Nonlinear impact load time sequence and position identification method based on deep learning

Abstract

The invention relates to the technical field of aircraft load identification, and discloses a non-linear impact load time sequence and position identification method based on deep learning, and the method comprises the steps: obtaining acceleration vibration signal data; constructing a neural network model comprising a convolutional encoder, a self-attention mechanism module and a convolutional decoder for identifying the size information of the impact load; meanwhile, a neural network model comprising an input layer module, a hidden layer module and a full-connection layer module is constructed and used for recognizing the position information of the impact load; the method has the advantages of being remarkable in recognition effect and high in accuracy when being used for complex and nonlinear impact load size reconstruction and position positioning.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft load identification, and particularly to a method for identifying the time sequence and position of non-linear impact loads based on deep learning. Background Art

[0002] In the aviation field, the safety performance of aircraft is a crucial consideration. Impact loads, as common external excitations during the take-off, flight, and landing of aircraft, have a significant impact on the safety of the aircraft structure. Directly measuring the loads acting on the aircraft structure is usually very difficult because it requires installing a large number of sensors on the structure, which is often infeasible or too costly in actual operation. Therefore, studying load identification technology, that is, inferring the magnitude and position of the loads acting on the aircraft structure through the response of the aircraft structure, namely the non-linear structure, has important reference significance for the safety, damage assessment, and strength design of aircraft. Load identification is an inverse problem in structural dynamics, and its goal is to determine the input excitation through the output response of the structure. In the field of aircraft structural health monitoring, existing load identification methods mainly include time-domain methods and frequency-domain methods. The frequency-domain method identifies loads by analyzing the frequency characteristics of the structural response and is suitable for the identification of periodic and steady-state loads. The time-domain method directly analyzes the structural response in the time domain and is suitable for the identification of non-stationary and transient loads. The time-domain method and the frequency-domain method have extensive applications in the field of load identification, but they may face difficulties such as large computational amounts, sensitivity to noise, and difficulty in capturing non-linear characteristics when dealing with complex dynamic systems. Summary of the Invention

[0003] Aiming at the above deficiencies in the prior art, the present invention provides a method for identifying the time sequence and position of non-linear impact loads based on deep learning, which is used to solve the problems existing in the existing load identification methods, such as large computational amounts, sensitivity to noise, and difficulty in capturing non-linear characteristics when dealing with complex dynamic systems.

[0004] In order to achieve the above invention purpose, the technical solution adopted by the present invention is as follows:

[0005] A method for identifying the time sequence and position of non-linear impact loads based on deep learning, comprising the following steps:

[0006] Obtain acceleration vibration signal data;

[0007] Construct a neural network model based on the self-attention mechanism, encoder, and decoder; the neural network model based on the self-attention mechanism, encoder, and decoder includes a convolutional encoder, a self-attention mechanism module, and a convolutional decoder;

[0008] Use the convolutional encoder to extract features from the acceleration vibration signal data to generate a first output feature;

[0009] After aggregating the spatial information of the first output feature using the self-attention mechanism module, a channel attention map is generated. By using the channel attention map as an important factor for each channel, the first output feature is rescaled to generate the second output feature;

[0010] Using a convolutional decoder to reconstruct the high-dimensional features of the second output feature to generate the magnitude information of the impact load;

[0011] Construct a deep convolutional and recursive neural network model based on the self-attention mechanism; the deep convolutional and recursive neural network model based on the self-attention mechanism includes an input layer module, a hidden layer module, and a fully connected layer module;

[0012] After using the input layer module to extract the spatial features of the acceleration vibration signal data, the temporal features of the spatial features are extracted;

[0013] Using the hidden layer module to construct a non-linear relationship for the temporal features to generate the first non-linear relationship feature;

[0014] Using the fully connected layer module to extract the coordinates of the first non-linear relationship feature to generate the position information of the impact load.

