Method and apparatus for providing data-based prediction model in consideration of dynamic characteristics
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2024-10-28
- Publication Date
- 2026-08-04
AI Technical Summary
然而，在预测模型的输入侧上被考虑的输入参量值的时间范围的扩大通常受限，因为这将显著增加预测模型的输入张量的维度
Smart Images
Figure CN122514764A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for providing predicted values through a data-based predictive model to map a physical system, for example, to implement a virtual sensor or perform anomaly detection. The invention further relates to a suitable training method for the data-based predictive model to take into account, particularly, the dynamic characteristics of the physical system in relation to the predicted parameters to be modeled. Background Technology
[0002] Data-based predictive models can be constructed, for example, as NARX models, which are based on neural networks. Such NARX models represent a scheme for using data-based models in dynamic applications where the predictive model forecasts time-series data.
[0003] For evaluation, the NARX model is fed input parameter values for multiple input parameters at the current time step and input parameter values for at least one input parameter at one or more past time steps. Furthermore, the output value of the prediction model can be fed back to the input of the prediction model, which models the predicted values for subsequent time steps, if necessary. Typically, such data-based sensor models are trained in a supervised manner, i.e., on a series of training datasets containing short time series of input parameters up to a specific time step, and labeled with the actual predicted values of the parameter to be modeled at the next time step.
[0004] By using the NARX model, it is therefore possible to consider the temporal correlation of different physical effects. However, with conventional supervised training of data-based sensor models, for example by means of loss values, specific characteristics of the system are not considered, which are given based on average error, such as mean squared error, log-likelihood, or mean absolute error.
[0005] Therefore, an extension scheme is needed to predict system behavior, which enables modeling of specific time-dependent characteristics. However, expanding the time range of input parameter values considered on the input side of the prediction model is generally limited, as this would significantly increase the dimensionality of the prediction model's input tensors.
[0006] Therefore, the object of the present invention is to provide an improved prediction model for predicting one or more prediction parameters for one or more future time steps, wherein the complexity of the prediction model can be maintained or even reduced while improving prediction accuracy. Summary of the Invention
[0007] This task is accomplished by the method for providing a data-based prediction model for at least one prediction parameter as described in claim 1, and by the corresponding device as described in the parallel claims.
[0008] Alternative design options are given in the dependent claims.
[0009] According to the first aspect, a method is provided for providing a data-based predictive model for modeling a technical system, wherein the predictive model is constructed to provide predicted values for time steps following an evaluation time step based on an input tensor (such as an input vector), wherein the input tensor includes, in addition to the value of at least one input parameter at the evaluation time step, at least one value of the at least one input parameter at one or more time steps prior to the evaluation time step within a pre-given time range, wherein the predictive model is trained in training steps based on a corresponding total loss, particularly using a gradient-based training method, wherein for each training step, the total loss is determined using the following steps: - Provide a training dataset, which is formed by time series of state parameters of the technical system, wherein the training dataset includes an input tensor and label values, wherein the input tensor has at least one value of at least one state parameter at a specific reference time point and at least one value of at least one state parameter at a time step before the specific reference time point, wherein the label values depend on the state parameter at a time step after the reference time point; - The total loss is determined based on multiple training datasets within a training time window, the training datasets having successively acquired values of the state parameter, wherein the main term is determined by the average error, and at least one additional term is provided that maps the temporal characteristics of the time series of the input parameter within the training time window.
[0010] Predictive models can be used in many fields. Thus, models for virtual sensors used in regulation systems, or models for anomaly identification, rely on pre-determined state parameters for corresponding state values at one or more subsequent time steps, which give the state of the physical system. This enables improved regulation or earlier anomaly identification.
[0011] To map temporal dynamics, recurrent models or NARX models can be used to account for time spans. However, these approaches are very expensive in terms of training and computation, and provide highly variable predictions.
[0012] According to the aforementioned method for providing a predictive model, a NARX model is preferably used. This NARX model utilizes multiple past time steps to consider the values of one or more input parameters, and further uses at least one state parameter with respect to the output state value at the next time step for evaluating the input parameters in the next time step. Training of the data-based predictive model is performed in a manner known per se based on a loss value, which can be determined from labeled training data. The loss value can be used for gradient-based training, such as backpropagation.
[0013] Training is performed on a training dataset consisting of time series of input parameters, the time series of which ends at a specific reference time point (reference time step) and is assigned a label. In the training dataset, the label, or label value, gives the target value of the model output and may correspond to one of the input parameters or a state parameter different from the input parameters, and involves time steps after the reference time point.
