CN116842385B - LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics - Google Patents
LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics Download PDFInfo
- Publication number
- CN116842385B CN116842385B CN202310802518.6A CN202310802518A CN116842385B CN 116842385 B CN116842385 B CN 116842385B CN 202310802518 A CN202310802518 A CN 202310802518A CN 116842385 B CN116842385 B CN 116842385B
- Authority
- CN
- China
- Prior art keywords
- acceleration
- acceleration data
- road surface
- lstm
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 230000001133 acceleration Effects 0.000 claims abstract description 182
- 238000012549 training Methods 0.000 claims abstract description 20
- 238000004458 analytical method Methods 0.000 claims abstract description 14
- 238000012360 testing method Methods 0.000 claims abstract description 8
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000005070 sampling Methods 0.000 claims description 61
- 238000010586 diagram Methods 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 8
- 238000012544 monitoring process Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 abstract description 10
- 230000006403 short-term memory Effects 0.000 description 60
- 230000006870 function Effects 0.000 description 20
- 230000008569 process Effects 0.000 description 6
- 239000000284 extract Substances 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000015654 memory Effects 0.000 description 3
- 230000004044 response Effects 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000003595 spectral effect Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000002238 attenuated effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000011217 control strategy Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 239000000725 suspension Substances 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B21/00—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant
- G01B21/30—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant for measuring roughness or irregularity of surfaces
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
An LSTM road surface unevenness recognition method based on crawler vibration characteristics belongs to the technical field of neural networks. Comprising the following steps: collecting A, B, C, D acceleration data of four road surface grades; analyzing acceleration data characteristics of each level of road surface by using a time-frequency domain analysis method, and determining the size and the step length of a sliding window when an acceleration sample is manufactured; preprocessing the acquired acceleration data, and establishing a database; and (3) taking the database established in the step three as the input of the LSTM network, taking the road surface unevenness as the output of the LSTM network, and identifying the grade of the road surface unevenness by training and testing the LSTM network. The advantages are that: the LSTM is firstly applied to the recognition of the road surface unevenness of the tracked vehicle, and the accuracy is high; the method has the advantages that the vibration characteristics of the tracked vehicle are analyzed by the Fourier transform and wavelet transform methods, the size and the step length of the sliding window are determined according to the distribution of the vibration characteristics in the time domain, the data are conveniently sampled, and the final recognition accuracy can be improved.
Description
Technical Field
The invention belongs to the technical field of neural networks, and particularly relates to an LSTM (Short Term Memory) road surface irregularity recognition method based on vibration characteristics of a crawler.
Background
The crawler chassis (hereinafter referred to as crawler) is widely used in various modern engineering vehicles, and has large contact area between the crawler and the ground, good adhesion performance, and easy passage of obstacles such as soft ground and large ditches. The road surface unevenness is the main excitation of vibration in the running process of the tracked vehicle, and the excitation of the left and right tracks is transmitted to the vehicle body after being attenuated by the suspension, so that the bidirectional stability of the tracked vehicle is adversely affected, and the accurate reconstruction of the road surface unevenness is a key part of a dynamic model of the tracked vehicle during running and is also a basis for verifying the effectiveness of a control strategy.
With the increasing popularity of artificial intelligence, students at home and abroad begin to use neural networks to identify road surface irregularities. The neural network can rapidly extract deep features and has good robustness, and achieves good effect in recognition of road surface unevenness, but most of the prior art aims at wheeled vehicles, and compared with wheeled vehicles, the tracked vehicles have more complex structures, vibration responses generated by the tracked vehicles when driving on uneven road surfaces are also more complex, the road surface unevenness recognition is difficult to be performed on the tracked vehicles by the prior neural network recognition road surface technology, and in fact, the problem of recognition of the road surface unevenness between the traveling of the tracked vehicles can not be solved by the prior art, in view of the technical problem, the applicant has made beneficial designs, and the technical scheme to be introduced below is generated under the background.
Disclosure of Invention
The invention aims to provide an LSTM road surface unevenness identification method based on the vibration characteristics of a crawler, which is easy to operate and high in identification accuracy.
The invention aims at achieving the aim, namely an LSTM road surface unevenness identification method based on the vibration characteristics of a crawler, which is characterized by comprising the following steps:
step one: collecting A, B, C, D acceleration data of four road surface grades;
step two: analyzing acceleration data characteristics of each level of road surface by using a time-frequency domain analysis method, and determining the size and the step length of a sliding window when an acceleration sample is manufactured;
step three: preprocessing the acquired acceleration data, and establishing a database;
step four: and (3) taking the database established in the step three as the input of the LSTM network, taking the road surface unevenness as the output of the LSTM network, and identifying the grade of the road surface unevenness by training and testing the LSTM network.
In a specific embodiment of the present invention, in the first step, acceleration data of A, B, C, D pavement classes are acquired by:
on the road surfaces of four grades A, B, C, D, monitoring points are arranged on the mass center of the vehicle body to obtain the acceleration of the crawler during running in the directions of X axis, Y axis and Z axis;
setting sampling step sizes of more than two monitoring points to obtain acceleration data under different sampling step sizes; and then the acceleration data of the crawler during running under various working conditions are obtained by changing the running speed of the crawler.
