A data-driven based evaluation method for vibration comfort of special vehicle seat

Abstract

The application discloses a kind of based on data driving's special vehicle seat vibration comfort evaluation method, it is related to seat comfort evaluation technical field, the 9 channel vibration information collected by sensor is input into multi-channel basic model to carry out first feature extraction and data fusion;In addition, before data fusion, channel priority convolution attention mechanism is added in model to extract feature information;Then the feature information after extraction and fusion is generated new network model by improvement residual contraction module to carry out seat vibration comfort evaluation, the evaluation model proposed in the application shows higher accuracy in evaluating seat vibration comfort degree.Aspects.The application uses deep learning to be applied to special vehicle vibration comfort evaluation, improves the accuracy of comfort evaluation, reduces the comfort evaluation error caused by human subjective factors, compared with traditional subjective evaluation method, improves the efficiency of special vehicle seat comfort evaluation.

Description

Technical Field

[0001] This invention belongs to the field of seat comfort assessment technology, specifically relating to a data-driven method for evaluating the vibration comfort of special vehicle seats. Background Technology

[0002] Special vehicles are vehicles used to perform special tasks in specific environments. They typically have special designs and functions, and may differ significantly from ordinary vehicles in terms of structure, power, and control systems to meet special work requirements. Special vehicles operate in harsh environments for extended periods, and the strong vibrations and impacts generated by the vehicles not only affect the work efficiency of personnel but also pose long-term potential health hazards. Therefore, evaluating the vibration comfort of special vehicle seats to guide vehicle seat design, thereby alleviating driver and passenger fatigue and improving efficiency, is crucial.

[0003] In seat comfort evaluation, subjective feelings are directly mapped to objective evaluation indicators. Establishing a relationship between subjective feelings and objective data through personnel may lead to information loss. Furthermore, the comfort assessment only focuses on the comfort of a single axis of the seat and does not consider information from other parts and directions of the seat. This cannot accurately express the nonlinear mapping between the original vibration data and the seat vibration comfort. In addition, the operating environment of special vehicles is complex and diverse. Comfort test evaluations based on ordinary civilian vehicles and trains cannot truly reflect the complex operating environment of special vehicles. In order to more realistically and efficiently evaluate the vibration comfort of special vehicle seats, a data-driven evaluation method for special vehicle seat vibration comfort is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the vibration comfort of special vehicle seats by providing a data-driven method for evaluating the vibration comfort of special vehicle seats.

[0005] This method analyzes the vibration data of special vehicle seats and then interprets and studies the vibration comfort of special vehicle seats, providing a novel and effective method for evaluating and testing the comfort of special vehicle seats.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0007] 1. A data-driven method for evaluating the vibration comfort of seats in special vehicles, characterized by comprising the following steps:

[0008] The first step is to collect a set of 9 axial vibration signals from 3 measuring points on the seat by using vibration acceleration sensors installed at 3 locations on the seat back, seat support surface and foot support surface.

[0009] The second step is to use the ISO2631 standard to calculate the total ride value of the nine axial vibration signals collected, which serves as a comfort score label for a set of vibration signals.

[0010] First, before calculating the total occupancy value, the root mean square value of the total weighted acceleration of the seats needs to be calculated:

[0011] ;

[0012] In the formula, , , These are the root mean square values ​​of the accelerations at the corresponding measurement points on the x, y, and z axes, respectively. , , The corresponding weighting factor;

[0013] Secondly, the total passenger load is further calculated as a comfort score label. The calculation method is as follows:

[0014] ;

[0015] In the formula, The root mean square of the total weighted acceleration at different measurement points of the signal;

[0016] The third step is to preprocess the vibration signals from the three measuring points of the seat, which cover a total of nine axes. First, 2048 sampling points are selected from the collected vibration signals using the sliding window method as one sample, with a 30% overlap rate. Then, this sample is divided into four sub-samples. Finally, the four sub-samples are combined to form a new sample, thereby expanding the data. Each expanded data set is labeled with a comfort score, and the data is shuffled. The training set and test set are divided in a 4:1 ratio.