[0014] Traditionally, a loss value is derived for each training dataset from the average error, such as a mean squared method, which gives the difference between the label value and at least one predicted value of the prediction model for successive time steps. Preferably, training datasets are randomly selected from the time series of training data, such that training datasets are chosen as sequentially as possible to avoid temporally consecutive training datasets. However, this disrupts the sequential structure of the training data.
[0015] In order to take into account the following time characteristics, which depend on the time development of a larger time period than the time range that can be covered by the past time steps, at least one additional term may be considered in addition to the main term mentioned above during the training of the data-based prediction model to determine the loss value.
[0016] It can be specified that a first additional term is determined based on the training dataset that determines the value of the primary term, wherein the first additional term is determined from the training dataset of the training time window based on the sum of the signed differences between the model output and the label value of the training dataset of the training time window. In particular, a first normalization function can be applied to the summation result, the first normalization function being particularly capable of including forming an absolute value (absolute value calculation) or squaring.
[0017] The first additional term can be determined based on the training dataset used to determine the value of the principal term. The additional term is formed from a training dataset of the same time series as the training dataset selected from the defined training time window, which ends at a reference time point used to determine the principal term. The length of the training time window is preferably chosen such that it covers a longer time period than the time period mapped by the time steps of the input parameters used on the input side of the data-based prediction model. Here, the first additional term for the loss value is determined after model evaluation on multiple training datasets within the training time window.
[0018] Furthermore, the first additional term can correspond to the sum of the differences between the model output and the label values of all training datasets considered for the training time window, wherein the summation result can be subjected to a first normalization function, which can correspond to either absolute value (absolute value calculation) or squared value calculation. Summing these differences enables the balancing of positive and negative biases in the predictions.
[0019] It can be specified that a second additional term is determined from the training dataset of the training time window, wherein the second additional term takes into account the variance of the model output of the prediction model evaluated against the training dataset of the training time window, or corresponds to the variance.
[0020] The second supplementary term can be determined based on the training dataset that determines the value of the principal term, similar to the first supplementary term. The second supplementary term is formed from a training dataset of the same time series as the training dataset selected from the defined training time window, which ends at the reference time point of the training dataset used to determine the principal term. The length of the training time window is preferably chosen such that it covers a longer time period than the time period mapped by the time steps of the input parameters used on the input side of the data-based prediction model. Here, the second supplementary term for the loss value is determined after model evaluation on multiple training datasets within the training time window.
[0021] The second additional term may include an unsupervised term that considers, or corresponds to, the variance of the model output evaluated by the prediction model for the training dataset within the training time window. The variance of the evaluated model output may be additionally fed into a second normalization function to penalize high variance of predictions within the training time window during training, i.e., to increase the total loss.
[0022] Furthermore, the total loss can be determined by a weighted sum of the value of the main item and the values of at least the first and second additional items.
[0023] According to one implementation, the training time window may include a larger number of time steps than the number of time steps corresponding to the time range.
[0024] It can be specified that the training time window for the training step is randomly selected from the available acquisition sequences of the state parameters. Therefore, the training time window is randomly selected from the time series of the training dataset, where each training time window takes into account the existing training dataset used to determine the loss value.
[0025] According to another aspect, the use of the prediction model provided by the above method is provided, the prediction model serving as a virtual sensor for estimating the temperature of components of the technical system, particularly an electric motor, based on time series of state parameters of the technical system, particularly current, speed of motion, and / or driving torque or driving force, in order to model the temperature values inside the components of the electric motor, particularly the rotor.
[0026] According to another aspect, an apparatus for performing the above-described method is provided. Attached Figure Description
[0027] The embodiments are described in more detail below with reference to the accompanying drawings. Wherein: Figure 1 A schematic diagram of the NARX model used to predict values for the first time step is shown; and Figure 2 A flowchart illustrating the methods used to train data-based prediction models is shown. Detailed Implementation
[0028] Figure 1 A schematic diagram of a data-based predictive model is shown, constructed as NARX model 1, for evaluating time series of multiple input parameters E1, E2, E3 (of arbitrary number). The architecture of the NARX model is based on a deep neural network. Such time series of input parameters can be provided, for example, by a technical system and represent measurement and / or state parameters of the technical system.