In another specific embodiment of the present invention, in the second step, the method further includes the following steps:
firstly, carrying out frequency domain analysis on acceleration data by utilizing Fourier transformation to obtain a spectrogram of the acceleration data, and finding out a characteristic frequency range of the acceleration data of each level of pavement through the analysis of the spectrogram;
and then, performing time-frequency domain analysis on the acceleration data by utilizing wavelet transformation to obtain a time-frequency diagram of the acceleration data, and determining the size and the step length of a sliding window when an acceleration sample is manufactured according to the time distribution condition of the characteristic frequency.
In yet another specific embodiment of the present invention, in the third step, the method further includes the steps of:
firstly, sliding sampling is carried out on acceleration data in a window sliding mode, wherein the sliding window size and the step size under different sampling step sizes are different;
setting a grade label of the corresponding road unevenness for the sampled sample;
then dividing an initial sample of acceleration data into a training set and a testing set according to a certain proportion, wherein each sample contains the acceleration data and a label corresponding to the acceleration data, extracting the acceleration data independently for normalization processing, and scaling the value to be between-1 and 1;
finally, the arrangement order of the samples is disturbed.
In a further specific embodiment of the present invention, in the normalization process in the third step, the maximum value X of the acceleration data is calculated first max And a minimum value X min Then the acceleration data is normalized using the following formula,
wherein X is i Representing normalized acceleration sample vector, X max Representing an acceleration sample vector x i Maximum value of X min Representing an acceleration sample vector x i Is the minimum value of (a).
In a further specific embodiment of the present invention, in the fourth step, the LSTM network includes an input layer, three LSTM layers, three batch normalization (english: batch Normalization, english: BN) layers, and a full connection layer as output; the method comprises the steps of using a softmax (normalized index) cross entropy as a loss function, wherein the softmax cross entropy loss function comprises a softmax function and a cross entropy function, the softmax function is used for converting the output of a network into probability distribution, the predicted value of the output of the network is between 0 and 1, the sum of the predicted value of the output of the network and the predicted value of the output of the network is 1, the cross entropy function is used for measuring the difference between the predicted result of the network and a real label, the softmax cross entropy loss function is used for calculating the error between the output of acceleration data after passing through an LSTM network and the real label, and the LSTM network parameters are updated according to the obtained error so that the classification difference of the acceleration data is minimized.
In a further specific embodiment of the present invention, in the fourth step, when training the LSTM network, the optimal input scheme is determined by comparing the classification accuracy of different inputs under eight working conditions of 5km/h, 10km/h, 15km/h, 20km/h, 25km/h, 30km/h, 35km/h and 40km/h, wherein the input scheme is divided into vertical acceleration in two or more sampling steps of time and three directional acceleration in two or more sampling steps of time; when the input is vertical acceleration, the characteristic dimension of the LSTM network input acceleration data is 1; when the input is acceleration in three directions, the characteristic dimension of the LSTM network input acceleration data is 3, and the data respectively represent the acceleration data in the three directions of X axis, Y axis and Z axis.
Due to the adoption of the structure, compared with the prior art, the invention has the beneficial effects that: firstly, applying the LSTM network to the road surface unevenness recognition of the tracked vehicle for the first time, judging whether the LSTM network extracts effective features or not by using the accuracy index of the LSTM network, and determining an optimal input scheme by comparing the recognition accuracy of the LSTM network when the acceleration of different sampling step sizes and different dimensions are used as input, so that the LSTM network can extract the effective features to the greatest extent when the road surface unevenness is recognized, thereby determining the optimal input scheme and ensuring the recognition accuracy; secondly, for the problem of how to select the size and the step length of the sliding window, most of the existing researchers adopt a trial-and-error method or self experience to determine, but rarely adopt a theoretical analysis method.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of a multi-body dynamics model of an in-flight crawler;
FIG. 3a is a schematic view of acceleration in three directions of a class A road surface at a vehicle speed of 40 km/h;
FIG. 3B is a schematic view of acceleration in three directions of a B-class road surface at a vehicle speed of 40 km/h;
FIG. 3C is a schematic view of acceleration in three directions of a C-class road surface at a vehicle speed of 40 km/h;
FIG. 3D is a schematic view of acceleration in three directions of a grade D road surface at a vehicle speed of 40 km/h;
FIG. 4a is a graph of the vehicle vertical acceleration (0.5 second sample step) when traveling at 40km/h on a class A road;
FIG. 4B is a graph of the vehicle vertical acceleration (0.5 second sample step) when traveling at 40km/h on a B-level road;
FIG. 4C is a graph of the vehicle vertical acceleration (0.5 second sample step) when traveling at 40km/h on a grade C road;
FIG. 4D is a graph of the vehicle vertical acceleration (0.5 second sample step) when driving at a vehicle speed of 40km/h on a grade D road;
FIG. 5a is a wavelet time-frequency plot of vehicle body vertical acceleration (0.5 second sampling step) when traveling at 40km/h on a class A road surface;
FIG. 5B is a wavelet time-frequency plot of vehicle body vertical acceleration (0.5 second sampling step) when traveling at 40km/h on a B-level road surface;
FIG. 5C is a wavelet time-frequency plot of vehicle body vertical acceleration (0.5 second sampling step) when traveling at 40km/h on a grade C road;
FIG. 5D is a wavelet time-frequency plot of vehicle body vertical acceleration (0.5 second sampling step) when traveling at a vehicle speed of 40km/h on a grade D road;