[0015] The present invention has the following beneficial effects:

[0016] A method for identifying the time series and position of non-linear impact loads based on deep learning proposed by the present invention combines the strong feature extraction ability of convolutional neural networks, the feature selection ability of self-attention mechanisms, and the time series processing ability of gated recurrent units, and shows significant advantages in dealing with impact load reconstruction tasks. The convolutional neural network, self-attention mechanism, and gated recurrent unit are alternately arranged to construct a neural network model based on the self-attention mechanism, encoder, and decoder to reconstruct the time history of the impact load; a deep convolutional and recursive neural network model based on the self-attention mechanism is constructed to achieve impact load positioning. This method has the advantages of significant recognition effect and high accuracy when facing complex and non-linear impact load magnitude reconstruction and position positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of a method for identifying the time series and position of non-linear impact loads based on deep learning proposed by the present invention;

[0018] Figure 2 It is a schematic diagram of the position layout of each sensor in the embodiment;

[0019] Figure 3 It is a schematic diagram of the neural network model structure based on the self-attention mechanism, encoder, and decoder in the embodiment;

[0020] Figure 4 Schematic diagram of the neural network model structure of deep convolution and recursion based on the self-attention mechanism in the embodiment;

[0021] Figure 5 Schematic diagram of the structure of the gated recurrent unit in the embodiment. Specific implementation manners

[0022] The following describes the specific implementation manners of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation manners. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0023] As Figure 1 shown, a method for identifying the time series and position of non-linear impact loads based on deep learning includes the following steps S1 - S9:

[0024] S1. Obtain acceleration vibration signal data.

[0025] In this embodiment, in order to identify the magnitude and position information of the impact load in the acceleration vibration signal data, by dividing the training load points and test load points on the aircraft tail wing, arranging acceleration vibration sensors to collect acceleration vibration signal data, arranging force sensors to collect excitation signal data, and at the same time applying excitation to the force sensors to obtain the sampling data of all training load points and all test load points; so as to use the sampling data of all training load points and all test load points to train the neural network model based on the self-attention mechanism, encoder and decoder and the neural network model of deep convolution and recursion based on the self-attention mechanism, thereby obtaining the trained models, which are respectively used to identify the magnitude information and position information of the impact load.

[0026] Among them, the sampling data of all training load points includes acceleration vibration signal data and excitation signal data; the sampling data of all test load points only includes acceleration vibration signal data; and the acquisition process of the sampling data of all training load points and all test load points is as Figure 2 shown, specifically:

[0027] (1) Construct a signal acquisition and analysis system: Uniformly arrange 4 acceleration vibration sensors in the x, y, and z directions on the aircraft tail wing model to obtain the vibration response (acceleration vibration signal data) after excitation. The force sensor can be movably installed at each acquisition point to record the excitation signal; at the same time, connect the sensors to the data acquisition instrument and computer analysis software to construct a complete signal acquisition and analysis system. For the specific arrangement of the sensors, refer to Figure 2 ,Figure 2 Perform mesh division on the aircraft tail structure, design a complete mesh of 40 mm × 40 mm to obtain more impact points; among them, #1, #2, #3, and #4 are the layout positions of the three-axis acceleration vibration sensors;

[0028] (2) Set the system continuous sampling time to 5 s and the sampling frequency to 12800 Hz. Use an impact hammer to apply excitation to the force sensor installed at the preset acquisition point (training load point or test load point). Control the excitation application magnitude within 5 - 55 N, and collect the dynamic responses (acceleration vibration signal data) of the vibration signals of 12 channels of 4 three-axis acceleration vibration sensors in the x, y, and z directions under the excitation;

[0029] Refer to Figure 2 , divide all acquisition points into training load points and test load points, and repeat the data acquisition process in (1) at each point in turn. Among them, 50 data are collected at each training load point, and 10 data are collected at each test load point to obtain all sample data. 6200 training samples are collected from 124 training load points, of which 4960 training samples are used as the training data set for training, and 1240 training samples are used as the validation data set for validation. Here, each training sample data includes the excitation signal data and the vibration response data (acceleration vibration signal data) after excitation;

[0030] In addition, during the mesh division process, take the intersection points of all the grid lines of the divided meshes as the training load points; if there are incomplete meshes among all the divided meshes, the points where the incomplete meshes intersect with the aircraft tail are not used as training load points. At the same time, among all the divided meshes, exclude the four meshes where the acceleration vibration sensors are located and the one mesh to the left of these four meshes to obtain the remaining meshes, and take the centers of the remaining meshes as the test load points; if there are incomplete meshes among the remaining meshes, the centers of the incomplete meshes are not used as test load points; by sequentially arranging a force sensor at each training load point and applying excitation to the force sensor, obtain the acceleration vibration signal data collected by each acceleration vibration sensor and the excitation signal data collected by the force sensor to obtain the sampling data of all training load points; by sequentially arranging a force sensor at each test load point and applying excitation to the force sensor, obtain the acceleration vibration signal data collected by each acceleration vibration sensor to obtain the sampling data of all test load points.