[0029] Input parameters are selected from the state parameters of the technical system. To evaluate the temporal dynamics, the values of input parameters E1, E2, and E3 for the current time step t are fed into the data-based prediction model. Furthermore, for at least one of the input parameters, one or more input parameter values from past time steps t-1...tn can be anlegenized into the prediction model, where n can be pre-defined as the time range. Therefore, for one or more input parameters, the input tensor of the data-based prediction model can consider not only the value at the current time step but also the values for one or more previous time steps t-1...tn. tn...t represents the time range of the temporal dynamics considered in the input tensor of the prediction model.
[0030] Prediction model 1 can be constructed to determine, for the prediction parameter P(t+1) to be modeled, the value of the state parameter to be predicted at the next time step after the reference time point, or to determine the difference between the value of the state parameter to be predicted at the current time step and the value of the state parameter to be predicted at the next time step t+1 after the reference time point, the difference giving the development of the prediction parameter P.
[0031] Typically, such NARX models are designed to consider a time range of n time steps for one or more input parameters E1, E2, E3, sufficient to map the system's unmittelbare dynamic behavior while keeping the dimension of the input tensor within the bounds. While it is possible to map the simple temporal dynamics of the system using a relatively small number of temporally successive values of the input parameters E1, E2, E3, there are long-term temporal characteristics of the system that cannot be described by purely dynamic behavior. To map these temporal characteristics, and when it is desirable to avoid expanding the time range and thus the dimension of the input tensor, the training method for the data-based prediction model 1 specifies that the loss function used to train the data-based prediction model is expanded. To train the data-based prediction model, the following steps are performed according to... Figure 2 The flowchart describes that method in more detail. This method is performed using conventional data processing equipment.
[0032] In step S1, time series of input parameters E1, E2, and E3 are acquired. These time series may include measurement parameters and / or state parameters, and correspond to physical parameters or model parameters. Furthermore, time series of prediction parameters to be modeled by the prediction model are provided synchronously. These prediction parameters may correspond to state parameters on the input side or other state parameters.
[0033] In step S2, a training dataset is first provided from the time series provided in step S1. Each of the training datasets relates to a reference time point (reference time step), wherein the input tensor is formed according to the following time range: the input side of the prediction model is designed for this time range, and the input tensor includes, in addition to the current value of the input parameter with respect to the reference time point, the historical values of the input parameter considered in prediction model 1. For the input tensor, the value of the input parameter or state parameter to be predicted at the next time step t+1 after the reference time point t is assigned as a label value, or the difference between the value of the input parameter or state parameter to be predicted at the current time step and the value of the input parameter or state parameter to be predicted at the next time step t+1, to determine the state parameter or output parameter to be modeled for the time steps after the reference time point.
[0034] The label can also be extended to multiple subsequent time steps.
[0035] For the example above, these training datasets correspond to input parameters E1(tn … t), E2(tn … t) E3(tn … t), where the label values correspond to the parameter P(t+1) to be modeled, and n gives the time range, which is used to consider the dynamic effects in the technical system by predicting model 1.
[0036] Based on the collected time series and each reference time point t, it is possible to create the time order (Abfolge) of the training dataset across the common collection sequence of input parameters / state parameters and label values (a measurement sequence that lasts for a specific time over a time range n).
[0037] First, in step S3, a training dataset with reference time points is randomly selected from the time series. For this training dataset, historical values of the input / state parameters and label values exist, tracing back (in the past) to a defined training time window (with a pre-given number of time steps). This training time window is larger than the time range considered by the prediction model using the input tensor. Preferably, the duration of the acquisition sequence is an integer multiple of the duration of the training time window.
[0038] Randomization can be achieved by randomly selecting a training time window from the acquired sequence. Another loss term could, for example, correspond to the variance of the model output within that training time window.
[0039] Now, in step S4, in order to train the data-based prediction model, a total loss value is determined, which is used within the scope of the backpropagation process to train the data-based prediction model.
[0040] The total loss can be considered as the principal term, determined in a manner known per se. This average error can be given as mean squared error (Euclidean norm) or summation norm, where mean squared error represents the average squared error between the model output and the corresponding label value. The average error is formed over successive time steps across the training time window.
[0041] Furthermore, a first additional loss term can be considered additively, which is derived as the summed signed difference between the model output and the label values on the training dataset for the considered training time window. Additionally, a first normalization function can be applied to the sum as a result.
[0042] A second additional loss term can be set, which takes into account, or corresponds to, the variance of the model output evaluated by the predictive model on the training dataset for the training time window. The variance of the model output evaluated by the predictive model can be additionally fed into a second normalization function to obtain the value of the second additional term. It can be specified that high variance of the model output within the training time window should lead to a higher total loss. In this way, high variance of the predicted values within the training time window can be penalized during training, as the high variance increases the total loss. The second additional loss term is unsupervised and can be used to identify sequences of state parameters that must be assumed to be anomalous due to their high variance.
