A vehicle motion state quantity prediction method based on local model identification and candidate model screening

Abstract

A vehicle motion state quantity prediction method based on local model identification and candidate model screening belongs to the technical field of vehicle motion state prediction and intelligent data processing. In order to solve the problems of insufficient matching of vehicle motion state quantity prediction model and current local evolution law, unclear candidate model screening and inaccurate fusion weight distribution, the application obtains a vehicle historical motion state quantity time sequence, constructs a local model library, respectively trains an independent identification network and a corresponding prediction model, screens a candidate local model according to a normalized identification probability and a preset cumulative probability threshold, and performs weighted fusion on a candidate prediction result to obtain a vehicle future motion state quantity prediction result, so that the matching degree of the prediction process and the current local evolution law can be improved, and the application is suitable for the fields of vehicle motion state prediction, vehicle motion control, path planning, collision warning and vehicle running state analysis.

Description

Technical Field

[0001] This invention belongs to the field of vehicle motion state prediction and intelligent data processing technology, specifically involving vehicle motion state quantity prediction technology based on local model recognition and candidate model screening. Background Technology

[0002] With the development of vehicle sensor systems and vehicle measurement and control systems, motion state variables such as vehicle position, vehicle speed, vehicle acceleration, and vehicle attitude angle can be continuously collected and formed into historical motion state variable time series. Predicting future vehicle motion state variables based on these time series is a fundamental technical step in vehicle motion control, path planning, collision warning, and vehicle operating status analysis. Current technologies typically utilize historical vehicle motion state variable data to build predictive models and extrapolate changes in motion state variables over future periods based on historical input sequences.

[0003] In existing technologies, common prediction methods include kinematic model-based prediction methods, statistical model-based prediction methods, and data-driven prediction methods based on neural networks. Kinematic models typically extrapolate based on pre-defined laws such as uniform velocity, uniform acceleration, or periodic changes; statistical models usually establish predictive relationships using the lag correlation of historical sequences; and neural network methods learn the mapping relationship between historical input sequences and future output sequences through models such as long short-term memory networks and encoder-decoder structures. Some solutions also combine multiple prediction models to adapt to the different variation patterns of vehicle motion state variables in different local intervals.

[0004] However, during actual driving or movement, the time series of a vehicle's motion state variables are often not continuously dominated by a single changing pattern. Instead, they may exhibit different evolutionary characteristics in different local intervals, such as uniform speed, acceleration, changes in acceleration, or periodic disturbances. If a single prediction model is always used, prediction bias is likely to occur due to the mismatch between the model assumptions and the current local evolutionary patterns. If multiple models are used but only fused with fixed weights or a single optimal model is selected, it is difficult to adaptively select more suitable candidate local models based on the current historical motion state variable sequence segments. When local patterns alternate or boundaries are unclear, the prediction results lack stability.

[0005] Furthermore, existing multi-model prediction methods typically focus on directly fusing the outputs of multiple models, failing to adequately identify the matching relationships between current historical motion state sequence segments and various local models, and neglecting to fully utilize recognition probabilities to screen and weight candidate local models. Therefore, it is necessary to propose a prediction method for vehicle motion state time series. This method involves constructing a local model library, training independent recognition networks and corresponding prediction models separately, and filtering candidate local models based on normalized recognition probabilities and preset cumulative probability thresholds. Finally, the candidate prediction results are fused to improve the matching degree between the prediction process of future vehicle motion state variables and the current local evolution patterns. Summary of the Invention