[0031] Then, the sampled data of all training load points are normalized and proportionally divided into a training data set and a test data set. At the same time, the sampled data of all test load point data are normalized and used as the test data set. Among them, the labels of the training data set and the test data set are excitation signal data. Here, the purpose of dividing the training data set and the test data set is to use the training data set and the test data set to train a neural network model based on the self-attention mechanism, encoder and decoder, and a neural network model based on the deep convolution and recursion of the self-attention mechanism, so as to obtain the trained models for respectively identifying the magnitude and position information of the impact load. The training process is the same as the subsequent identification process. The difference is that in the training process of each model, first, the training data set is input for several model trainings of each model. After training, the test data set is input into these several models to screen the optimal model and use it as the final model for identifying the magnitude or position information of the impact load. Among them, the training parameters of each model are set as follows: the number of training epochs is set to 100 times, the batch-size is set to 128, the learning rate is set to 0.05, and the optimal weight model is saved after training. This model will be used for the effect test of impact load magnitude and position identification. Specifically:

[0032] First, obtain the prediction results of the impact load magnitude and the impact load position.

[0033] Then, calculate the peak error, main impact error, full impact error, main impact root mean square error and full impact root mean square error between the prediction result of the impact load magnitude and the true value, that is:

[0034]

[0035] Among them, PE represents the peak error, N represents the number of samples in the test data set, represents the true value of the impact load magnitude of the i-th sample, represents the prediction result of the impact load magnitude of the i-th sample, MIE represents the main impact error, T1 represents the start time of the main impact, T2 represents the end time of the main impact, represents the true value of the impact load magnitude of the i-th sample at the t-th moment, represents the prediction result of the impact load magnitude of the i-th sample at the t-th moment, TIE represents the full impact error, T3 represents the secondary impact, MI-RMSE represents the main impact root mean square error, and TI-RMSE represents the full impact root mean square error;

[0036] And judge whether each error is less than the corresponding error threshold. If so, the prediction result of the impact load magnitude is accurate. Otherwise, repeat the training of the neural network model based on the self-attention mechanism, encoder and decoder.

[0037] The principle is as follows: The selected peak error is calculated only using the maximum peak of the true value and the maximum peak of the predicted value. This method reflects the accuracy of the model's peak prediction by focusing on the extreme value difference of the impact load. The selected main impact error not only includes the maximum peaks of the true value and the predicted value, but also includes several key sampling points near the peak in the error calculation range. This method can more comprehensively evaluate the reconstruction accuracy of the model in the main impact area. The selected full impact error compares all sampling points covering the entire impact process. This evaluation method can more comprehensively show the overall performance of the model in time series reconstruction and reflect its performance in time series details. The selected root mean square error amplifies larger deviations by calculating the square of the error, so it is more sensitive to outliers and can well reflect the overall level of the model's prediction deviation. The selected main impact root mean square error can more centrally evaluate the reconstruction performance of the model in the main impact area by including the maximum peaks of the true value and the predicted value and several key sampling points around them in the error calculation range. The selected full impact root mean square error covers all sampling points of the entire impact process for error calculation, thus more comprehensively showing the overall performance of the model in time series reconstruction.

[0038] Specifically, compare the peak error between the predicted result of the impact load magnitude and the true value. If the peak error is less than 10%, it meets the threshold requirement. Similarly, if the main impact error is less than 10%, it meets the threshold requirement. Similarly, if the full impact error is less than 10%, it meets the threshold requirement. If the main impact root mean square error is less than 20%, it meets the threshold requirement. If the full impact root mean square error is less than 20%, it meets the threshold requirement. Therefore, when all errors are less than their corresponding thresholds, the model prediction result is accurate.

[0039] At the same time, calculate the mean absolute error and misjudgment rate between the predicted result of the impact load position and the true value, that is:

[0040]

[0041] Among them, MAE represents the mean absolute error. respectively represent the x-axis coordinate and y-axis coordinate of the predicted result of the impact load position of the i-th sample, x i 、y i respectively represent the x-axis coordinate and y-axis coordinate of the true value of the impact load position of the i-th sample, MR represents the misjudgment rate, and M represents the number of misjudged samples.