[0017] The fourth step is to establish a seat vibration comfort assessment model. The assessment model is based on a multi-channel CNN model. The model input data is an expanded 9-axis vibration acceleration signal. The vibration signal of each axis passes through a convolutional layer, a pooling layer, a channel-first convolutional attention layer, and a fusion layer. After processing by the fusion layer, the feature is extracted through a deep residual shrinking network module, and finally the comfort prediction value is output.

[0018] The multi-channel CNN model is responsible for combining the multi-axis vibration information of the seat. The convolutional layer is responsible for performing convolution operations on the input information, and the pooling layer uses max pooling to reduce the dimensionality of the data from the upper layer.

[0019] The channel-priority convolutional attention layer is responsible for autonomously learning the sensitivity of vibration signal features to comfort, dynamically distributing the obtained weights on the channel and spatial dimensions, and performing the first feature extraction.

[0020] The extracted features are fused using a weighted fusion algorithm. The fusion process can be represented as follows:

[0021] ;

[0022] In the formula, n represents the number of input channels; k is the number of convolution kernels; This represents the input of the nth channel of the input layer; This represents the weight of the k-th convolutional kernel in convolutional layer l; It is a bias; the feature sequence it generates is denoted by F;

[0023] In the weighted fusion algorithm, the weights refer to the axis weighting coefficients for different measurement points on the seat given in the ISO2631 standard. The weight allocation for each channel is defined based on these axis weighting coefficients, and then fusion is performed. The result of the first feature fusion is as follows:

[0024] ;

[0025] In the formula, α, β, and γ represent the triaxial weighting coefficients for the same measuring point. , and This indicates the characteristics of triaxial vibration information at the same measuring point. For the feature fusion result of the same measurement point, i represents b, s, and f;

[0026] After completing the first feature fusion, a second feature fusion was performed at different measurement points. The fusion result is as follows:

[0027] ;

[0028] In the formula, F is the feature sequence after the fusion of 3 measurement points;

[0029] After the fused feature sequence is passed through a convolutional layer, redundant signals are eliminated using an improved deep residual shrinkage module. This improved deep residual shrinkage module is improved using a soft thresholding function as follows: ;

[0030] In the formula, x is the input, y is the output, and λ is the threshold;

[0031] The signal, after being processed by the improved deep residual shrinkage module, is subjected to convolution operation by the convolutional layer, max pooling by the pooling layer, and then a second feature extraction is performed. Finally, a fully connected layer is used for output.

[0032] The fifth step involves inputting the training set into the proposed model for training. The batch size is set to 32, the learning rate to 1e-3, and the number of training epochs to 200. The loss function is then applied. The trained seat vibration comfort evaluation model is obtained during network training.

[0033] The sixth step is to input the test set into the trained seat vibration comfort evaluation model to evaluate the seat vibration comfort.

[0034] In summary, the present invention has the following beneficial effects:

[0035] This invention provides a data-driven method for evaluating the vibration comfort of special vehicle seats, which effectively improves the efficiency of comfort assessment. Furthermore, by employing a channel-first convolutional attention mechanism and an improved residual contraction module to generate a new multi-channel seat comfort assessment model, the accuracy of comfort assessment can be improved. This invention provides a novel and effective method for the comfort assessment and detection of special vehicle seats. Attached Figure Description

[0036] Figure 1 This is a structural diagram of the seat comfort evaluation model in this invention.

[0037] Figure 2 This is the data augmentation method in this invention.

[0038] Figure 3 This is a diagram of the channel-first convolutional attention mechanism used in this invention.

[0039] Figure 4 This is a schematic diagram of multi-channel fusion in this invention.

[0040] Figure 5 This is a comparison chart of the soft threshold function before and after the improvement in this invention.

[0041] Figure 6 This is a comparison chart of seat comfort evaluation in this invention.

[0042] Figure 7 This is a graph of the loss function of the evaluation model in this invention.

[0043] Figure 8 This is a residual plot of the evaluation results of the evaluation model in this invention.

[0044] Figure 9 This is the regression fitting graph of the evaluation model in this invention. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings.