[0006] A method for predicting vehicle motion state variables based on local model identification and candidate model screening, the method comprising: S1. Obtain the historical motion state time series of the vehicle and preprocess it. Divide the preprocessed historical motion state time series into segments according to the sliding window method to obtain the historical motion state sequence segments to be predicted. The historical motion state time series is collected by the vehicle sensor system or the vehicle measurement and control system. The observed values ​​in the historical motion state time series are physical state quantities that reflect the motion state of the vehicle. S2. Construct a local model library, and generate training sample sets for training the corresponding independent recognition network and the corresponding prediction model based on each local model in the local model library; S3. Establish an independent recognition network for each local model in the local model library, and train the independent recognition network using the corresponding training sample set. The independent recognition network is used to output the recognition probability that the historical motion state sequence segment to be predicted belongs to the corresponding local model. S4. For each local model in the local model library, a corresponding prediction model is established, and the corresponding prediction model is trained using the corresponding training sample set. The corresponding prediction model is used to output the candidate prediction results under the corresponding local model. S5. Input the sequence fragments of the historical motion state variables to be predicted into all independent recognition networks to obtain the recognition probability set corresponding to each local model. S6. Sort the recognition probabilities in the recognition probability set in descending order, and accumulate the recognition probabilities in sequence according to the sorting results until the sum of the first m recognition probabilities is greater than or equal to the cumulative probability threshold, and the sum of the first m-1 recognition probabilities is less than the cumulative probability threshold. Then, determine the first m local models participating in the accumulation as candidate local models. S7. Only call the prediction model corresponding to each of the selected candidate local models to obtain multiple sets of candidate prediction results; S8. Normalize the recognition probability of each candidate local model selected to obtain the fusion weight corresponding to each candidate prediction result. S9. The multiple candidate prediction results are weighted and fused according to the fusion weight to obtain the prediction result of the vehicle's future motion state.

[0007] Furthermore, the physical state quantities mentioned in S1 include one or more of the following: vehicle position, vehicle speed, vehicle acceleration, and vehicle attitude angle.

[0008] Furthermore, the local model library mentioned in S2 includes a uniform speed model, a uniform acceleration model, a uniform judder model, and a periodic model. The uniform speed model, uniform acceleration model, uniform judder model, and periodic model are used to characterize the different evolution laws of the vehicle motion state time series within a local interval.

[0009] Further, S2, which generates training sample sets for training corresponding independent recognition networks and corresponding prediction models based on each local model in the local model library, includes: generating time series samples of vehicle motion state quantities of corresponding types based on each local model, adding a perturbation term to the generated time series samples of vehicle motion state quantities, and using the time series samples of vehicle motion state quantities with the perturbation term as the corresponding training sample set.

[0010] Furthermore, the independent recognition network described in S3 adopts a long short-term memory network structure based on an attention mechanism. The long short-term memory network structure based on an attention mechanism includes an input layer, a first long short-term memory layer, a second long short-term memory layer, an attention layer, a linear mapping layer, and a probability output layer.

[0011] Furthermore, the independent recognition network output recognition probability described in S3 includes: outputting a hidden state sequence through a second long short-term memory layer, weighting the hidden states at each time step in the hidden state sequence through an attention layer to obtain a global representation vector, and outputting the recognition probability of the predicted historical motion state sequence fragment belonging to the corresponding local model through a linear mapping layer and a probability output layer.

[0012] Furthermore, the corresponding prediction model in S4 adopts an encoder-decoder long short-term memory network structure with a one-dimensional convolutional feature extraction layer. The one-dimensional convolutional feature extraction layer is used to perform one-dimensional convolution on the sequence segment of the historical motion state quantity to be predicted to obtain local feature representation. The encoder is used to output the hidden state and memory unit state according to the local feature representation. The decoder is used to output the candidate prediction result under the corresponding local model according to the hidden state and memory unit state.

[0013] Furthermore, S7's statement that only the prediction models corresponding to the selected candidate local models are called includes: the prediction models corresponding to local models that are not determined as candidate local models do not generate candidate prediction results and do not participate in the fusion calculation of the prediction results of the vehicle's future motion state.

[0014] Furthermore, the normalization of the recognition probabilities corresponding to each selected candidate local model in S8 includes: generating fusion weights based on the proportional relationship between the recognition probabilities corresponding to each selected candidate local model, and performing weighted fusion of the multiple candidate prediction results based on the fusion weights to obtain the prediction result of the vehicle's future motion state.

[0015] Based on the same inventive concept, the present invention also proposes a computer storage medium storing a computer program, which, when executed by a processor, implements the above-described method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By acquiring and preprocessing the historical motion state time series of the vehicle, the preprocessed historical motion state time series is divided into segments of the historical motion state sequence to be predicted using a sliding window method. This enables the physical state quantities such as position, velocity, acceleration, and attitude angle collected by the vehicle sensor system or vehicle measurement and control system to form a unified model input, thereby improving the input consistency of subsequent local model recognition and prediction processing.

[0017] By constructing a local model library including uniform speed model, uniform acceleration model, uniform judder model and periodic model, and generating training sample sets for training the corresponding independent recognition network and the corresponding prediction model based on each local model, the evolution law of vehicle motion state time series in different local intervals can be classified and modeled, thereby reducing the prediction bias caused by the mismatch between the single prediction model and the current local evolution law.