[0042] And judge whether the mean error is less than the mean error threshold and whether the misjudgment rate is less than the misjudgment rate threshold. If so, the predicted result of the impact load position is accurate; otherwise, repeat the training of the deep convolutional and recursive neural network model based on the self-attention mechanism.

[0043] Specifically, by comparing the mean absolute error between the predicted result and the true value of the impact load position, if the mean absolute error is less than 10%, the threshold requirement is met; similarly, if the misjudgment rate is less than 10%, the threshold requirement is met; therefore, when all errors are less than their corresponding thresholds, the model prediction result is accurate, and the model can be used to identify the position information of the impact load subsequently. Otherwise, the model is retrained.

[0044] S2. Construct a neural network model based on the self-attention mechanism, encoder, and decoder; the neural network model based on the self-attention mechanism, encoder, and decoder includes a convolutional encoder, a self-attention mechanism module, and a convolutional decoder.

[0045] In this embodiment, a neural network model (ESD-CGRU) based on the self-attention mechanism, encoder, and decoder is constructed to reconstruct the time history of the impact load; among them, the reconstruction of the time history of the impact load is a typical sequence-to-sequence task. The ESD-CGRU model combines the strong feature extraction ability of the convolutional network, the feature selection ability of the self-attention mechanism, and the time series processing ability of the gated recurrent unit (GRU), and shows significant advantages in dealing with the task of reconstructing the time history of the impact load; the model includes a convolutional encoder module, a self-attention mechanism module, and a decoder module. Specifically, the structure of the ESD-CGRU model is as Figure 3 shown, including a convolutional encoder, a self-attention mechanism module, and a convolutional decoder; the convolutional encoder includes a first one-dimensional convolutional layer, a first gated recurrent unit, a second one-dimensional convolutional layer, and a bidirectional gated recurrent unit; the convolutional decoder includes a first one-dimensional transposed convolutional layer, a second gated recurrent unit, and a second one-dimensional transposed convolutional layer.

[0046] Among them, the convolutional encoder, the self-attention mechanism module, and the convolutional decoder are connected in sequence, the first one-dimensional convolutional layer, the first gated recurrent unit, the second one-dimensional convolutional layer, and the bidirectional gated recurrent unit are connected in sequence, and the first one-dimensional transposed convolutional layer, the second gated recurrent unit, and the second one-dimensional transposed convolutional layer are connected in sequence.

[0047] In addition, both the first gated recurrent unit and the second gated recurrent unit are gated recurrent units, and their structures are as Figure 5 shown, including a reset gate, an update gate, and a candidate hidden state.

[0048] S3. Use the convolutional encoder to extract features from the acceleration vibration signal data to generate a first output feature, including:

[0049] Input the acceleration vibration signal data into the first one-dimensional convolutional layer of the convolutional encoder for initial feature extraction, then input it into the first gated recurrent unit for temporal feature extraction of the initial features, then input it into the second one-dimensional convolutional layer for deep feature extraction, and then input it into the bidirectional gated recurrent unit for forward and backward temporal feature processing to generate the first output feature of the convolutional encoder.

[0050] Specifically, the formula for the first one-dimensional convolutional layer to perform initial feature extraction is:

[0051]

[0052] where l represents the length of the input signal, y j represents the initial feature of the j-th segment of the input signal, represents the output value of the j-th segment of the input signal on the convolutional kernel, *, [·] respectively represent the one-dimensional convolution operation and the concatenation operation, tanh represents the hyperbolic tangent activation function, n represents the total number of convolutional kernels, k ic represents the i-th convolutional kernel, s represents the size of the convolutional kernel, m represents the stride of the convolutional kernel, x m(j-1):s+m(j-1) represents the j-th segment of the input signal, b c represents the bias vector of the convolutional kernel.