FIG. 6a is a graph of the vehicle vertical acceleration (0.1 second sampling step) when traveling at 40km/h on a class A road;
FIG. 6B is a graph of the vehicle vertical acceleration (0.1 second sample step) when traveling at 40km/h on a B-level road;
FIG. 6C is a graph of the vehicle vertical acceleration (0.1 second sample step) when traveling at 40km/h on a grade C road;
FIG. 6D is a graph of the vehicle vertical acceleration (0.1 second sample step) when driving at a vehicle speed of 40km/h on a grade D road;
FIG. 7a is a wavelet time-frequency plot of vehicle body vertical acceleration (0.1 second sampling step) when traveling at 40km/h on a class A road surface;
FIG. 7B is a wavelet time-frequency plot of vehicle body vertical acceleration (0.1 second sampling step) when traveling at 40km/h on a B-level road surface;
FIG. 7C is a wavelet time-frequency plot of vehicle body vertical acceleration (0.1 second sampling step) when traveling at 40km/h vehicle speed on a grade C road surface;
FIG. 7D is a wavelet time-frequency plot of vehicle body vertical acceleration (0.1 second sampling step) when traveling at a vehicle speed of 40km/h on a grade D road;
FIG. 8 is a schematic diagram of an LSTM frame with acceleration in three directions of X-axis, Y-axis and Z-axis as input;
FIG. 9 is a histogram of classification accuracy for different input schemes under eight conditions;
FIG. 10 is a box plot of classification accuracy for different input schemes at a vehicle speed of 40 km/h;
FIG. 11 is a schematic diagram of a t-SNE visualization result with acceleration in three directions at a sampling step of 0.1 seconds as input at a vehicle speed of 40 km/h;
FIG. 12 is a schematic diagram of a confusion matrix with acceleration in three directions at a sampling step of 0.1 seconds as input at a vehicle speed of 40 km/h.
Detailed Description
The following detailed description of specific embodiments of the invention, while given in connection with the accompanying drawings, is not intended to limit the scope of the invention, and any changes that may be made in the form of the inventive concepts described herein, without departing from the spirit and scope of the invention.
In the following description, all concepts related to the directions (or azimuths) of up, down, left, right, front and rear are directed to the position states where the drawings are being described, so as to facilitate public understanding, and thus should not be construed as being particularly limiting to the technical solutions provided by the present invention.
The invention relates to an LSTM road surface unevenness recognition method based on crawler vibration characteristics, which is characterized in that monitoring points are arranged on the mass center of a vehicle body, and the acceleration data of a crawler in the advancing process under different road surface grades are measured under the conditions of different sampling step sizes and different speed grades respectively.
Step one: acceleration data for four road grades were collected A, B, C, D.
According to the method, the road surface power spectral density can divide the road surface unevenness into 8 grades according to the description of the national standard document 'mechanical vibration road surface spectral measurement data report', and road surface unevenness coefficient intervals corresponding to the road surfaces of A, B, C, D grades required by the invention are respectively [8,32], [32,128], [128,512] and [512,2048]. Acceleration data are acquired on the four grades of road surfaces of A, B, C, D respectively, and monitoring points are arranged on the mass center of the vehicle body to acquire the acceleration of the crawler in the X-axis, Y-axis and Z-axis directions of the traveling crawler.
Considering that the feature information contained in different sampling steps may be different, the classification accuracy under different sampling steps needs to be compared to judge which sampling step has more extracted features, and the method is more suitable for identifying the road surface unevenness of the tracked vehicle. In this embodiment, the sampling step sizes of the monitoring points are set to 0.5 seconds and 0.1 seconds, respectively, so that acceleration data under different sampling step sizes are obtained. Considering that the running speeds of the tracked vehicle are different, the vibration responses of the tracked vehicle are also different, so that eight working conditions of 5km/h, 10km/h, 15km/h, 20km/h, 25km/h, 30km/h, 35km/h and 40km/h are selected by changing the running speed of the tracked vehicle, the acceleration data of the tracked vehicle during running under the eight working conditions are obtained, and the acceleration data acquisition work of A, B, C, D road surface grades is finally completed.
Step two: and analyzing the acceleration data characteristics of each level of road surface by using a time-frequency domain analysis method, and determining the size and the step length of a sliding window when an acceleration sample is manufactured.
Firstly, carrying out frequency domain analysis on acceleration data by utilizing Fourier transformation to obtain a spectrogram of the acceleration data, and finding out a characteristic frequency range of each level of road surface acceleration data by analyzing the spectrogram.
Then, time-frequency domain analysis is carried out on the acceleration data by utilizing wavelet transformation, 4-order Daubechies wavelets are selected, scale parameters are set to be 500, a time-frequency diagram of the acceleration data is obtained, and the size and the step length of a sliding window when the acceleration sample is manufactured are determined according to the distribution condition of characteristic frequencies in time and the basic principle that each acceleration sample can contain enough characteristic frequencies.
Finally, after time-frequency domain analysis is carried out on acceleration data under eight working conditions, the size and the step length of the sliding window determined under the working conditions of 40km/h can enable acceleration samples of all working conditions to contain enough characteristic frequencies, so that the acceleration data of all working conditions uniformly adopts the sliding window determined under the working conditions of 40km/h for sliding sampling, the specific parameter of the sliding window is that the size of the sliding window under the sampling step length of 0.5 seconds is 48 time step lengths, namely 24 seconds, and the sliding step length is 5 time step lengths, namely 2.5 seconds; the sliding window size at a sampling step of 0.1 seconds is 301 time steps, i.e. 30.1s, and the sliding step is 5 time steps, i.e. 0.5s.