[0018] By establishing an independent recognition network for each local model and training the independent recognition network using the corresponding training sample set, the matching relationship between the sequence fragment of the historical motion state quantity to be predicted and each local model can be represented by the recognition probability, thereby improving the problem of insufficient recognition of the current local evolution pattern by existing methods.

[0019] By establishing a corresponding prediction model for each local model and training the corresponding prediction model using the corresponding training sample set, each prediction model can generate candidate prediction results under the corresponding local model, thereby enhancing the correspondence between the candidate prediction results and different local evolution laws.

[0020] By inputting the sequence fragments of the historical motion state variables to be predicted into all independent recognition networks and globally normalizing the recognition probability set corresponding to each local model, the recognition probabilities output by different independent recognition networks can be compared on a unified scale, thus providing a clear probabilistic basis for candidate local model selection and fusion weight allocation.

[0021] By sorting the normalized recognition probabilities and selecting candidate local models from the sorting results according to a preset cumulative probability threshold, the prediction of future vehicle motion state variables no longer depends on a single optimal model or a fixed combination, thereby reducing the risk of model misselection when local modes alternate or mode boundaries are unclear.

[0022] By only calling the prediction models corresponding to each selected candidate local model to generate multiple sets of candidate prediction results, the prediction models that are not determined as candidate local models do not participate in the generation of candidate prediction results and the final fusion calculation, thereby reducing the interference of low-matching local models on the prediction results of the vehicle's future motion state.

[0023] By normalizing the normalized recognition probabilities of each selected candidate local model again to generate fusion weights, and then weighting and fusing multiple candidate prediction results according to the fusion weights, the candidate local models with higher recognition probabilities will have corresponding weights in the final prediction results, thereby improving the matching degree between the prediction results of the vehicle's future motion state and the current local evolution law.

[0024] This invention features clear local model identification, clear candidate local model screening, and a fusion weight allocation corresponding to the identification probability. It can improve the matching degree between the prediction process of future vehicle motion state and the current local evolution law, and is applicable to fields such as vehicle motion state prediction, vehicle motion control, path planning, collision warning, and vehicle operation state analysis. Attached Figure Description

[0025] Figure 1 This is a flowchart of the vehicle motion state quantity prediction method based on local model identification and candidate model screening as described in the implementation method. Figure 2 This is a schematic diagram of the independent recognition network structure corresponding to each local model described in the implementation method; Figure 3 This is a schematic diagram of the prediction model structure corresponding to each local model described in the implementation method; Figure 4 This is a diagram showing the fusion prediction results of vehicle motion state quantities as described in the implementation method. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0027] Implementation Method 1 like Figure 1 As shown, a method for predicting vehicle motion state variables based on local model identification and candidate model screening is described, the method comprising: S1. Obtain the historical motion state time series of the vehicle and preprocess it. Divide the preprocessed historical motion state time series into segments according to the sliding window method to obtain the historical motion state sequence segments to be predicted. The historical motion state time series is collected by the vehicle sensor system or the vehicle measurement and control system. The observed values ​​in the historical motion state time series are physical state quantities that reflect the motion state of the vehicle. S2. Construct a local model library, and generate training sample sets for training the corresponding independent recognition network and the corresponding prediction model based on each local model in the local model library; S3. Establish an independent recognition network for each local model in the local model library, and train the independent recognition network using the corresponding training sample set. The independent recognition network is used to output the recognition probability that the historical motion state sequence segment to be predicted belongs to the corresponding local model. S4. For each local model in the local model library, a corresponding prediction model is established, and the corresponding prediction model is trained using the corresponding training sample set. The corresponding prediction model is used to output the candidate prediction results under the corresponding local model. S5. Input the sequence fragments of the historical motion state variables to be predicted into all independent recognition networks to obtain the recognition probability set corresponding to each local model. S6. Sort the recognition probabilities in the recognition probability set in descending order, and accumulate the recognition probabilities in sequence according to the sorting results until the sum of the first m recognition probabilities is greater than or equal to the cumulative probability threshold, and the sum of the first m-1 recognition probabilities is less than the cumulative probability threshold. Then, determine the first m local models participating in the accumulation as candidate local models. S7. Only call the prediction model corresponding to each of the selected candidate local models to obtain multiple sets of candidate prediction results; S8. Normalize the recognition probability of each candidate local model selected to obtain the fusion weight corresponding to each candidate prediction result. S9. The multiple candidate prediction results are weighted and fused according to the fusion weight to obtain the prediction result of the vehicle's future motion state.

