The first slip degree and the second slip degree are added and averaged to obtain the bandwidth adjustment coefficient; Based on the deviation between the comprehensive prediction value of each feature and its expected value, the relative position of the target AGV is controlled.

2. The AGV relative position control method based on kernel function prediction according to claim 1, characterized in that, The feature vector comprises: The relative position information, the relative heading angle information, and the respective linear velocity and angular velocity information between the target AGV and the reference AGV.

3. The AGV relative position control method based on kernel function prediction according to claim 1, characterized in that, The adaptive kernel similarity is obtained by: A pre-window containing the current time is obtained, and any time except the current time in the pre-window is selected as a target time; According to the statistical information of the feature error sequence of each feature in the pre-window, a comprehensive weight is assigned to each feature; the comprehensive weight is the average of the sum of the first similarity weight and the second similarity weight; According to the comprehensive weight of each feature, the reciprocal of the bandwidth adjustment coefficient, and the initial bandwidth value of the Gaussian kernel function, the adaptive kernel similarity between the feature vectors of the current time and the target time is calculated.

4. The AGV relative position control method based on kernel function prediction according to claim 1, characterized in that, The adaptive kernel similarity is obtained by: A pre-window containing the current time is obtained, and any time except the current time in the pre-window is selected as a target time; Based on the reciprocal of the bandwidth adjustment coefficient of each feature, the initial bandwidth of the Gaussian kernel function is adjusted to obtain the adjusted Gaussian kernel function bandwidth; According to the adjusted Gaussian kernel function bandwidth, the distance components of each feature between the current time and the target time are calculated; The negative exponential function value of the sum of the distance components of all features is calculated to obtain the adaptive kernel similarity between the feature vectors of the current time and the target time.

5. The AGV relative position control method based on kernel function prediction according to claim 1, characterized in that, The weight in the weighted fusion is obtained by: For each cycle length, a pre-window containing the current time is obtained, the mean of all adaptive kernel similarities in the pre-window of the current time is calculated, and the ratio of the mean of the adaptive kernel similarity under the single cycle length to the sum of the means of the adaptive kernel similarities under all cycle lengths is taken as the prediction confidence weight corresponding to the cycle length.

6. The AGV relative position control method based on kernel function prediction according to claim 3, characterized in that, The obtaining of the first similarity weight comprises: The standard deviation of the feature error sequence of each feature in the pre-window is calculated, and the ratio of the standard deviation of the feature error sequence of each feature to the average of the standard deviations of the feature error sequences of all features is taken as the first similarity weight of the feature.

7. The AGV relative position control method based on kernel function prediction according to claim 3, characterized in that, The obtaining of the second similarity weight comprises: The cumulative value of the absolute error value of the feature error sequence of each feature in the pre-window is calculated, and the ratio of the cumulative value of each feature to the average of the cumulative values of all features is taken as the second similarity weight of the feature.

8. The AGV relative position control method based on kernel function prediction according to claim 4, characterized in that, The adjusted Gaussian kernel function bandwidth is the product of the inverse of the bandwidth adjustment coefficient and the initial bandwidth.

Images