Step three: preprocessing the acquired acceleration data, and establishing a database.
The method comprises the steps of firstly adopting a window sliding mode to conduct sliding sampling on acceleration data. The basic idea of window sliding sampling is to divide acceleration data into a plurality of windows, wherein each window contains a certain number of time steps, and then the acceleration data in each window is used as an independent acceleration data sample for neural network training and testing, so that the task of classifying the acceleration data is realized. The sliding window size and the step size under different sampling step sizes are different, the sliding window size under the sampling step size of 0.5 seconds is 48 time step sizes, and the sliding step size is 5 time step sizes; the sliding window size at the 0.1 second sampling step is 301 time steps and the sliding step is 5 time steps.
Setting a grade label corresponding to the road unevenness for the sampled acceleration sample to obtain an initial sample set of acceleration dataWherein x is i Represents the ith acceleration sample vector, y under a certain working condition i And (3) representing a label corresponding to the ith acceleration sample vector, wherein n represents the sample number of a certain working condition.
The initial sample set of acceleration data is then divided into a training set and a test set in a ratio, here preferably 8:2, for each acceleration sample vector x i And (3) carrying out normalization processing, scaling the value between-1 and 1, wherein the calculation process is shown in the following formula (1):
wherein X is i Representing normalized acceleration sample vector, X max Representing an acceleration sample vector x i Maximum value of X min Representing an acceleration sample vector x i Is the minimum value of (a). Finally, the arrangement sequence of the samples is disturbed, so that the establishment of the database is completed.
Step four: and (3) taking the database established in the step three as the input of the LSTM network, taking the road surface unevenness as the output of the LSTM network, and identifying the grade of the road surface unevenness by training and testing the LSTM network. The specific method is that,
the LSTM network is composed of one input layer, three LSTM layers, three BN layers, and one fully connected layer as output.
The LSTM layer is a recurrent neural network with gating mechanism for processing time series data. Compared to a common recurrent neural network, the LSTM network introduces three gates: an input gate, a forgetting gate and an output gate to control the flow of information in the network, thereby avoiding the problems of gradient extinction or gradient explosion. The input gate determines which information is to be added to the cell state; forgetting gate determines whether certain information needs to be deleted from the cell state; the output gate determines what information is to be left from the cell state as the output of the current time step of the network, and the calculation is as follows, as shown in equations (2) - (7):
i t ＝σ(W i ·[h t-1 ,x t ]+b i ) (3)
f t ＝σ(W f ·[h t-1 ,x t ]+b f ) (4)
o t ＝σ(W o ·[h t-1 ,x t ]+b 0 ) (5)
h t ＝o t ·tanh(C t ) (7)
wherein i is t 、f t 、o t Respectively represent the output of the input gate, the forget gate and the output gate,representing candidate memory cells at time t, C t Representing memory cells at time t, C t-1 Indicating t-1 time memory cell, h t Represents the hidden state at the moment t, h t-1 Represents the hidden state at the moment t-1, x t Representing acceleration data input at time t, W c 、W i 、W f 、W o Representing a weight matrix, b c 、b i 、b f 、b o Representing the bias term, σ represents the sigmoid activation function, and tanh represents the hyperbolic tangent function.
The BN is a regularization method commonly used in neural networks. When constructing an LSTM network, the output of the LSTM layer has three dimensions, the first dimension representing the number of acceleration samples contained in each batch, the second dimension representing the length of the input acceleration samples, and the third dimension representing the hidden state dimension of the LSTM layer. Taking each LSTM layer output as an input to the BN layer, the BN layer will process the characteristic dimension (i.e., the dimension of the hidden state) of the LSTM layer output as shown in equation (8) below:
where x represents the output of the LSTM layer, y represents the output of the BN layer, ε represents a small constant to avoid dividing by zero, μ represents the mean value of the LSTM layer output hidden state dimensions, σ 2 Representing the variance of the LSTM layer output hidden state dimension, γ represents scaling, and β represents the amount of translation.
After the three LSTM layers and the corresponding BN layers, the fully-connected layer selects the output of the last time step as the input of the fully-connected layer, and the calculation formula is shown in (9):
f＝θ(β·X+b) (9)
where f represents the output of the fully connected layer, θ (·) represents the ReLU activation function, β represents the weight coefficient, X represents the output of the last time step of the BN layer, and b represents the bias.
Next, softmax cross entropy was used as the loss function. The softmax cross entropy loss function consists essentially of two parts: the invention relates to a softmax function and a cross entropy function, wherein the softmax function is used for converting the output of a network into probability distribution, so that the predicted value of the output of the network is between 0 and 1, the sum of the predicted value of the output of the network and the predicted value of the output of the network is 1, the cross entropy function is used for measuring the gap between the predicted result of the network and a real label, the error between the output of acceleration data after passing through an LSTM network and the real label is calculated by using the softmax cross entropy loss function, and a calculation formula is shown as (10):
wherein,representing the degree of difference between the predicted output of the LSTM network and the real label, N represents the batch size of the acceleration training samples, y i The true label representing the ith acceleration sample, L represents road grade, I {. Cndot. } represents the indication function, when the true label y of the ith acceleration sample i In the case of category L, I {.cndot.1, otherwise 0, ω n The weight matrix representing the nth road class, T representing the matrix transpose, f representing the output of the last layer of the LSTM network, and b representing the bias.