[0028] Furthermore, the physical state quantities mentioned in S1 include one or more of the following: vehicle position, vehicle speed, vehicle acceleration, and vehicle attitude angle.

[0029] By defining the physical state quantities mentioned above, the time series of the vehicle's historical motion state quantities can correspond to the vehicle's actual motion state, thus providing a clear data foundation for local model identification and prediction of the vehicle's future motion state quantities.

[0030] Furthermore, the local model library mentioned in S2 includes a uniform speed model, a uniform acceleration model, a uniform judder model, and a periodic model. The uniform speed model, uniform acceleration model, uniform judder model, and periodic model are used to characterize the different evolution laws of the vehicle motion state time series within a local interval.

[0031] Preferably, the local model library is used to classify and express the different evolution patterns that the vehicle motion state time series may exhibit within local intervals, so that the subsequent independent identification network and corresponding prediction model can be established separately around different local models.

[0032] The aforementioned local model library enables the local evolution patterns of vehicle motion state time series to be categorized and described, thereby enhancing the correspondence between subsequent candidate local model selection and candidate prediction result generation.

[0033] Further, S2, which generates training sample sets for training corresponding independent recognition networks and corresponding prediction models based on each local model in the local model library, includes: generating time series samples of vehicle motion state quantities of corresponding types based on each local model, adding a perturbation term to the generated time series samples of vehicle motion state quantities, and using the time series samples of vehicle motion state quantities with the perturbation term as the corresponding training sample set.

[0034] Preferably, the time series samples of vehicle motion state variables after adding the disturbance term are used to train the corresponding independent identification network and the corresponding prediction model, so that the independent identification network and the corresponding prediction model are both trained based on the data foundation corresponding to the same local model.

[0035] By using the above-mentioned training sample set generation method, the time series samples of vehicle motion state quantities corresponding to each local model can be used to train the corresponding independent recognition network and the corresponding prediction model, thereby improving the correlation between the recognition probability output and the generation of candidate prediction results.

[0036] Furthermore, the independent recognition network described in S3 adopts a long short-term memory network structure based on an attention mechanism. The long short-term memory network structure based on an attention mechanism includes an input layer, a first long short-term memory layer, a second long short-term memory layer, an attention layer, a linear mapping layer, and a probability output layer.

[0037] Preferably, each local model corresponds to an independent recognition network, which is used to identify the matching relationship between the sequence fragment of the historical motion state quantity to be predicted and the corresponding local model.

[0038] Through the aforementioned independent identification network structure, the sequence fragments of historical motion state variables to be predicted can be matched and judged with each local model, thereby providing a basis for the formation of the subsequent normalized identification probability set.

[0039] Furthermore, the independent recognition network output recognition probability described in S3 includes: outputting a hidden state sequence through a second long short-term memory layer, weighting the hidden states at each time step in the hidden state sequence through an attention layer to obtain a global representation vector, and outputting the recognition probability of the predicted historical motion state sequence fragment belonging to the corresponding local model through a linear mapping layer and a probability output layer.

[0040] By using the above-mentioned recognition probability output method, the hidden states of the historical motion state sequence fragments to be predicted at different time steps can be weighted by the attention layer to form a global representation vector, so that the recognition probability output by the independent recognition network can reflect the degree of matching between it and the corresponding local model.

[0041] Furthermore, the corresponding prediction model in S4 adopts an encoder-decoder long short-term memory network structure with a one-dimensional convolutional feature extraction layer. The one-dimensional convolutional feature extraction layer is used to perform one-dimensional convolution on the sequence segment of the historical motion state quantity to be predicted to obtain local feature representation. The encoder is used to output the hidden state and memory unit state according to the local feature representation. The decoder is used to output the candidate prediction result under the corresponding local model according to the hidden state and memory unit state.

[0042] Through the aforementioned corresponding prediction model structure, the sequence fragment of the historical motion state quantity to be predicted can first be processed by one-dimensional convolution to extract local feature representations, and then the encoder and decoder output candidate prediction results, thereby enhancing the correlation between the candidate prediction results and the corresponding local model.

[0043] Furthermore, S6, which involves sequentially accumulating the normalized recognition probabilities according to the sorting results, includes: determining the first m local models participating in the accumulation as candidate local models until the sum of the first m normalized recognition probabilities is greater than or equal to the cumulative probability threshold, and the sum of the first m-1 normalized recognition probabilities is less than the cumulative probability threshold.