[0043] By selecting a training time window with a larger number of time steps compared to the time steps considered on the input side of a data-based prediction model, it is possible to consider time characteristics or correlations beyond the purely dynamic nature of the technical system.
[0044] Additional loss terms can be applied to the main term, in particular with appropriate weighting (e.g., additively), in order to obtain the total loss.
[0045] Now, in step S5, a training step is performed based on the total loss.
[0046] Then, the method continues with step S4 to perform additional training steps until a known termination criterion is met.
[0047] The aforementioned predictive model can be used in many technical fields. For example, it can be used as a virtual sensor to estimate the temperature of components of a motor, such as the motor itself, based on time series of state parameters like motor current, speed, and torque, in order to model the temperature values inside the motor components, such as the rotor. This allows for monitoring of the motor temperature, enabling appropriate current limiting or shutdown measures to be taken in case of overheating, i.e., when the temperature limit is reached.
[0048] Especially in regulation, predictive models are suitable as observational models for the parameters to be regulated.
Claims
1. A computer-implemented method for providing a data-based prediction model (1) for modeling a technical system, wherein the prediction model (1) is constructed to provide predictions for time steps following an evaluation time step based on an input tensor, wherein the input tensor includes, in addition to the values of at least one input parameter (E1, E2, E3) at the evaluation time step, at least one value of the at least one input parameter (E1, E2, E3) at one or more time steps prior to the evaluation time step within a pre-given time range, wherein the prediction model is trained in a training step based on a corresponding total loss, particularly using a gradient-based training method (S5), wherein for each training step, the total loss is determined using the following steps: - Provide training datasets (S1, S2), which are formed by time series of state parameters of the technical system, wherein the training datasets include input tensors and label values, wherein the input tensors have at least one value of at least one state parameter at a specific reference time point (t) and at least one value of at least one state parameter at a time step (tn) before the specific reference time point (t), wherein the label values depend on the state parameter at a time step (t+1) after the reference time point (t); - The total loss is determined (S4) based on multiple training datasets within a training time window, the training datasets having successively acquired values of the state parameters, wherein the main term is determined using the average error, and at least one additional term is set, the additional term mapping the temporal characteristics of the time series of the input parameters (E1, E2, E3) within the training time window.
2. The method of claim 1, wherein a first additional term is determined based on a training dataset that determines the value of the primary term, wherein the first additional term is determined from the training dataset of the training time window based on the sum of signed differences between the model output and the label value of the training dataset of the training time window, wherein in particular a first normalization function is applied to the summation result, the first normalization function being particularly capable of including forming an absolute value (taking an absolute value) or taking a square.
3. The method according to any one of claims 1 to 2, wherein a second additional term is determined from the training dataset of the training time window, wherein the second additional term takes into account the variance of the model output of the prediction model evaluated for the training dataset of the training time window, or corresponds to the variance, wherein in particular a second normalization function is additionally applied to the variance of the model output of the prediction model evaluated.
4. The method of claim 3, wherein the total loss is determined by a weighted sum of the value of the main item and the values of at least the first and second supplementary items.
5. The method according to any one of claims 1 to 4, wherein the training time window comprises a larger number of time steps than the number of time steps in the time range.
6. The method of claim 5, wherein a training time window for the training step is randomly selected (S3) from the available acquisition sequences of the state parameters.
7. The method according to claims 1 to 6, wherein the total loss is determined by a weighted sum of the value of the main item and the values of at least the first supplementary item and the second supplementary item.
8. The method according to claims 1 to 7, wherein the mean error is given as mean squared error (Euclidean norm), log-likelihood, or mean absolute error, the mean squared error being the mean squared error between the model output and the corresponding label value.
9. An application of a prediction model (1) provided by the method according to any one of claims 1 to 8, the prediction model being used as a virtual sensor for estimating the temperature of components of a technical system, particularly an electric motor, based on time series of state parameters of the technical system, particularly current, speed of motion and / or driving torque or driving force, so as to model the temperature values of components of the electric motor, particularly the rotor.
10. An apparatus for performing the method according to any one of claims 1 to 8.
11. A computer program product comprising instructions that, when executed by at least one data processing device, cause the data processing device to perform the steps of the method according to any one of claims 1 to 8.
12. A machine-readable storage medium comprising instructions that, when executed by at least one data processing device, cause the data processing device to perform the steps of the method according to any one of claims 1 to 8.