[0053] Specifically, the formula for the first gated recurrent unit to perform temporal feature extraction of the initial features is:

[0054]

[0055] where h t represents the final hidden state of the first gated recurrent unit at the t-th moment, z t represents the output of the update gate at the t-th moment, h t-1 represents the final hidden state of the first gated recurrent unit at the (t - 1)-th moment, r t represents the output of the reset gate at the t-th moment, represents the candidate hidden state at the t-th moment, σ represents the sigmoid activation function, x t represents the input vector at the t-th moment, that is, the initial feature input by the first one-dimensional convolutional layer, W z 、W r 、W h respectively represent the input weight matrices for calculating the update gate, reset gate, and candidate hidden state, U z 、U r 、U h respectively represent the hidden state weight matrices for calculating the update gate, reset gate, and candidate hidden state, b z 、b r 、b hrespectively represent the bias vectors used to calculate the update gate, reset gate, and candidate hidden state.

[0056] In this embodiment, in the convolutional encoder module, first, the initial features are extracted through the first one-dimensional convolutional layer, then the preliminary extracted temporal features are processed by the gated recurrent unit, and then the features are further extracted through the second one-dimensional convolutional layer, and the forward and backward sequence dependencies are simultaneously processed by the bidirectional gated recurrent unit (BiGRU); where the one-dimensional convolution is the matrix multiplication operation of the convolution kernel (k ic ) and the input signal (x in ), where the input signal is divided into many segments; for the j-th segment input (x m(j-1):s+m(j-1) ), the convolutional output can be calculated according to the following formula: Therefore, the final output of the first one-dimensional convolutional layer is: The gated recurrent unit (GRU) is a variant of the RNN that can effectively solve the problem of gradient disappearance in general RNNs; because the GRU has a special gating structure, it can capture the forward and backward dependencies of sequence data and is very suitable for time series processing and prediction. The formula for extracting features of the GRU unit structure is where each W, U, and b are trainable parameters, and a GRU layer shares these parameters; the bidirectional GRU (BiGRU) layer is indispensable in the shock time history reconstruction. It can calculate data in both the forward and backward directions at the same time, enabling the model to capture the global dependencies of sequence data; the GRU and BiGRU layers are evenly distributed in the convolutional-encoder-self-attention mechanism-decoder structure, which means that the inputs of these RNN layers have lower dimensions at time steps, so the relationship between low-dimensional feature representations and output payloads can be learned at a lower computational cost; in addition, to avoid overfitting problems, the Dropout strategy is used in the RNN layer, and the parameter is set to 0.1, which means that during training, 10% of the neurons in the GRU unit are randomly discarded.

[0057] S4. After aggregating the spatial information of the first output feature using the self-attention mechanism module, a channel attention map is generated. By using the channel attention map as an important factor for each channel, the first output feature is rescaled to generate a second output feature, including:

[0058] The first output feature is input into the self-attention mechanism module. First, the global average pooling and global max pooling operations are used to aggregate the spatial information of the first output feature to generate two different first spatial context descriptors and second spatial context descriptors. At the same time, the first spatial context descriptor and the second spatial context descriptor are input into a shared network composed of a multi-layer perceptron and a hidden layer to generate a channel attention map, that is:

[0059] M = σ(W1(W0(Favg )) + W1(W0(F max )))

[0060] Among them, M represents the channel attention map, σ represents the sigmoid activation function of the hidden layer, W0 and W1 represent the weights of the first fully connected layer and the second fully connected layer respectively, and the first fully connected layer and the second fully connected layer form a multi-layer perceptron, F avg 、F max respectively represent the first spatial context descriptor and the second spatial context descriptor generated by global average pooling and global maximum pooling operations; among them, the multi-layer perceptron is composed of the first fully connected layer and the second fully connected layer.

[0061] Taking the channel attention map as an important factor for each channel, rescale the first output feature of the convolutional encoder to generate the second output feature of the self-attention mechanism module, that is:

[0062]

[0063] where F ′ represents the second output feature, F represents the first output feature of the convolutional encoder, represents the channel multiplication operation that broadcasts the channel attention value along the spatial dimension.