[0044] By using the above-mentioned candidate local model screening method, the candidate local model can be determined by the normalized identification probability ranking result and the cumulative probability threshold, thus providing a clear screening basis for subsequently calling only the prediction model corresponding to the candidate local model.

[0045] Furthermore, S7's statement that only the prediction models corresponding to the selected candidate local models are called includes: the prediction models corresponding to local models that are not determined as candidate local models do not generate candidate prediction results and do not participate in the fusion calculation of the prediction results of the vehicle's future motion state.

[0046] By using the above-mentioned prediction model invocation method, the prediction of the vehicle's future motion state is generated only based on the prediction model corresponding to the selected candidate local model, thereby reducing the interference of unselected local models on the final fusion calculation.

[0047] Furthermore, the normalization of the recognition probabilities corresponding to each selected candidate local model in S8 includes: generating fusion weights based on the proportional relationship between the recognition probabilities corresponding to each selected candidate local model, and performing weighted fusion of the multiple candidate prediction results based on the fusion weights to obtain the prediction result of the vehicle's future motion state.

[0048] By using the above-mentioned fusion weight generation and weighted fusion methods, the role of each candidate prediction result in the final prediction result corresponds to the normalized recognition probability of the corresponding candidate local model, thereby improving the matching degree between the prediction result of the vehicle's future motion state and the current local evolution law.

[0049] The method described in this embodiment can be executed by a processor calling a computer program, which can be stored in a computer storage medium. When the computer program is executed by the processor, the above-described method for predicting vehicle motion state variables based on local model recognition and candidate model screening can be implemented.

[0050] Implementation Method 2 This embodiment, based on Embodiment 1, further elaborates on the vehicle motion state quantity prediction method based on local model recognition and candidate model screening. In this embodiment, the historical motion state quantity time series of the vehicle is collected by the vehicle sensor system or vehicle measurement and control system, and the observed values ​​in the historical motion state quantity time series are physical state quantities reflecting the vehicle's motion state. The method constructs a local model library, establishes an independent recognition network and corresponding prediction model for each local model, and screens candidate local models based on the recognition probability ranking results and cumulative probability thresholds. Then, it performs weighted fusion of the candidate prediction results to obtain the predicted result of the vehicle's future motion state quantity.

[0051] Step 1: Obtain the historical motion state time series of the vehicle and perform preprocessing. Obtain the historical motion state time series of the vehicle, denoted as ,in, Indicates the total length of the time series. Indicates the first The observation value at each time point.

[0052] The historical motion state time series of the vehicle is preprocessed, including one or more of the following: resampling, denoising, outlier correction, and normalization. Normalization uses a max-min normalization method, expressed as:

[0053] in, This represents the normalized observation value. Represents the maximum value in the sample sequence. This represents the minimum value in the sample sequence.

[0054] The preprocessed historical motion state time series is divided into a historical input sequence and a predicted target sequence using a sliding window method. The historical window length is set to... The prediction step size is set to Then the first The input and output sequences of the group samples are as follows:

[0055]

[0056] Among them, the historical window length The range of values ​​is Predicting step size The range of values ​​is The range of values ​​for the sliding step size is: .

[0057] Step 2: Building a local model library To address the different evolution patterns that may occur in the time series of vehicle motion state variables within local intervals, a local model library is constructed. This local model library includes at least the following four types of local models: Uniform model

[0058] in, Indicates the initial value. Indicates a constant rate of change. Represents a time variable.

[0059] Uniform acceleration model

[0060] in, Indicates the initial rate of change. It represents constant acceleration.

[0061] Uniform quiescent model

[0062] in, Indicates the initial acceleration. This indicates a constant jump.

[0063] Periodic model

[0064] in, Indicates amplitude, Represents angular frequency. Indicates the initial phase. This indicates the bias term.

[0065] During sample generation, parameters of each local model are sampled within a preset range to form time series samples of vehicle motion state quantities for the corresponding local model. The permissible parameter range can be set according to the observation requirements of the vehicle motion state quantities, where: The range of values ​​is any real number within the range of sample observations; , , , , The range of values ​​is the real number interval that meets the requirements for the magnitude of sequence change; The range of values ​​is positive real numbers; The range of values ​​is positive real numbers; The range of values ​​is ; The range of values ​​is .