Based on the resulting error, LSTM network parameters are updated using an adaptive moment estimation (Adam) optimizer to minimize the classification differences of the acceleration data, with the optimization objective beingWherein θ is l For LSTM network parameters, the following formula (11) is obtained after optimization:
where η represents the learning rate.
The training process mainly comprises the following four steps: firstly, training all acceleration samples in batches, namely taking acceleration samples with batch size from an acceleration sample set during each training, and inputting the acceleration samples into an LSTM network for training; step two, calculating the actual output of the acceleration data after passing through the LSTM network, and transmitting the information from the input layer to the output layer through step-by-step conversion at the stage; thirdly, calculating the difference between the actually output acceleration data characteristic and the corresponding road surface grade label by using softmax cross entropy as a loss function; fourth, the LSTM network parameters are updated using Adam optimizers to minimize the classification differences of the acceleration data.
Further, in step four, when training the LSTM network, in order to enable the LSTM network to extract the effective features to the greatest extent, an optimal input scheme is determined by comparing the classification accuracy of different inputs, where the input scheme is divided into vertical acceleration at two or more sampling steps of time and acceleration at three directions at two or more sampling steps of time. In this embodiment, the following four input schemes are used for training: first, vertical acceleration at 0.5 second sampling step; second, vertical acceleration at 0.1 second sampling step; thirdly, acceleration in three directions under a sampling step length of 0.5 seconds; fourth, acceleration in three directions at a sampling step of 0.1 seconds. When the input is vertical acceleration, the characteristic dimension of the LSTM network input acceleration data is 1; when the input is acceleration in three directions, the characteristic dimension of the LSTM network input acceleration data is 3, and the data respectively represent the acceleration data in the three directions of X axis, Y axis and Z axis. The optimal input scheme is then determined by comparing the classification accuracy of the different inputs under eight conditions of 5km/h, 10km/h, 15km/h, 20km/h, 25km/h, 30km/h, 35km/h and 40 km/h. Experiments prove that the three-direction acceleration data are acquired by using a sampling step length of 0.1 second as input, and the time division performance is the best, so that the method is an optimal input scheme.
In order to verify the effectiveness of the technical scheme of the invention, the following simulation experiment is performed through specific examples.
Referring to fig. 1, the present embodiment follows the flowchart shown in fig. 1 and the steps described above.
Fig. 2 is a kinetic model of an ongoing crawler, using a turret as the kinetic model, with the following parameters: gun turret mass m 1 Distance L between turret rotation center and gun rotation center =5000 kg 1 =1.15m, artillery mass m 2 1950kg distance L between the centre of rotation of the cannon and the muzzle 2 =4.88 m, g=9.8 m/s 2 。
Table 1 road surface unevenness classifying table
The four grades of road surface unevenness coefficients G of A, B, C, D are shown in Table 1 q (n 0 ) Is a geometric mean of (c).
Fig. 3 (a), 3 (b), 3 (c) and 3 (d) show time and amplitude images of acceleration in three directions of A, B, C, D four road surface grades acquired respectively when the crawler travel speed is 40 km/h. Taking the Y-axis acceleration as an example, comparing fig. 3 (a), 3 (b), 3 (c) and 3 (d), it can be seen that the rougher the road surface, the greater the acceleration amplitude.
Fig. 4 (a), 4 (b), 4 (c) and 4 (d) show frequency charts of vertical acceleration of a vehicle body when the vehicle runs on a A, B, C, D four-grade road surface at a vehicle speed of 40km/h under a sampling step length of 0.5 seconds, and it can be seen from the figures that the higher the grade of the road surface is, the higher the amplitude of the frequency is, and each grade of road surface contains a plurality of frequency components, so that the frequency charts have more complex frequency spectrum characteristics, particularly, peak values appear every 0.0189-0.0756 Hz, and the frequency can be used as a characteristic frequency of the acceleration.
Fig. 5 (a), 5 (b), 5 (c) and 5 (d) show wavelet time-frequency diagrams of the vertical acceleration of the vehicle body when the vehicle runs on a A, B, C, D four-class road surface at a vehicle speed of 40km/h at a sampling step length of 0.5 seconds. In order to ensure that each sample contains as many characteristic frequencies as possible, and the characteristic frequencies of the pavement samples at each level are different so as to be convenient for distinguishing, the time range of the most obvious characteristic frequency of the pavement at each level is 15.5-28.5 s by taking the C-level pavement as a reference under comprehensive consideration, and the sliding window size should be 48 time steps, namely 24s, in order to be capable of covering the most obvious characteristic frequency. For other road surfaces, the sliding window of 24s can also contain enough characteristic frequency information, so that the sliding window size is finally determined to be 24s. In addition, it can be seen from the figure that the characteristic frequency of each level of road surface appears once every 2 s-10 s, and the interval of time is the most 2.5s, in order to increase the number of road surface samples of each level, an overlapping sampling method is adopted, and the step length of the sliding window takes 5 time steps, namely 2.5s.
FIGS. 6 (a), 6 (B), 6 (C) and 6 (D) show frequency charts of vertical acceleration of a vehicle body when the vehicle runs on A, B, C, D four grades of roads at a speed of 40km/h in a sampling step length of 0.1 seconds, frequency amplitude of an A-grade road acceleration signal is obvious between 0 and 1Hz and 2 and 4Hz, frequency amplitude of a B-grade road acceleration signal is obvious between 2 and 4Hz, frequency amplitude of a C-grade road acceleration signal is obvious between 3 and 5Hz, and frequency amplitude of a D-grade road acceleration signal is obvious between 2 and 5 Hz. Compared with the graphs in fig. 4 (a), 4 (b), 4 (c) and 4 (d), the characteristic frequencies of the road acceleration data of each level acquired under the sampling step length of 0.1 second are more obviously different, and the classification of the neural network is more facilitated.