[0046] This embodiment describes a data-driven method for evaluating the vibration comfort of special vehicle seats. First, it collects and expands vibration data from three measuring points on the seat, labeling the data with comfort scores. Second, it inputs the 9-channel vibration data from the three measuring points and three axes into a model for initial feature extraction and fusion. Then, the fused data is transmitted to an improved deep residual shrinkage network to eliminate redundant information and complete a second feature extraction. Finally, it outputs the seat comfort evaluation score, such as... Figure 1 The diagram shown is a schematic of the evaluation model of the present invention.

[0047] (1) Data processing methods and the establishment of comfort assessment models

[0048] First, the total ride value is calculated using the ISO2631 standard for the nine axial vibration signals collected, serving as a comfort score label for a set of vibration signals. Before calculating the total ride value, the root mean square value of the total weighted acceleration of the seat needs to be calculated:

[0049] ;

[0050] In the formula, , , These are the root mean square values ​​of the accelerations at the corresponding measurement points on the x, y, and z axes, respectively. , , These are the weighting factors;

[0051] Secondly, the total passenger load is further calculated as a comfort score label. The calculation method is as follows:

[0052] ;

[0053] In the formula, The root mean square of the total weighted acceleration at different measurement points of the signal;

[0054] Data augmentation specifically involves selecting 2048 sampling points from the acquired vibration signals using a sliding window technique, with a 30% overlap rate. This sample is then divided into four sub-samples, which are subsequently recombined to form new samples, thus augmenting the data. Each augmented data set is labeled with a comfort score, and the data is shuffled, with the training and test sets divided in a 4:1 ratio. Figure 2 As shown.

[0055] The training set data is input into the evaluation model for feature extraction. Vibration signals from each channel are input into convolutional layers and pooling layers to complete the first feature extraction. To effectively extract seat comfort-sensitive information, an attention mechanism is considered for feature recognition. Channel-priority convolutional attention is an efficient attention method that supports the dynamic distribution of attention weights across channels and spatial dimensions. By employing multi-scale deep convolutional modules, spatial relationships can be effectively extracted while preserving prior channel information, such as... Figure 3 As shown, the channel-first convolutional attention module performs channel attention and spatial attention sequentially. Given an intermediate feature map T as input, the channel attention module (CA) first infers a channel attention map. Then M c Element-wise multiplication with the input feature T broadcasts the channel attention values ​​along the spatial dimension, resulting in refined features T with channel attention. c Spatial Attention Module (SA) processes T c Generate spatial attention map The final output features It is M s With T c The result is obtained by multiplication. The entire attention process can be summarized as:

[0056] ;

[0057] ;

[0058] Secondly, to accurately represent the nonlinear mapping between seat vibration data and comfort, and to achieve minimal information loss and highly accurate evaluation results, after completing the first layer of feature extraction, data fusion is performed on the vibration features extracted from the 9 channels. This fusion is mainly done in two stages: the first fusion is the fusion of vibration signal features from the same seat measuring point along three axes, and the second fusion is the fusion of vibration signal features from three measuring points. Figure 3 As shown, the overall calculation process of fusion can be expressed as follows:

[0059] ;

[0060] In the formula, n represents the number of input channels; k is the number of convolution kernels; This represents the input of the nth channel of the input layer; This represents the weight of the k-th convolutional kernel in convolutional layer l; The bias is used; the characteristic sequence it generates is denoted by F.

[0061] To address the weighting issue in data fusion, considering the axial weighting coefficients specified in ISO 2631 for seat vibration evaluation (as shown in Table 1), this invention proposes using the axial weighting coefficients specified in ISO 2631 to fuse the extracted features. First, the axial weighting coefficients for each measurement point are obtained from the ISO 2631 standard. Then, the weighting allocation for each channel is defined based on these axial weighting coefficients. The first feature fusion result is as follows:

[0062] ;

[0063] In the formula, α, β, and γ represent the triaxial weighting coefficients for the same measuring point. , and This indicates the characteristics of triaxial vibration information at the same measuring point. The result is the feature fusion of the same measurement point, where i represents b, s, and f.

[0064] After the first feature fusion of the same measurement point according to the ISO2631 standard, a second feature fusion of different measurement points is performed. The fusion result is as follows:

[0065] ;

[0066] In the formula, F is the feature sequence after the fusion of 3 measurement points.