[0066] Step 3: Generate the training sample set corresponding to each local model Based on the local models described in step two, time series samples of vehicle motion state quantities of corresponding types are generated. A perturbation term is added to the generated samples to simulate measurement noise and local fluctuations during the acquisition process of the vehicle sensor system or vehicle measurement and control system. The sample sequence after perturbation is represented as follows:

[0067] in, This represents the disturbance term. A zero-mean random variable whose magnitude does not exceed the magnitude of the corresponding sequence. .

[0068] The samples generated by each local model are divided into training, validation, and test sets according to a set ratio. The ratio of training, validation, and test sets is 8:1:1, but can be adjusted within the range of 6:2:2 to 9:0.5:0.5 depending on the sample size.

[0069] Step 4: Establish the recognition network corresponding to each local model For each local model in the local model library, an independent recognition network is built. Let the total number of local models be... Then the corresponding establishment There are three recognition networks, denoted as follows: .

[0070] No. A network receives historical sequence fragments. As input, output the history sequence belonging to the first... The probability of a local model is expressed as

[0071] in, This indicates that the historical sequence segment belongs to the first... The recognition probability of a local model.

[0072] like Figure 2 As shown, the independent recognition network adopts a long short-term memory network structure based on the attention mechanism, including an input layer, a first long short-term memory layer, a second long short-term memory layer, an attention layer, a linear mapping layer, and a probability output layer.

[0073] In long short-term memory units, the first The calculation process for each time step is as follows:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] in, Indicates the first Input vector at each time step Indicates a hidden state. Indicates the state of the memory cell. Indicates the output of the forget gate. Indicates the input gate output. Indicates the output gate output. , , , This represents the input weight matrix. , , , Represents the cyclic weight matrix. , , , Indicates the bias term. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0080] The attention layer weights the hidden states across all time steps to obtain the global representation vector. Let the output hidden state sequence of the second long short-term memory layer be...

[0081] The attention layer assigns weights based on the importance of each time step, and the weighted result is input into the linear mapping layer and the probability output layer to obtain the recognition probability. .

[0082] The independent recognition network is trained using the cross-entropy loss function, and the Adam optimizer is used for optimization. The learning rate range is [range missing]. The hidden layer dimension range is The network layer range is The training rounds range is .

[0083] Step 5: Establish the prediction model corresponding to each local model. For each local model, an independent prediction model is established, and the corresponding prediction model is denoted as .

[0084] No. Each prediction model receives historical sequence fragments. As input, directly output the future length as The predicted sequence is represented as

[0085] in, Indicates the first Candidate prediction results corresponding to each local model.

[0086] like Figure 3 As shown, the prediction model employs an improved Long Short-Term Memory (LSTM) network structure, specifically an encoder-decoder LSM network with a one-dimensional convolutional feature extraction layer. First, a one-dimensional convolution is performed on the input sequence to obtain local feature representations.

[0087] in, This represents a one-dimensional convolution operation. This represents the feature sequence after convolution.

[0088] The convolutional feature sequence is input into the encoder, which outputs the hidden state and memory unit state at the last time step. .Will As the initial state of the decoder, the decoder gradually outputs the predicted values ​​for each future time step, eventually forming a sequence of length [value missing]. The predicted sequence.

[0089] The prediction model training uses the mean squared error loss function.

[0090] in, Indicates the loss value. Represents the true value. This represents the predicted value.

[0091] The permissible range for training parameters of the prediction model is as follows: the number of convolutional kernels ranges from [value missing]. The kernel size range is The encoder and decoder hidden layer dimensions range from 100 to 1000. The training learning rate range is .

[0092] Step 6: Calculate the recognition probability and filter candidate local models Historical sequence fragments to be predicted Enter all of them respectively A recognition network is used to obtain a set of recognition probabilities.

[0093] Sort the recognition probabilities from largest to smallest to obtain the sorted probability sequence.

[0094] in, Indicates the maximum recognition probability. This indicates the second-highest recognition probability.

[0095] Before filtering by cumulative probability threshold A number of highly correlated local models were selected based on the following criteria: and

[0096] Where Pr represents the preset cumulative probability threshold, This indicates the number of candidate local models ultimately selected. The cumulative probability threshold Pr ranges from [value missing]. More .