Fig. 7 (a), 7 (b), 7 (c) and 7 (d) show wavelet time-frequency diagrams of the vertical acceleration of the vehicle body when the vehicle was driven on a A, B, C, D four-class road surface at a vehicle speed of 40km/h in a sampling step of 0.1 seconds. The acceleration wavelet time-frequency diagram of the class-A pavement has obvious characteristic frequency in the time range of 34.2-35.2 s, and the sliding window size is 301 time steps, namely 30.1s, in order to enable the sample to contain the characteristic frequency. B. The characteristic frequency of the C, D-grade pavement is very dense at all time points, and if a sliding window of 30.1s is taken, the characteristic frequency of the B, C, D-grade pavement can be contained, so that the sliding window is finally determined to be 30.1s. In addition, the four grades of pavement can have a more obvious frequency every 0.5 s-3.5 s, wherein 0.5s is more, in order to increase the number of samples of each grade of pavement, an overlapping sampling method is adopted, and the step length of the sliding window takes 5 time steps, namely 0.5s.
TABLE 2 basic Structure and parameter Table of LSTM
| Network layer | Parameter information |
| Input layer | / |
| LSTM1 | Hidden layer node number: 64, BN |
| LSTM2 | Hidden layer node number: 64, BN |
| LSTM3 | Hidden layer node number: 64, BN |
| Full connection layer | Weight: 64×class number, bias: number of categories |
Referring to Table 2, in this embodiment, three layers of LSTM are stacked, each with 64 nodes, and each layer of LSTM is followed by a BN layer that keeps the inputs of each layer of the network equally distributed. After passing through the three LSTM layers and the corresponding BN layer, the output of the last time step is selected as the input of the fully connected layer, whose output is the final extracted feature. Next, a softmax function is used on the output of the output layer to obtain a prediction classification. In addition, the training frequency of the network was set to 200, and the learning rate was selected from {0.001,0.0001,0.00001} according to the training frequency, that is, the first 50 times selected the learning rate of 0.001, the second 50 times selected the learning rate of 0.00001, and the middle selected the learning rate of 0.0001.
FIG. 8 illustrates an LSTM framework with inputs of acceleration in three directions, X-axis, Y-axis, Z-axis, where the sample input dimension of LSTM is 3.
Table 3 classification accuracy table for different input schemes under eight conditions
Table 3 shows the classification accuracy of the different input schemes under eight working conditions, and it can be seen from the table that the classification accuracy obtained by adopting the sampling step length of 0.1 second under eight working conditions is higher than the classification accuracy obtained by adopting the sampling step length of 0.5 second. Under the same sampling step length, the classification accuracy obtained by taking the acceleration in three directions as input is lower than the classification accuracy obtained by taking the acceleration in the vertical direction only by 0.02 percent, and the classification accuracy obtained by taking the acceleration in the three directions as input is higher than or equal to the classification accuracy obtained by taking the acceleration in the vertical direction only as input. The best classification effect is a scheme taking acceleration in three directions under a sampling step length of 0.1 second as input, the classification accuracy is higher than 98% in eight working conditions and is obviously higher than other schemes, and the accuracy can reach 100% under two working conditions of 5km/h, 15km/h, 20km/h, 30km/h, 35km/h and 40 km/h.
Fig. 9 is a histogram of classification accuracy for eight different input schemes under eight conditions, which more clearly shows the comparison between the schemes.
Fig. 10 shows a classification accuracy box diagram of different input schemes at a vehicle speed of 40km/h, and the accuracy obtained by training 50 times after each method is selected during drawing, so that the accuracy box diagrams of the other three methods are all in a straight line except for the method taking the vertical acceleration with the sampling step length of 0.1 as the input, the accuracy median of the three methods is equal to the upper and lower quartiles, and the accuracy distribution is centralized. The upper quartile and median of the accuracy bin diagram obtained with the 0.1 sample step vertical acceleration as input coincide and there is no upper edge, which means that the accuracy is very concentrated in the distribution of the upper half, most of the accuracy is concentrated near the median and there is no significant outlier beyond the upper edge. It can also be seen from the distribution of four method accuracies that the three-direction acceleration at the 0.1 second sampling step has the optimal classification performance as input.
FIG. 11 shows the result of t-SNE (English full name: t-Distributed Stochastic Neighbor Embedding, chinese name: t distributed random neighborhood embedding) with three directional accelerations at 0.1 second sampling steps as input at a vehicle speed of 40 km/h. the t-SNE maps the extracted features to a two-dimensional space, so that the classification effect of the invention can be more intuitively presented, and all the categories can be clearly distinguished from the figure, so that the better classification performance of taking the acceleration in three directions under the sampling step length of 0.1 second as an input scheme is demonstrated.
Fig. 12 shows a confusion matrix with acceleration in three directions at a sampling step of 0.1 second as input at a vehicle speed of 40km/h, and it can be seen from the figure that all road surface unevenness levels can be correctly identified.