[0067] Table 1. Axis weighting factors recommended by ISO 2631

[0068]

[0069] The fused features undergo a second feature extraction and are output through a fully connected layer to obtain the comfort assessment result. Considering the redundancy that arises when fusing multiple features or information sources of vibration signals, this information does not contribute to improving the performance of the evaluation model and may even lead to overfitting or reduced generalization ability. Deep residual contraction networks, by combining soft thresholding and attention mechanisms, can effectively address the problem of redundant information after fusion. However, due to the constant bias between the input and output of the residual contraction unit, effective feature information is filtered out, resulting in reduced accuracy of the evaluation model. To address the shortcomings of the residual contraction unit, the semi-soft thresholding function is improved, thereby improving the deep residual contraction network. The improved semi-soft thresholding function is as follows:

[0070] ;

[0071] In the formula, x is the input, y is the output, and λ is the threshold.

[0072] when When the improved soft thresholding function converts features close to zero to 0, it is the same as the soft thresholding function, when... When the input and output exhibit a non-linear relationship, as x increases, the deviation between the input x and the output y gradually decreases, eliminating the constant deviation problem inherent in the soft threshold function and preserving the effective features to the greatest extent. Specific examples of the soft threshold function before and after the improvement are shown below. Figure 5 As shown.

[0073] (2) Vibration test data acquisition

[0074] To verify the effectiveness of the proposed evaluation method, experimental data was collected. Vibration acceleration signals were measured at three points (driver's seat support surface, seat back, and footwell) along nine axes while the vehicle was in motion. Test speeds were 15 km / h, 20 km / h, 25 km / h, 30 km / h, 35 km / h, and 40 km / h, with a test duration of at least 100 seconds. To obtain diverse excitations, both concrete and unpaved roads were selected.

[0075] (3) Vibration test data processing

[0076] The total seat vibration value given by the ISO2631 standard is used to score each vibration signal to reduce the error caused by human subjective evaluation. The data is expanded by data augmentation method and the expanded data is labeled with comfort value. Then the data is shuffled and divided into training set and test set in a 4:1 ratio.

[0077] (4) Analysis of experimental results

[0078] There are many hyperparameters during network training, and the settings of these hyperparameters affect the final result. This invention, based on the basic parameters, utilizes a validation set... The optimal hyperparameters of the network were found through a combination of numerical and experimental methods. The specific hyperparameter settings are shown in Table 2, where `filters` represents the number of convolutional kernels, `Kernel_size` represents the convolutional window size, `Padding` represents the padding method, `Pool_size` represents the pooling layer size, `Strides` represents the stride, `Out` represents the number of output channels, and `Units` represents the number of neurons. The training epochs were 200, the batch size was set to 32, the initial learning rate was 1e-3, and the loss function was MSE. The dataset was input into the evaluation model for training and testing. After 100 training iterations, the model basically met the training accuracy requirements, and the loss values ​​on both the training and validation sets approached zero, indicating that the model had essentially reached its optimal state.

[0079] Table 2 Basic parameters of the evaluation model

[0080]

[0081] Table 3 shows the results of 10 trials of the model. Figure 6To compare the actual and predicted scores of seat vibration comfort, Table 3 shows the determination coefficient R of the proposed model. 2 The average values ​​of the root mean square error (RMSE) and the mean square error (MAE) were 0.979, 0.049, and 0.035, respectively, all indicating small prediction errors. Figure 6 As can be seen, most actual values ​​basically coincide with the evaluated predicted values; the training loss function curve is as follows: Figure 7 The proposed model converges quickly during training, with the loss value approaching zero; the prediction error histogram on the validation dataset is shown below. Figure 8 As shown, the prediction error of almost all samples is between -0.1 and 0.1, which is a relatively low error, which can intuitively prove the high fidelity of the proposed model. Figure 9 As shown in the regression analysis curve, the predicted values ​​and actual values ​​are in high agreement. Most data points are closely clustered on the diagonal, showing a very strong linear relationship. The accuracy of the evaluation model is high, indicating that the model proposed in this invention can effectively evaluate seat vibration comfort.