[0097] By using multiple independent recognition networks to determine the degree of matching between the current historical motion state sequence segment to be predicted and each local model, and then selecting multiple candidate local models based on the ranking results and cumulative probability thresholds, instead of directly selecting a single maximum probability model, the prediction bias caused by pattern misjudgment is reduced.

[0098] Step 7: Calculate candidate prediction results For the first batch of samples selected in step six Each highly correlated local model is used to call its corresponding prediction model to obtain... The candidate prediction results are represented as

[0099] in, Indicates the first The prediction results corresponding to each candidate local model.

[0100] The prediction models corresponding to different local models generate prediction results of the vehicle's future motion state based on different assumptions of local evolution laws, thereby preserving future evolution information under multiple candidate modes and providing a basis for subsequent fusion.

[0101] Step 8: Weighted fusion of candidate prediction results The first batch selected in step six The recognition probabilities corresponding to each candidate local model are normalized to obtain the fusion weights.

[0102] in, Indicates the first The fusion weights corresponding to each candidate local model, and satisfying the following conditions:

[0103] Based on the normalized fusion weights, the candidate prediction results are weighted and summed to obtain the final prediction result.

[0104] in, This represents the predicted sequence of the final output.

[0105] This approach allows candidate models with higher recognition probabilities to have greater weight in the final prediction results, while retaining supplementary information from other highly relevant models, thereby improving the stability and accuracy of prediction results in complex time series scenarios.

[0106] In one specific embodiment, the local model library includes four types of local models: uniform velocity model, uniform acceleration model, uniform judder model, and periodic model. Random perturbations are added to each type of sample to generate training data, which is then divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0107] History window length Set the value to 50 and predict the step size. Take 10. When When the data is too small, historical information is insufficient, leading to a decrease in the accuracy of local pattern recognition; when... When the value is too large, redundant information increases, and computational complexity rises. (Take...) At the same time, it can better balance local feature extraction capability and computational efficiency. At the same time, it can meet the needs of short-term multi-step prediction while controlling error accumulation.

[0108] Figure 2 The independent recognition network shown adopts a structure that combines a two-layer long short-term memory network with an attention mechanism. The hidden dimension of the long short-term memory layer is 128, and the hidden dimension of the attention layer is 64. Figure 3 The prediction model shown employs an encoder-decoder long short-term memory network structure with a one-dimensional convolutional feature extraction layer. The hidden layer dimensions of both the encoder and decoder are 128. The independent recognition network uses cross-entropy loss as its loss function, while the prediction model uses mean squared error loss. The Adam optimizer is used for both optimization algorithms, with an initial learning rate of 0.001.

[0109] The cumulative probability threshold Pr is set to 0.85. When Pr is too low, the number of candidate models is too small, easily degenerating into a single model prediction; when Pr is too high, low-relevance models are introduced, affecting the fusion effect. At the same time, it can better balance the number of candidate models and prediction accuracy.

[0110] In this embodiment, the prediction results of the present invention are illustrated using the scenario corresponding to the uniform acceleration model in the local model library as an example. This time series is... , and The sequence consists of three state components, which together form the changes in the vehicle's motion state in the state space. Within the indicated prediction interval, the sequence exhibits continuous and smooth acceleration characteristics, and can be regarded as a local pattern sequence dominated by a uniform acceleration model.

[0111] Depend on Figure 4 As can be seen, the predicted sequence of a single model is relatively close to the actual sequence in the initial stage of prediction, but gradually deviates as time progresses. In contrast, the fused predicted sequence of this embodiment is closer to the actual sequence in both overall trend and local curvature. This embodiment demonstrates that this implementation can effectively improve prediction accuracy in uniform acceleration scenarios. Furthermore, this embodiment is not limited to uniform acceleration scenarios; for vehicle motion state time series corresponding to other local patterns in the local model library, prediction can also be achieved through local model identification, candidate model screening, and weighted fusion, obtaining good prediction results.

[0112] The embodiments of the present invention have been described above. These embodiments are used to help understand the technical solutions of the present invention and should not be construed as limiting the scope of the technical solutions of the present invention. Equivalent substitutions or conventional modifications made by those skilled in the art to the embodiments without departing from the technical concept of the present invention should all fall within the scope of the technical solutions of the present invention.