According to the invention, the LSTM network is applied to the road surface unevenness recognition of the tracked vehicle for the first time, the vibration response of the tracked vehicle is different from that of the wheeled vehicle, and the technical problem solved by the invention is that how to select the input scheme of acceleration data so that the LSTM network can extract effective characteristics to the greatest extent. The accuracy index of the LSTM network can judge whether the LSTM network extracts effective features, and the specific embodiment shows that the detection method is truly effective.
Claims (4)
1. An LSTM road surface unevenness identification method based on crawler vibration characteristics is characterized by comprising the following steps:
step one, collecting A, B, C, D acceleration data of four road surface grades,
in the first step, A, B, C, D four road surface grade acceleration data are collected,
on the road surfaces of four grades of A, B, C, D, monitoring points are arranged on the mass center of the vehicle body to obtain the acceleration of the crawler during running in the directions of X axis, Y axis and Z axis,
setting sampling step sizes of more than two monitoring points to obtain acceleration data under different sampling step sizes; obtaining acceleration data of the crawler during running under various working conditions by changing the running speed of the crawler;
step two, analyzing the acceleration data characteristics of the road surface of each grade by using a time-frequency domain analysis method, determining the size and the step length of a sliding window when an acceleration sample is manufactured,
in the second step, the method further comprises the following steps,
firstly, carrying out frequency domain analysis on acceleration data by utilizing Fourier transformation to obtain a spectrogram of the acceleration data, and finding out a characteristic frequency range of the acceleration data of each level of pavement through the analysis of the spectrogram;
then, performing time-frequency domain analysis on the acceleration data by utilizing wavelet transformation to obtain a time-frequency diagram of the acceleration data, and determining the size and the step length of a sliding window when an acceleration sample is manufactured according to the time distribution condition of the characteristic frequency, wherein the size and the step length of the sliding window under different sampling step lengths are different;
step three, preprocessing the collected acceleration data, and establishing a database;
step four, taking the database established in the step three as the input of the LSTM network, taking the road surface unevenness as the output of the LSTM network, identifying the grade of the road surface unevenness by training and testing the LSTM network,
in the fourth step, when training the LSTM network, determining an optimal input scheme by comparing classification accuracy of different inputs under eight working conditions of 5km/h, 10km/h, 15km/h, 20km/h, 25km/h, 30km/h, 35km/h and 40km/h, wherein the input scheme is divided into vertical acceleration under sampling step sizes of two or more times and three directional acceleration under sampling step sizes of two or more times; when the input is vertical acceleration, the characteristic dimension of the LSTM network input acceleration data is 1; when the input is acceleration in three directions, the characteristic dimension of the LSTM network input acceleration data is 3, and the data respectively represent the acceleration data in the three directions of X axis, Y axis and Z axis.
2. The LSTM road surface unevenness identification method based on a vibration characteristic of a crawler according to claim 1, wherein in the third step, the method further comprises the steps of:
firstly, sliding sampling is carried out on acceleration data in a window sliding mode;
setting a grade label of the corresponding road unevenness for the sampled sample;
then dividing an initial sample of acceleration data into a training set and a testing set according to a certain proportion, wherein each sample contains the acceleration data and a label corresponding to the acceleration data, extracting the acceleration data independently for normalization processing, and scaling the value to be between-1 and 1;
finally, the arrangement order of the samples is disturbed.
3. The LSTM road surface unevenness recognition method based on the vibration characteristics of the crawler according to claim 2, wherein the normalization in the third step is performed by calculating the maximum value X of the acceleration data max And a minimum value X min Then the acceleration data is normalized using the following formula,
wherein X is i Representing normalized acceleration sample vector, X max Representing an acceleration sample vector x i Maximum value of X min Representing an acceleration sample vector x i Is the minimum value of (a).
4. The LSTM road surface unevenness recognition method based on a vibration feature of a crawler according to claim 1, wherein in the fourth step, the LSTM network includes an input layer, three LSTM layers, three batch normalization layers, and a full connection layer as an output; the method comprises the steps of using softmax cross entropy as a loss function, wherein the softmax cross entropy loss function comprises a softmax function and a cross entropy function, the softmax function is used for converting output of a network into probability distribution, so that predicted values of the output of the network are between 0 and 1, the sum of the predicted values is 1, the cross entropy function is used for measuring the difference between a network predicted result and a real label, calculating an error between the output of acceleration data after passing through an LSTM network and the real label by using the softmax cross entropy loss function, and updating LSTM network parameters according to the obtained error so that classification difference of the acceleration data is minimized.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310802518.6A CN116842385B (en) | 2023-06-30 | 2023-06-30 | LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310802518.6A CN116842385B (en) | 2023-06-30 | 2023-06-30 | LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116842385A CN116842385A (en) | 2023-10-03 |