[0064] In this embodiment, a self-attention mechanism module is used to enhance informative features and suppress redundant features; first, by using global average pooling (GAP) and global maximum pooling (GMP) operations to aggregate the spatial information of the feature map, two different spatial context descriptors are generated; then, both of these descriptors are forwarded to a shared network to generate a channel attention map C represents the channel; the shared network consists of a multi-layer perceptron and a hidden layer; in order to reduce the parameter overhead, the hidden activation size is set to After applying the shared network to each descriptor, the output feature vectors are merged using element-wise summation; the calculation formula for the channel attention map M is: M = σ(W1(W0(F avg )) + W1(W0(F max ))), where and respectively represent the weights of the first fully connected layer and the second fully connected layer, and and respectively represent the first spatial context descriptor and the second spatial context descriptor generated by global average pooling and global maximum pooling operations; then, the channel attention map is regarded as an important factor for each channel, and the input is rescaled, that is: and respectively represent the input and refined output of the self-attention module, Represents channel multiplication that broadcasts channel attention values along the spatial dimension.

[0065] S5. Use a convolutional decoder to reconstruct high-dimensional features from the second output feature, generating the magnitude information of the impact load, including:

[0066] Input the second output feature into the first one-dimensional transposed convolution layer for high-dimensional feature reconstruction, then input it into the second gated recurrent unit for intermediate feature extraction of the reconstruction, and finally input it into the second one-dimensional transposed convolution layer to complete the high-dimensional reconstruction and generate the magnitude information of the impact load.

[0067] In this embodiment, the decoder module is responsible for mapping low-dimensional features back to the high-dimensional space. Specifically: First, start reconstructing the high-dimensional features of the signal through the first one-dimensional transposed convolution layer, then process the reconstructed intermediate features by the second gated recurrent unit, and finally complete the high-dimensional reconstruction of the signal through the second one-dimensional transposed convolution layer; the decoder module is responsible for mapping low-dimensional features back to the high-dimensional space; where the one-dimensional transposed convolution layer is similar to the one-dimensional convolution layer, but can be used for upsampling features after training to reconstruct the features of the hidden layer; when the stride is 1 / 2, the output length of the feature vector will double; the tanh activation function is introduced in the convolution and transposed convolution layers to enhance the non-linear representation ability; in the encoder module, the stride of the convolution is set to 2 to obtain the low-dimensional feature representation of the input signal; in the decoder module, the stride of the transposed convolution is also set to 2 to map the features to the high-dimensional space.

[0068] S6. Construct a deep convolutional and recursive neural network model based on the self-attention mechanism; the deep convolutional and recursive neural network model based on the self-attention mechanism includes an input layer module, a hidden layer module, and a fully connected layer module.

[0069] In this embodiment, construct a deep convolutional and recursive neural network model (SC-BiGRU) based on the self-attention mechanism for impact load localization; where the position localization of the impact load is a sequence-to-point task, mainly extracting the key features in the input acceleration sequence and at the same time needing to learn the spatial and temporal relationships of the acceleration signal. The structure of the SC-BiGRU model is as Figure 4 shown, including an input layer module, a hidden layer module, and a fully connected layer module; the input layer module includes a first one-dimensional convolutional layer, a first bidirectional gated recurrent unit, a self-attention mechanism module, a second one-dimensional convolutional layer, and a second bidirectional gated recurrent unit; the fully connected layer module (FC) includes a first fully connected layer and a second fully connected layer.

[0070] S7. After using the input layer module to extract the spatial features from the acceleration vibration signal data, perform the temporal feature extraction of the spatial features, including:

[0071] Input the acceleration vibration signal data into the first one-dimensional convolutional layer of the input layer module for spatial feature extraction, and use max pooling operation to compress the spatial features to generate a compressed spatial feature sequence. Then input it into the first bidirectional gated recurrent unit for forward and backward temporal feature extraction of the compressed spatial feature sequence. Then input it into the self-attention mechanism module for enhancing the impact signal information features and suppressing redundant features including noise and irrelevant frequency components. Then input it into the second one-dimensional convolutional layer for spatial feature extraction again and use max pooling operation to compress the spatially extracted features again. Then input it into the second bidirectional gated recurrent unit for forward and backward temporal feature extraction of the spatially compressed features again to generate the temporal features of the spatial features.

[0072] In this embodiment, the first one-dimensional convolutional layer is used to extract spatial features, and the feature sequence is further compressed through max pooling operation to enhance spatial features and mitigate overfitting. Then, the BiGRU learns the temporal relationships of these spatial features. The BiGRU can process forward and backward data simultaneously to capture the global dependencies of sequence data. Then, through the self-attention mechanism module, the information features are enhanced and redundant features are suppressed. Subsequently, the previous operations are repeated through the second one-dimensional convolutional layer and max pooling layer to further extract and compress features, and the temporal relationships are captured again through BiGRU.

[0073] S8. Use the hidden layer module to construct a non-linear relationship for the temporal features to generate the first non-linear relationship feature, including:

[0074] Input the temporal features into the hidden layer to construct a non-linear relationship for the position coordinates of the acceleration vibration signal to generate the first non-linear relationship feature.

[0075] S9. Use the fully connected layer module to extract coordinates from the first non-linear relationship feature to generate the position information of the impact load, including:

[0076] Input the first non-linear relationship feature into the first fully connected layer of the fully connected layer module to extract the x-axis coordinate of the impact load position, and input the first non-linear relationship feature into the second fully connected layer of the fully connected layer module to extract the y-axis coordinate of the impact load; finally generate the position information of the impact load.

[0077] In this embodiment, the x and y coordinates of the impact load position are predicted through the first fully connected layer and the second fully connected layer respectively; the combination of multiple levels and multiple modules enables SC-BiGRU to efficiently process complex time series data and ensure the stability and performance of the model.

[0078] In summary, a method for identifying the time series and position of non-linear impact loads based on deep learning proposed by the present invention combines the strong feature extraction ability of convolutional neural networks, the feature selection ability of self-attention mechanisms, and the time series processing ability of gated recurrent units, and shows significant advantages in processing impact load reconstruction tasks. By alternately arranging convolutional neural networks, self-attention mechanisms, and gated recurrent units, a neural network model based on self-attention mechanisms, encoders, and decoders is constructed to reconstruct the time history of impact loads; a neural network model of deep convolution and recursion based on self-attention mechanisms is constructed to achieve impact load positioning. This method has the advantages of significant recognition effect and high accuracy when facing the reconstruction of the magnitude and positioning of complex and non-linear impact loads.

[0079] Specific embodiments are applied in the present invention to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0080] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A method for identifying the time series and position of non-linear impact loads based on deep learning, characterized in that, It includes the following steps: Obtain acceleration vibration signal data; Construct a neural network model based on self-attention mechanism, encoder and decoder; The neural network model based on self-attention mechanism, encoder and decoder includes a convolutional encoder, a self-attention mechanism module, and a convolutional decoder; Use the convolutional encoder to extract features from the acceleration vibration signal data to generate a first output feature; After aggregating the spatial information of the first output feature using the self-attention mechanism module, generate a channel attention map, and by taking the channel attention map as an important factor for each channel, rescale the first output feature to generate a second output feature; Use the convolutional decoder to reconstruct high-dimensional features of the second output feature to generate the magnitude information of the impact load; Construct a neural network model of deep convolution and recursion based on self-attention mechanism; The neural network model of deep convolution and recursion based on self-attention mechanism includes an input layer module, a hidden layer module, and a fully connected layer module; After using the input layer module to extract spatial features from the acceleration vibration signal data, perform temporal feature extraction of the spatial features; Use the hidden layer module to construct a non-linear relationship for the temporal features to generate a first non-linear relationship feature; Use the fully connected layer module to extract coordinates from the first non-linear relationship feature to generate the position information of the impact load.

2. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, wherein Using the convolutional encoder to extract features from the acceleration vibration signal data to generate a first output feature, including: Input the acceleration vibration signal data into the first one-dimensional convolutional layer of the convolutional encoder for initial feature extraction, then input it into the first gated recurrent unit for temporal feature extraction of the initial features, then input it into the second one-dimensional convolutional layer for deep feature extraction, and then input it into the bidirectional gated recurrent unit for forward and backward temporal feature processing to generate the first output feature of the convolutional encoder.

3. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 2, wherein The formula for the first one-dimensional convolutional layer to perform initial feature extraction is: Among them, l represents the length of the input signal, and y j represents the initial feature of the j-th segment of the input signal, represents the output value of the j-th segment of the input signal in the convolutional kernel. *, [·] respectively represent the one-dimensional convolution operation and the splicing operation, tanh represents the hyperbolic tangent activation function, n represents the total number of convolutional kernels, and k ic represents the i-th convolutional kernel, s represents the size of the convolutional kernel, m represents the stride of the convolutional kernel, and x m(j-1):s+m(j-1) represents the j-th segment of the input signal, and b c represents the bias vector of the convolutional kernel.

4. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 3, wherein The formula for the first gated recurrent unit to perform temporal feature extraction of the initial features is: Among them, h t represents the final hidden state of the first gated recurrent unit at the t-th moment, z t represents the output of the update gate at the t-th moment, h t-1 represents the final hidden state of the first gated recurrent unit at the (t - 1)-th moment, r t represents the output of the reset gate at the t-th moment, represents the candidate hidden state at the t-th moment, σ represents the sigmoid activation function, x t represents the input vector at the t-th moment, that is, the initial feature input by the first one-dimensional convolutional layer, W z 、W r 、W h respectively represent the input weight matrices for calculating the update gate, reset gate, and candidate hidden state, U z 、U r 、U h respectively represent the hidden state weight matrices for calculating the update gate, reset gate, and candidate hidden state, b z 、b r 、b h respectively represent the bias vectors for calculating the update gate, reset gate, and candidate hidden state.

5. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, characterized in that, After aggregating the spatial information of the first output feature using the self-attention mechanism module, generate a channel attention map, and by taking the channel attention map as an important factor for each channel, rescale the first output feature to generate a second output feature, including: Input the first output feature into the self-attention mechanism module. First, use global average pooling and global max pooling operations to aggregate the spatial information of the first output feature to generate two different first spatial context descriptors and second spatial context descriptors. At the same time, input the first spatial context descriptor and the second spatial context descriptor into a shared network composed of a multi-layer perceptron and a hidden layer to generate a channel attention map; Take the channel attention map as an important factor for each channel and rescale the first output feature of the convolutional encoder to generate the second output feature of the self-attention mechanism module.

6. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 5, wherein The formula for the channel attention map is: M = σ(W1(W0(F avg )) + W1(W0(F max ))) Among them, M represents the channel attention map, σ represents the sigmoid activation function of the hidden layer, W0 and W1 respectively represent the weights of the first fully connected layer and the second fully connected layer, and the first fully connected layer and the second fully connected layer form a multi-layer perceptron, F avg and F max respectively represent the first spatial context descriptor and the second spatial context descriptor generated by global average pooling and global max pooling operations; among them, the multi-layer perceptron is composed of the first fully connected layer and the second fully connected layer.

7. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, wherein Using the convolutional decoder to reconstruct high-dimensional features of the second output feature to generate the magnitude information of the impact load, including: The second output feature is input into the first one-dimensional transposed convolutional layer for high-dimensional feature reconstruction, then input into the second gated recurrent unit for intermediate feature extraction of the reconstruction, and then input into the second one-dimensional transposed convolutional layer to complete the high-dimensional reconstruction, generating the impact load magnitude information.

8. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, characterized in that, After using the input layer module to extract the spatial features of the acceleration vibration signal data, the temporal features of the spatial features are extracted, including: The acceleration vibration signal data is input into the first one-dimensional convolutional layer of the input layer module for spatial feature extraction, and the maximum pooling operation is used to compress the spatial features, generating a compressed spatial feature sequence. Then it is input into the first bidirectional gated recurrent unit for forward and backward temporal feature extraction of the compressed spatial feature sequence. Then it is input into the self-attention mechanism module for impact signal information feature enhancement and suppression of redundant features including noise and irrelevant frequency components. Then it is input into the second one-dimensional convolutional layer for spatial feature extraction again, and the maximum pooling operation is used to compress the spatially re-extracted features. Then it is input into the second bidirectional gated recurrent unit for forward and backward temporal feature extraction of the spatially re-compressed features, generating the temporal features of the spatial features.

9. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, wherein Using the hidden layer module to construct the non-linear relationship of the temporal features, generating the first non-linear relationship feature, including: The temporal features are input into the hidden layer to construct the non-linear relationship of the acceleration vibration signal position coordinates, generating the first non-linear relationship feature.

10. The method for identifying the time series and position of non-linear impact loads based on deep learning according to claim 1, wherein Using the fully connected layer module to extract the coordinates from the first non-linear relationship feature, generating the position information of the impact load, including: The first non-linear relationship feature is input into the first fully connected layer of the fully connected layer module to extract the x-axis coordinate of the impact load position. The first non-linear relationship feature is input into the second fully connected layer of the fully connected layer module to extract the y-axis coordinate of the impact load. Finally, the position information of the impact load is generated.

Images