Claims

1. A vehicle motion state quantity prediction method based on local model identification and candidate model screening, characterized by, The method includes: S1. Obtain the historical motion state time series of the vehicle and preprocess it. Divide the preprocessed historical motion state time series into segments according to the sliding window method to obtain the historical motion state sequence segments to be predicted. The historical motion state time series is collected by the vehicle sensor system or the vehicle measurement and control system. The observed values ​​in the historical motion state time series are physical state quantities that reflect the motion state of the vehicle. S2. Construct a local model library, and generate training sample sets for training the corresponding independent recognition network and the corresponding prediction model based on each local model in the local model library; S3. Establish an independent recognition network for each local model in the local model library, and train the independent recognition network using the corresponding training sample set. The independent recognition network is used to output the recognition probability that the historical motion state sequence segment to be predicted belongs to the corresponding local model. S4. For each local model in the local model library, a corresponding prediction model is established, and the corresponding prediction model is trained using the corresponding training sample set. The corresponding prediction model is used to output the candidate prediction results under the corresponding local model. S5. Input the sequence fragments of the historical motion state variables to be predicted into all independent recognition networks to obtain the recognition probability set corresponding to each local model. S6. Sort the recognition probabilities in the recognition probability set in descending order, and accumulate the recognition probabilities in sequence according to the sorting results until the sum of the first m recognition probabilities is greater than or equal to the cumulative probability threshold, and the sum of the first m-1 recognition probabilities is less than the cumulative probability threshold. Then, determine the first m local models participating in the accumulation as candidate local models. S7. Only call the prediction model corresponding to each of the selected candidate local models to obtain multiple sets of candidate prediction results; S8. Normalize the recognition probability of each candidate local model selected to obtain the fusion weight corresponding to each candidate prediction result. S9. The multiple candidate prediction results are weighted and fused according to the fusion weight to obtain the prediction result of the vehicle's future motion state.

2. The method of claim 1, wherein, The physical state quantities mentioned in S1 include one or more of the following: vehicle position, vehicle speed, vehicle acceleration, and vehicle attitude angle.

3. The method of claim 1, wherein, The local model library mentioned in S2 includes a uniform speed model, a uniform acceleration model, a uniform judder model, and a periodic model. The uniform speed model, uniform acceleration model, uniform judder model, and periodic model are used to characterize the different evolution laws of the time series of vehicle motion state quantities within a local interval.

4. The method of claim 1, wherein, S2 The step of generating training sample sets for training corresponding independent recognition networks and corresponding prediction models based on each local model in the local model library includes: generating time series samples of vehicle motion state quantities of corresponding types based on each local model, adding a perturbation term to the generated time series samples of vehicle motion state quantities, and using the time series samples of vehicle motion state quantities with the perturbation term as the corresponding training sample set.

5. The method of claim 1, wherein, The independent recognition network described in S3 adopts a long short-term memory network structure based on an attention mechanism. The long short-term memory network structure based on an attention mechanism includes an input layer, a first long short-term memory layer, a second long short-term memory layer, an attention layer, a linear mapping layer, and a probability output layer.

6. The method of claim 1, wherein, The independent recognition network output recognition probability described in S3 includes: outputting a hidden state sequence through a second long short-term memory layer, weighting the hidden states at each time step in the hidden state sequence through an attention layer to obtain a global representation vector, and outputting the recognition probability of the predicted historical motion state sequence fragment belonging to the corresponding local model through a linear mapping layer and a probability output layer.

7. The method of claim 1, wherein, The corresponding prediction model described in S4 adopts an encoder-decoder long short-term memory network structure with a one-dimensional convolutional feature extraction layer. The one-dimensional convolutional feature extraction layer is used to perform one-dimensional convolution on the sequence fragment of the historical motion state quantity to be predicted to obtain local feature representation. The encoder is used to output the hidden state and memory unit state according to the local feature representation. The decoder is used to output the candidate prediction result under the corresponding local model according to the hidden state and memory unit state.

8. The method of claim 1, wherein, S7 The method of only calling the prediction model corresponding to each selected candidate local model includes: the prediction model corresponding to the local model that is not determined as a candidate local model does not generate candidate prediction results and does not participate in the fusion calculation of the prediction results of the future motion state of the vehicle.

9. The method of claim 1, wherein, S8 describes normalizing the recognition probabilities corresponding to each selected candidate local model, which includes generating fusion weights based on the proportional relationship between the recognition probabilities corresponding to each selected candidate local model, and weighting and fusing the multiple candidate prediction results based on the fusion weights to obtain the prediction result of the vehicle's future motion state.

10. A computer storage medium, characterized in that The computer storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images