[0082] Table 3 Evaluation Indicators for Seat Vibration Comfort

[0083]

[0084] (4) Comfort assessment and verification

[0085] Ten sets of seat vibration signals were randomly selected and input into the model proposed in this invention. The results are shown in Table 4. It can be seen that the evaluation results obtained by the comfort evaluation model proposed in this invention are close to the actual evaluation results obtained by ISO2631. The average error of 10 verifications is 2.02%, indicating that the model proposed in this invention has high evaluation accuracy. The evaluation model proposed in this invention can effectively extract the sensitive features affecting comfort from vibration data, and thus estimate the degree of subjective feeling.

[0086] Table 4. Comfort scores of 10 sample groups

[0087] .

Claims

1. A data-driven method for evaluating the vibration comfort of seats in special vehicles, characterized in that, Includes the following steps: The first step is to collect a set of 9 axial vibration signals from 3 measuring points on the seat by using vibration acceleration sensors installed at 3 locations on the seat back, seat support surface and foot support surface. The second step is to use the ISO2631 standard to calculate the total ride value of the nine axial vibration signals collected, which serves as a comfort score label for a set of vibration signals. First, before calculating the total occupancy value, the root mean square value of the total weighted acceleration of the seats needs to be calculated: ； In the formula, , , These are the root mean square values ​​of the accelerations at the corresponding measurement points on the x, y, and z axes, respectively. , , These are the weighting factors; Secondly, the total passenger load is further calculated as a comfort score label. The calculation method is as follows: ； In the formula, The root mean square of the total weighted acceleration at different measurement points of the signal; The third step is to preprocess the vibration signals from the three measuring points and nine axes of the seat. First, 2048 sampling points are selected from the collected vibration signals using the sliding window method as one sample. The overlap rate of the sliding window samples is 30%. Then, the sample is divided into four sub-samples on an average basis. Finally, the four sub-samples are combined to form a new sample to expand the data. Each group of expanded data is labeled with a comfort score, and the data is shuffled. The training set and test set are divided in a 4:1 ratio. The fourth step is to establish a seat vibration comfort assessment model. The assessment model is based on a multi-channel CNN model. The model input data is an expanded 9-axis vibration acceleration signal. The vibration signal of each axis is passed through a convolutional layer, a pooling layer, a channel-first convolutional attention layer, and a fusion layer. After processing by the fusion layer, the feature is extracted through a deep residual shrinkage network module. Finally, a fully connected layer is used to output the comfort prediction value. The multi-channel CNN model is responsible for combining the multi-axis vibration information of the seat. The convolutional layer is responsible for performing convolution operations on the input information, and the pooling layer uses max pooling to reduce the dimensionality of the data from the upper layer. The channel-priority convolutional attention layer is responsible for autonomously learning the sensitivity of vibration signal features to comfort, dynamically distributing the obtained weights on the channel and spatial dimensions, and performing the first feature extraction. The extracted features are fused using a weighted fusion algorithm. The fusion process is represented as follows: ； In the formula, n represents the number of input channels; k is the number of convolution kernels; This represents the input of the nth channel of the input layer; This represents the weight of the k-th convolutional kernel in convolutional layer l; It is a bias; the feature sequence it generates is denoted by F; In the weighted fusion algorithm, the weights are based on the axis weighting coefficients for different measurement points on the seat given in the ISO2631 standard. The weight allocation for each channel is defined according to these axis weighting coefficients, and then fusion is performed. The result of the first feature fusion is as follows: ； In the formula, α, β, and γ represent the triaxial weighting coefficients for the same measuring point. , and This indicates the characteristics of triaxial vibration information at the same measuring point. For the feature fusion result of the same measurement point, i represents b, s, and f; After completing the first feature fusion, a second feature fusion was performed at different measurement points. The fusion result is as follows: ； In the formula, F is the feature sequence after the fusion of 3 measurement points; After the fused feature sequence is passed through a convolutional layer, redundant signals are eliminated using an improved deep residual shrinkage module. This improved deep residual shrinkage module is improved using a soft thresholding function as follows: ； In the formula, x is the input, y is the output, and λ is the threshold; The fifth step involves inputting the training set into the proposed model for training. The batch size is set to 32, the learning rate to 1e-3, and the number of training epochs to 200. The loss function is then applied. Used in training the network to obtain a well-trained seat vibration comfort evaluation model; The sixth step is to input the test set into the trained seat vibration comfort evaluation model to evaluate the seat vibration comfort.

Images