| CN116842385B true CN116842385B (en) | 2024-03-19 |
Family
ID=88173806
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310802518.6A Active CN116842385B (en) | 2023-06-30 | 2023-06-30 | LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116842385B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111597639A (en) * | 2020-05-19 | 2020-08-28 | 北京理工大学 | Coupling analysis method for crawler belt and cross-country road surface |
| CN111967308A (en) * | 2020-07-01 | 2020-11-20 | 清华大学 | Online road surface unevenness identification method and system |
| CN112836576A (en) * | 2020-12-30 | 2021-05-25 | 北京理工大学 | Rapid analysis method for cross-country road surface model |
| CN113932758A (en) * | 2021-09-15 | 2022-01-14 | 同济大学 | Road surface flatness prediction method and device |
| CN115294449A (en) * | 2022-07-21 | 2022-11-04 | 中国北方车辆研究所 | Ground identification method for crawler robot |
| CN115758212A (en) * | 2022-11-10 | 2023-03-07 | 四川轻化工大学 | Mechanical equipment fault diagnosis method based on parallel network and transfer learning |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP6032574B2 (en) * | 2012-08-26 | 2016-11-30 | 国立大学法人大阪大学 | Fourier transform spectroscopy, spectroscopic device and spectroscopic measurement program for improving spectral resolution and spectral accuracy |
| US9507042B2 (en) * | 2012-08-31 | 2016-11-29 | Lumina Geophysical LLC | System and method for constrained least-squares spectral processing and analysis of seismic data |
| US11928894B2 (en) * | 2012-09-18 | 2024-03-12 | Origin Wireless, Inc. | Method, apparatus, and system for wireless gait recognition |
-
2023
- 2023-06-30 CN CN202310802518.6A patent/CN116842385B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111597639A (en) * | 2020-05-19 | 2020-08-28 | 北京理工大学 | Coupling analysis method for crawler belt and cross-country road surface |
| CN111967308A (en) * | 2020-07-01 | 2020-11-20 | 清华大学 | Online road surface unevenness identification method and system |
| CN112836576A (en) * | 2020-12-30 | 2021-05-25 | 北京理工大学 | Rapid analysis method for cross-country road surface model |
| CN113932758A (en) * | 2021-09-15 | 2022-01-14 | 同济大学 | Road surface flatness prediction method and device |
| CN115294449A (en) * | 2022-07-21 | 2022-11-04 | 中国北方车辆研究所 | Ground identification method for crawler robot |
| CN115758212A (en) * | 2022-11-10 | 2023-03-07 | 四川轻化工大学 | Mechanical equipment fault diagnosis method based on parallel network and transfer learning |
Non-Patent Citations (7)
| Title |
|---|
| Assessment of acceleration responses of a railway bridge using wavelet analysis;Mohamed A. Sayed et al.;《Structural Engineering》;20160902;第21卷;全文 * |
| Comparative study of Hilbert–Huang transform, Fourier transform and wavelet transform in pavement profile analysis;A. Y. Ayenu-Prah et al.;《International Journal of Vehicle Mechanics and Mobility》;20090318;第47卷;全文 * |
| Road Surface Classification Using a Deep Ensemble Network with Sensor Feature Selection;Jongwon Park et al.;《Sensors 2018》;20181209;全文 * |
| S变换谱分解技术在深反射地震弱信号提取中的应用;邓攻;梁锋;李晓婷;赵俊猛;刘红兵;王洵;;地球物理学报;20151215(12);全文 * |
| 基于履带车辆车体动态响应的行驶路面不平度识别;凌启辉 等;《中国机械工程》;20220131;第33卷;全文 * |
| 时频分析在发动机振动监测与故障诊断中的应用;李飞行;王涛;张群岩;;测控技术;20110218(02);全文 * |
| 自行火炮行进间射击误差影响因素仿真分析;王军 等;《系统仿真技术及其应用》;20201231;第21卷;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116842385A (en) | 2023-10-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111192284B (en) | Vehicle-mounted laser point cloud segmentation method and system | |
| CN112462346A (en) | Ground penetrating radar roadbed defect target detection method based on convolutional neural network | |
| CN111898634A (en) | Intelligent fault diagnosis method based on depth-to-reactance-domain self-adaption | |
| CN109345007B (en) | Advantageous reservoir development area prediction method based on XGboost feature selection | |
| CN113009447B (en) | Road underground cavity detection and early warning method based on deep learning and ground penetrating radar | |
| CN114783183A (en) | Monitoring method and system based on traffic situation algorithm | |
| CN111292534A (en) | Traffic state estimation method based on clustering and deep sequence learning | |
| WO2021013190A1 (en) | Meteorological parameter-based high-speed train positioning method and system in navigation blind zone | |
| CN114564982A (en) | Automatic identification method for radar signal modulation type | |
| CN114675261B (en) | Urban underground disease body full life cycle management method based on ground penetrating radar | |
| CN106874950A (en) | A kind of method for identifying and classifying of transient power quality recorder data | |
| CN112100930B (en) | Formation pore pressure calculation method based on convolutional neural network and Eaton formula | |
| CN106023592A (en) | Traffic jam detection method based on GPS data | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN112884014A (en) | Traffic speed short-time prediction method based on road section topological structure classification | |
| CN111461183A (en) | Flat wheel fault detection method for optimizing BP neural network by improving artificial fish swarm algorithm | |
| CN114882069A (en) | Taxi track abnormity detection method based on LSTM network and attention mechanism | |
| CN113033899A (en) | Unmanned adjacent vehicle track prediction method | |
| CN111967308A (en) | Online road surface unevenness identification method and system | |
| CN110673208B (en) | First arrival picking method and system for high-dimensional feature constraint under machine learning framework | |
| CN113064133B (en) | Sea surface small target feature detection method based on time-frequency domain depth network | |
| CN114332722A (en) | Real-time estimation method for adhesion coefficient of mixed ice and snow road surface based on video data | |
| CN116842385B (en) | LSTM road surface unevenness identification method based on tracked vehicle vibration characteristics | |
| CN116466408B (en) | Artificial neural network superbedrock identification method based on aeromagnetic data | |
| CN112860658A (en) | High-speed train electromagnetic emission prediction method based on data mining |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |