CN112985830B - Automatic abs result judging algorithm - Google Patents
Automatic abs result judging algorithm Download PDFInfo
- Publication number
- CN112985830B CN112985830B CN202110161343.6A CN202110161343A CN112985830B CN 112985830 B CN112985830 B CN 112985830B CN 202110161343 A CN202110161343 A CN 202110161343A CN 112985830 B CN112985830 B CN 112985830B
- Authority
- CN
- China
- Prior art keywords
- data
- layer
- abs
- vehicle
- training
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M17/00—Testing of vehicles
- G01M17/007—Wheeled or endless-tracked vehicles
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Regulating Braking Force (AREA)
Abstract
The invention provides an abs result automatic judging algorithm, which comprises the following steps of S1: acquiring measurement data and additional information related to the detection of the ABS braking performance of the automobile in the braking process; s2, carrying out data preprocessing on the acquired measurement data and the additional information; s3: making sample data with labels; s4: the training of the non-supervision deep learning network model utilizes the cloud database to accumulate real vehicle measurement and corresponding indoor whole vehicle test data, and can more comprehensively detect the braking execution conditions of vehicles assembled with ABS on various road surfaces, thereby ensuring the safety of vehicle braking, reducing traffic accidents, having higher practical value.
Description
[ technical field ]
The invention relates to the technical field of ABS braking performance detection, in particular to an ABS result automatic judging algorithm with outstanding application effect.
[ background Art ]
With the wider application of anti-lock brake systems (ABS for short) on automobiles, in the whole automobile detection line of automobiles, corresponding devices for detecting the brake performance of the ABS of automobiles are required to be added, and the devices judge the working performance of the ABS by detecting the wheel speed, the speed of the automobile body, the pedal, the pressure of a pipeline and the like when the automobile brakes. Since the common detection personnel does not have the relevant expertise, the automobile ABS brake performance detection device can automatically judge the working state of the ABS of the detected automobile.
In the massive monitoring data, because the occurrence probability of faults is very low compared with the normal working condition and the manual marking of information is difficult, a large number of high-value marked sample sets are generally difficult to construct. How to perform effective state monitoring and fault identification under the condition of limited or lacking training sample label information is a key problem which is not only challenging but also has important application value.
[ summary of the invention ]
In order to overcome the problems in the prior art, the invention provides an abs result automatic judging algorithm with outstanding application effect.
The technical problem to be solved by the invention is to provide an abs result automatic judging algorithm which comprises the following steps,
s1: acquiring measurement data and additional information related to the detection of the ABS braking performance of the automobile in the braking process;
s2, carrying out data preprocessing on the acquired measurement data and the additional information;
s3: making sample data with labels;
s4: training of an unsupervised deep learning network model.
Preferably, the measured data in the step S1 includes vehicle speed, wheel speed, pedal force, abs braking time, interval, brake line pressure data; the additional information comprises vehicle type, vehicle number, vehicle age, road position and vehicle type technical parameters.
Preferably, the step S1 further includes dividing the acquired measurement data and additional information into structured data and unstructured data; the structured data comprises numerical data and a database; the unstructured data includes text-type data or time domain waveform diagrams.
Preferably, the step S2 specifically includes the following steps, A1: intercepting data of the length related to abs braking action in the waveform time curve of the measured data as original waveform data; a2: removing baseline drift noise by adopting a high-pass filter; a3: confirming whether the noise is too high or not based on a standard variance and a threshold method, and removing noise interference by using a low-pass Butterworth filter when the noise is too high; a4: and carrying out normalization processing on all data according to the characteristics of multi-source and heterogeneous data sources.
Preferably, in step S3, measured data of the Abs braking system running under different working conditions have a certain similarity, and the time domain distribution diagram of the fault data has a certain similarity, that is, the source domain and the target domain have a common part.
Preferably, in the step S3, a tag is added to the processed data under different working conditions and is denoted as Source Domain (SD); the evaluation conclusion of the integral state of the ABS is that the label comprises normal, fault and deficiency; the preprocessed data collected in real time is denoted as a Target Domain (TD).
Preferably, the unsupervised deep learning network model in step S4 includes a primary network and a secondary network; the main network is provided with a feedforward connection structure and a feedback connection structure and has local memory capacity, and the main network is sequentially composed of an input layer, a bearing layer, an intermediate layer and an output layer; the secondary network is used for primary network initialization, and the secondary network comprises a visual layer and an implicit layer.
Preferably, the training process of step S4 is divided into two phases: firstly, adopting a bottom-up unsupervised training mode; secondly, adopting a top-down supervised learning mode; the bottom-up training is layered training with no calibration data or with calibration data; firstly, inputting training samples, learning weights of a first layer of a network until a model n-1 layer, wherein the output of the training samples is used as the model n-layer input, so that neuron parameters of each layer are obtained, and the training process is unsupervised; the top-down learning is to further train the network through the data with the labels after the first training process is finished, so that errors are transmitted layer by layer from top to bottom, fine adjustment is carried out on the network parameters obtained through pre-training, and finally network parameters of each layer of the model are determined.
Preferably, in the vehicle type and/or ABS braking system with abundant accumulated fault data, under the condition of limited or lacking training sample label information, fault characteristics are comprehensively and effectively extracted from measured vehicle data and additional information of multiple sources such as different vehicle types, vehicle numbers, vehicle ages, road positions, numerical values, characters, pictures and the like, so that the system is suitable for automatic judgment of the ABS working state of the vehicle under the condition of class label missing; the method comprises the following specific steps: A. unstructured data is used as input of a convolutional neural network, and structured data is used as input of a deep neural network; B. the method comprises the steps that through a feature fusion layer comprising a plurality of hidden layers, a full-connection layer of CNN and neurons in a hidden layer of the last layer of DNN are fully connected with neurons in a hidden layer of the first layer of the feature fusion layer, and fault features extracted from unstructured data by CNN and fault features extracted from structured data by DNN are seamlessly integrated through full-connection operation; C. and inputting the output of the last hidden layer of the feature fusion layer into a softmax classifier to classify faults.
Preferably, under different data sets of different working conditions, implicit representative features are automatically extracted, and a general fault diagnosis model is established, so that fault diagnosis of the abs system can be realized on the premise of changing the working conditions; the method comprises the following specific steps: A. acquiring ABS brake performance detection data, analyzing an ABS detection result, and selecting m characteristic parameters capable of reflecting the ABS working performance of the automobile; B. making sample data with labels; C. performing data cleaning processing on n sample data (n=ns+nt) of the characteristic parameters; D. and constructing an unsupervised deep learning network model.
Compared with the prior art, the ABS result automatic judging algorithm utilizes the cloud database to accumulate real vehicle measurement and corresponding indoor whole vehicle test data, and can more comprehensively detect the braking execution conditions of vehicles assembled with the ABS on various road surfaces, thereby ensuring the safety of vehicle braking, reducing traffic accidents and having higher practical value.
[ description of the drawings ]
Fig. 1 is a block diagram of a database application system of the present invention.
Fig. 2 is an unsupervised deep learning model of the present invention.
Fig. 3 is an ABS braking performance test result automatic determination algorithm 1 of the present invention.
Fig. 4 is an ABS braking performance test result automatic determination algorithm 2 of the present invention.
FIG. 5 (a) is a graph of vehicle speed/wheel speed contrast for normal operation of the ABS of the present invention.
Fig. 5 (b) is a vehicle speed/wheel speed comparison curve at the time of ABS failure of the present invention.
FIG. 6 (a) is a graph showing the comparison of vehicle speed and wheel speed for the normal operation of the ABS under the same road surface with the same adhesion coefficient according to the present invention.
FIG. 6 (b) is a graph showing the comparison of vehicle speed and wheel speed when the ABS braking force is insufficient under the same road surface with the adhesion coefficient according to the present invention.
FIG. 7 is a flow chart of the ABS brake data dimension reduction process of the present invention.
FIG. 8 is a flow chart of an abs result automatic determination algorithm according to the present invention.
Detailed description of the preferred embodiments
For the purpose of making the technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Referring to the following views, an abs braking fault diagnosis system database application system of an automobile corresponding to an abs result automatic determination algorithm of the present invention, as described in the following figure 1,
the method specifically comprises the following steps: the system comprises a private data interface 1, a public data interface 2, an Ethernet 3, a brake platform electric control cabinet 4, an abs brake detection platform 5, a tested vehicle 6, an ODB adapter 7 and a station computer 8.
The station computer 8 is connected with the brake table electric control cabinet 4 and the ODB adapter 7 through a field bus (RS 232/CAN bus and the like); and can communicate with the private data interface 1 and the public data interface 2 through the Ethernet 3.
The other end of the OBD adapter 7 is connected with a vehicle-mounted OBD interface of the tested vehicle 6 through a vehicle-mounted diagnosis protocol (K line/CAN bus) and is communicated with an ABS Electronic Control Unit (ECU) on the OBD interface.
The brake table electric control cabinet 4 is provided with an embedded control board card and an A/D signal conversion board card and is connected with the abs brake detection table 5 through a signal line.
By way of example, these components may have the following functions:
1. and (3) a station computer:
a. the system is communicated with an electric control cabinet of the brake platform, coordinates the tested vehicle and an abs brake detection platform for detection, and receives abs brake test data;
b. the method comprises the steps of indicating a tested vehicle to brake, and receiving abs braking operation data of the tested vehicle through an OBD adapter;
c. embedding a deep learning algorithm model to realize database application of the deep neural network model and automatically judging abs braking performance of the tested vehicle;
d. an embedded human-machine interface (HMI), a visual chart and the like are used for realizing daily intelligent monitoring and interaction of industrial equipment;
e. and the system is connected and communicated with the private data interface and the public data interface, and is used for issuing data and receiving data feedback.
2. abs braking detection table and braking table electric control cabinet
a. Performing abs braking detection;
b. collecting abs braking test data;
c. the collected data is conditioned into a digital signal which can be identified by a computer.
3. Tested vehicle and ODB adapter
a. Executing abs braking;
b. collecting abs braking operation data;
c. the collected data is conditioned into a digital signal which can be identified by a computer.
4. Private data interface and public data interface
a. Connecting a private cloud platform and a public cloud platform;
b. receiving test data and test results issued by a station computer;
c. receiving an abs braking data annotation sample set aiming at a specific vehicle type and a specific vehicle number;
d. and feeding back the trained deep neural network model to the station computer aiming at the specific vehicle model and the specific vehicle number, and/or finishing an abs braking annotation sample set required by the training of the deep neural network model by the station computer.
The invention relates to an automatic judging algorithm of an ABS braking performance test result based on an unsupervised deep learning model, which is shown in the figure 2, and specifically comprises the following steps:
1. acquiring measurement data and additional information related to the detection of the ABS braking performance of the automobile in the braking process;
a. the detection data comprise measurement data of braking processes such as vehicle speed, wheel speed, pedal force, abs braking time, interval, braking pipeline pressure and the like;
b. the additional information includes, but is not limited to: vehicle type, number, age of vehicle, road location; vehicle model technical parameters, such as: the method comprises the steps of obtaining dynamic data such as wheelbase, wheel track, preparation quality, tire specification, ABS form, ABS signals, power, torque and the like from a preset technical parameter database of vehicle types and vehicle numbers;
c. in one embodiment of the invention, the acquired measurement data and additional information are distinguished into structured data and unstructured data. Structured data, e.g., numerical data, databases, etc.; unstructured data, e.g., literal data, time domain waveforms, etc.;
2. the data preprocessing of the acquired measurement data and the additional information comprises the following steps:
a. intercepting data of the length related to abs braking action in the waveform time curve of the measured data as original waveform data;
b. removing baseline drift noise by adopting a high-pass filter;
c. confirming whether the noise is too high or not based on a standard variance and a threshold method, and removing noise interference by using a low-pass Butterworth filter when the noise is too high;
d. aiming at the characteristics of multi-source and heterogeneous data sources in the invention, all data are normalized;
3. making labeled sample data:
the measured data of the abs braking system running under different working conditions have certain similarity, the time domain distribution diagram of the fault data has certain similarity, namely the source field and the target field have public parts, which is the premise of adopting an unsupervised deep learning network model in the invention.
b. In one embodiment of the invention, the processed data under different working conditions are added with labels and marked as Source Domain (SD for short), for example, the labels comprise normal, fault and deficiency; the preprocessed data collected in real time is denoted as a Target Domain (TD).
4. Training of unsupervised deep learning network model
a. The feature extraction of the input data is a very important step in the machine learning algorithm and is also a stage of the most time consumption; the characteristic indexes are extracted by relying on human experience and a large amount of experimental research before fault identification, the workload is large, and whether the selected characteristic indexes can completely express the running mode of the abs braking process cannot be verified;
the invention adopts an unsupervised deep learning network model, skips the stage of feature design in the whole learning process, can directly learn the input signals in different modes, and learns all feature information of the signals through layer-by-layer feature transformation in the model learning process, thereby obtaining a global rather than locally optimal fault recognition result;
the model may include: a primary network and a secondary network; the main network is provided with a feedforward connection structure and a feedback connection structure and has local memory capacity, and the main network is sequentially composed of an input layer, a bearing layer, an intermediate layer and an output layer; the secondary network is used for initializing the primary network and comprises a visual layer and an implicit layer;
the tasks of the model may be classification or regression, related to the modeled problem itself and the form of the labels. The classification task in the invention can be the automatic judgment of abs braking performance detection results; the regression task can be to predict the next abs braking performance fault type and occurrence time of the tested vehicle according to the current detection result.
b. The training process of the unsupervised deep learning network model is divided into two stages: firstly, adopting a bottom-up unsupervised training mode; secondly, adopting a top-down supervised learning mode;
the bottom-up training is layered training by using no calibration data (the calibration data can be used); firstly, inputting training samples, learning weights of a first layer of a network until a model n-1 layer, wherein the output of the training samples is used as the model n-layer input, so that neuron parameters of each layer are obtained, and the training process is unsupervised;
in the training process, the model has the capability of learning all information contained in massive input data due to the characteristics of capacity limitation, sparsity constraint and the like of the depth model; the method is the greatest difference from the traditional neural network, and the model initial value is obtained in the internal structure of the learning training sample and is closer to the global optimum of the network, so that more accurate effects can be obtained than the neural network;
the top-down learning is to further train the network through the data with the labels after the first training process is finished, so that errors are transmitted layer by layer from top to bottom, fine adjustment is carried out on the network parameters obtained through pre-training, and finally network parameters of each layer of the model are determined.
The training process of the unsupervised deep learning network model depends on big data, and the data acquisition mode in the invention comprises road test, bench test, vehicle-mounted tracking measurement, available database or other ways.
In the case of a vehicle type and/or ABS braking system with accumulated rich fault data, as shown in the following fig. 3, the algorithm embodiment of the invention aims at comprehensively and effectively extracting fault characteristics from measured data and additional information of a detected vehicle with multiple structures from multiple sources such as different vehicle types, vehicle numbers, vehicle ages, road positions, numerical values, characters, pictures and the like under the condition of limited or lack of training sample label information, so that the algorithm is suitable for automatic judgment of the ABS working state of the vehicle under the condition of class label deficiency; the method comprises the following specific steps:
1. firstly, unstructured data is used as input of a convolutional neural network, and structured data is used as input of a deep neural network;
a. the Convolutional Neural Network (CNN) consists of a convolutional layer, a sub-sampling layer and a full-connection layer, and the output of the full-connection layer is the fault characteristic extracted from unstructured data through operations such as convolution, pooling and the like;
b. the Deep Neural Network (DNN) has a plurality of hidden layers, the first of which extracts basic low-level features from the raw data, the subsequent hidden layers convert them layer by layer into more abstract high-level features that can describe the data distribution more accurately; DNN can adaptively learn some deep hiding rules from sample data without requiring specialized domain expertise; the output of the last hidden layer of DNN is the fault feature extracted from the structured data;
2. the method comprises the steps that through a feature fusion layer comprising a plurality of hidden layers, a full-connection layer of CNN and neurons in a hidden layer of the last layer of DNN are fully connected with neurons in a hidden layer of the first layer of the feature fusion layer, and fault features extracted from unstructured data by CNN and fault features extracted from structured data by DNN are seamlessly integrated through full-connection operation;
a. assuming that the output dimension of the CNN full connection layer is NC, the feature vector Vc epsilon R 1×Nc (Vc belongs to a space in dimension 1 XNc); if the output dimension of the hidden layer of the last DNN layer is Nd, the feature vector Vd epsilon R 1×Nd (Vd belongs to a space of 1 XNd dimension); the feature vector Vin e R constructed after the full connection operation 1×(Nc+Nd) (Vin belongs to a space of 1× (nc+nd) dimension);
b. taking Vin as the input of a first hidden layer of a feature fusion layer, and carrying out fusion mapping on a feature vector Vin in a plurality of hidden layers of the feature fusion layer;
3. and finally, inputting the output of the last hidden layer of the feature fusion layer into a softmax classifier to classify faults.
The invention further provides an algorithm embodiment, which aims to automatically extract hidden representative characteristics from different data sets of different working conditions, establish a general fault diagnosis model, and enable the model to realize fault diagnosis of an abs system on the premise of changing the working conditions, such as split road condition simulation of an indoor test bench, automatic determination of vehicle-mounted real-time measurement data under a complex road surface and the like; the method comprises the following steps:
1. acquiring ABS brake performance detection data, analyzing an ABS detection result, and selecting m characteristic parameters capable of reflecting the ABS working performance of the automobile; for example: the slip rate, the adhesion coefficient utilization rate and the like are selected as judging indexes of the ABS detection result;
a. the slip rate represents the difference degree between the speed and the wheel speed of the automobile in the braking process; the calculation formula is as follows:
wherein S represents slip ratio, V represents vehicle speed, r represents wheel radius, and ω represents wheel angular velocity; the control principle of the automobile abs to the braking process is that the automobile is kept near the optimal slip rate during the braking process so as to obtain larger ground adhesion force;
b. adhesion coefficient utilization rate: the road adhesion coefficient refers to the ratio of adhesion to the normal pressure of the wheel, which can be regarded as the static friction coefficient between the wheel and the road, the greater this coefficient, the greater the adhesion available, the less likely the wheel will slip. The utilization rate of the attachment coefficient refers to the effective utilization degree of the maximum adhesion force of the whole vehicle to the ground in the braking process, and the ABS control braking efficiency is embodied on a certain road surface with a certain attachment coefficient;
2. making sample data with labels;
a. determining a label parameter; for example, the ABS overall state is evaluated, with (0, 1) representing a normal state, with (1, 0) representing a failure state, and with (0, 0) representing a state of insufficient braking force;
b. adding labels Ys to sample data sets Xs under different working conditions and marking the labels as Source Domain (SD) for short (see the table below);
table 4: source field sample data
c. The method comprises the steps of recording unlabeled sample data Xt acquired in real time as a target field (TD for short);
d. through the steps, obtaining the data input Xs of the source field SD, the label output Ys thereof and the data input Xt of the target field TD; let xs=xt, ys=yt; and the edge distribution of the source data and the target data is different (i.e., P (Xs) noteqp (Xt)), and the condition distribution is different (i.e., Q (ys|xs) noteq (yt|xt)); the problem to be solved is that a classifier trained by using source data predicts the label output Yt of the target data input Xt;
3. performing data cleaning processing on n sample data (n=ns+nt) of the characteristic parameters;
a. performing linear normalization processing on the non-binarized sample data to enable the non-binarized sample data to be distributed between 0 and 1; the calculation formula is as follows: x' = (X-Xmin)/(Xmax-Xmin);
wherein X represents a certain sample value in a certain data attribute, xmax represents a maximum value in the certain data attribute, xmin represents a minimum value in the certain data attribute, and X' represents a value after normalization processing;
b. removing noise contained in the normalized data through Fourier transformation;
4. constructing an unsupervised deep learning network model, as shown in fig. 5;
a. the unsupervised deep learning network model is structurally the same as a traditional multi-layer neural network, and comprises an input layer, a plurality of hidden layers and an output layer: the neurons of each layer are not connected, and the layers are fully connected; for the abs braking system general fault diagnosis model, the node numbers of the input layer and the output layer respectively correspond to the input attribute and the category number of the data set;
the front part of the hidden layer is formed by stacking a plurality of layers of Automatic Encoders (AE), and a layer-by-layer training method is adopted during training, namely, the output of the previous layer is used as the input of the next layer to train sequentially; after each layer of automatic encoder extracts the characteristics, common information between the source domain and the target domain hidden under the data set is extracted through edge distribution adaptation and condition distribution adaptation, so that characteristic representation which remarkably reduces the difference between the edge distribution and the condition distribution of the source domain and the target domain is obtained;
the last layer of the hidden layer is a classification layer representing expected output variables, preferably a Softmax classifier suitable for nonlinear multi-classification problems, wherein the output of the Softmax classifier is probability values of corresponding samples belonging to different label states respectively, and the state with the maximum probability value is the final diagnosis result;
the training process is divided into two stages of pre-training and fine tuning, wherein the pre-training adopts sample data of a source field SD or a target field TD as input of a network, and the initialization of AE parameters of a plurality of layers at the front part is completed through a BP algorithm; when the AE parameters of each layer are pre-trained, the parameters of other layers are fixed and kept unchanged, and the fine tuning is to adopt the source data with labels
The BP algorithm is used for simultaneously adjusting the parameters of the whole network including the classification layer, so that the discrimination performance of the network is optimal;
b. the automatic encoder AE comprises an output layer, an implicit layer and an input layer, wherein the output layer and the input layer have the same scale; the feature transformation process from the input layer to the hidden layer is called encoding, and the feature transformation process from the hidden layer to the output layer is called decoding;
the coding function is defined as f (x) =s f (wx+p), the decoding function is defined as g (h) =s g (W T h+q), wherein: s is S f 、S g Preferably a sigmoid function, W represents a weight matrix between the input layer and the hidden layer, W T Representing a weight matrix between the hidden layer and the output layer; p represents the bias vector of the hidden layer; q represents the offset vector of the output layer; the AE parameters are marked as θ;
let s= { X be the training sample set 1 ，…，X n The process of training AE is essentially a process of training the parameter θ with S; the specific method comprises the following steps:
J AE+sp representing sparse self-coding, L (x, y) is a reconstruction error function; beta is a weight coefficient for controlling the sparsity penalty term; ρ is a sparsity parameter;
the abs braking system general fault diagnosis model operates under different working conditions, the edge distribution of source data and target data is inconsistent, and the distance between the edge distribution of a source domain and a target domain needs to be further shortened, and the specific method comprises the following steps:
J M representing edge distribution adaptation, X S ，X T Representing feature representations from source and target domains; learning feature transformation matrix a by using edge distribution adaptation to obtain new feature representation z=a T X；
Training of a general fault diagnosis model of an abs braking system, wherein a classifier trained by source data is required to be used for predicting a label of a target source, the difference of conditional distribution between the source data and the target data is considered, the class conditional probability is required to be minimized to reach the target of minimizing the conditional probability, and the maximum average difference (maximize mean discrepancy, abbreviated as MMD) is still used for shortening the distance between the conditional distribution of a source domain and the conditional distribution of a target domain, and the specific method is as follows:
J C representing condition distribution adaptation, X S ，X T Representing feature representations from source and target domains, C is a class conditional probability, Q (X S |Y S ＝C)、Q(X T |Y T =c), C e {1,2 …, n }; learning the feature transformation matrix a by using conditional distribution adaptation to obtain a new feature representation z=a T X；
e. Under the condition that the edge distribution and the condition distribution of the source data and the target data are large in difference, the classifier trained by the source data is used for predicting the output Yt of the target data, and MMD distances of the edge distribution and all kinds of condition distribution are required to be added for optimization, specifically:
f. The BP algorithm repeatedly iterates the following specific steps:
through pre-training, each layer of the model is regarded as a self-coding network, and the input data is continuously coded and decoded until the output layer of the depth model. Meanwhile, calculating a training sample error by continuously utilizing a back propagation algorithm, and optimizing a loss function in each layer according to the gradient of the error to obtain an optimized weight and an optimized bias parameter, specifically;
wherein W is the weight before adjustment, W' is the weight after adjustment, E is the error, and eta is the learning rate;
calculating error variation sigma of two successive iterations, and stopping the iteration process of the back propagation algorithm when sigma is more than or equal to 0 and less than or equal to H;
finally, back propagation is utilized again to transfer the error between the expected output and the actual output of the system to each layer, so that the overall model parameters are optimized;
g. the Softmax classifier is trained through the feature vectors output by the front multi-layer automatic encoder; assuming that there are k classification categories in total, the output of the Softmax classifier is a first-order probability matrix, and a probability value p is estimated for class labels from 1 to k, and the system equation is as follows:
each row of the matrix is a parameter of a class label corresponding to the classifier, and the total k rows are represented as a loss function:
where l {.cndot } is an indicator function, i.e., when the value in brackets is true, the function value is 1, otherwise it is 0.
The partial derivative function of the loss function with respect to the parameter θ is as follows:
and obtaining the parameter value of the system by using a gradient descent method according to the training sample, the loss function and the partial derivative function.
5. The construction steps of the general fault diagnosis model for the abs braking system are as follows;
a. pre-training an algorithm model by using a label sample in the source field to obtain a weight and a bias parameter of the model;
b. respectively selecting the same number of sample data in the source field and the target field as input, and adjusting the weight and the bias parameters of the model again to obtain corresponding characteristic representations of the data;
c. the feature representation of the source domain is used to train a Softmax classifier, resulting in a trained classification model,
d. taking the characteristic representation of the target field as the input of a Softmax classifier to obtain a classification label of each sample, thereby obtaining an automatic judging algorithm model of the automobile abs braking performance;
next, an example of ABS detection parameters and detection methods thereof that can be employed in the present invention will be described.
(example 1) description is given of a characteristic determination index using a deceleration ratio as an ABS detection result.
Fig. 5 schematically shows a vehicle speed/wheel speed comparison curve at ABS braking. The deceleration ratio refers to the ratio of the deceleration of the wheels to the deceleration of the vehicle body during braking, and under the normal working condition of the ABS (fig. 5 (a)), the magnitude of the deceleration of the wheels is basically consistent with that of the deceleration of the vehicle body, and under the failure condition of the ABS (fig. 5 (b)), the wheels are rapidly locked, the wheel speed is reduced to zero in a short time, the speed of the vehicle body is reduced relatively slowly, and a large difference is formed between the two decelerations. Thus, the larger the deceleration ratio, the worse the ABS turndown is explained;
(example 2) a characteristic determination index using the braking deceleration as the ABS detection result is described.
Fig. 6 schematically shows a comparison curve of vehicle speed/wheel speed at ABS braking under the same road surface with an adhesion coefficient, and the braking deceleration reflects the rate of decrease in vehicle speed at braking, and thus the braking deceleration reflects the braking performance of the braking system to some extent. If the deceleration value is small on the road surface of the same attachment coefficient, a situation may occur in which the braking force is insufficient as shown in fig. 6 (b) (in the case where fig. 6 (b) is smaller in braking initial speed than fig. 6 (a), the braking distance is 8 times that of fig. 6 (a));
(example 3) illustrates a method and example of reconstructing raw data of ABS braking by a deep learning algorithm.
Fig. 7 schematically shows a flow of the dimension reduction processing of the raw data of the ABS brake by the automatic encoder AE.
The characteristic parameters reflecting the ABS working performance are more, and the ABS working performance is to be evaluated according to all the parameters, so that the training of the neural network is more complicated, and an accurate conclusion cannot be obtained due to multiple correlations among the characteristic parameters. Therefore, for a general fault diagnosis model of a desired abs braking system, the automatic encoder AE can be adopted to perform dimension reduction processing on the data source, and the method has the following beneficial effects:
the automatic encoder AE can automatically learn features from unmarked data, is a neural network aiming at reconstructing input information, has stronger feature learning capability, and can give better feature description than original data;
compared to the conventional principal component analysis (principal component analysis, PCA) algorithm, the automatic encoder AE can characterize both linear and nonlinear transformations and can be used as a layer for constructing a deep learning network. By setting appropriate dimensions and sparse constraints, the self-encoder AE can learn more meaningful data projections than with PCA and other techniques;
(example 4) an embodiment of judging abs performance by vehicle speed and wheel speed is described.
1. Raw data preparation
1.1, first the preparation part of the data.
In the actual operation of this embodiment, we choose the vehicle body speed and the wheel speed as the study object. The method takes the brake initial speed of 40km/h as a reference, selects 1 point every 10 milliseconds for data sampling frequency, 30 seconds for sampling time, and 1 point every 100 milliseconds for the last 20 seconds, and is mainly used for detecting insufficient braking force.
1.2 test conditions
According to the requirement of GB/T13594-2003 on an ABS road test, ABS performance detection mainly comprises working conditions such as a road surface with a high adhesion coefficient, a road surface with a low adhesion coefficient, a split road surface, a butt joint road surface and the like.
1.3 tag data
Referring to fig. 5 and 6, evaluation tags for the ABS overall state, such as a normal state, a failure state, and a braking force shortage, are added to the sample data of the source field.
2. Deep learning model training
2.1, establishing an abs braking performance automatic judging algorithm model
And establishing a deep learning neural network consisting of a plurality of layers of automatic encoders, and adopting a Softmax regression algorithm as a top classifier to obtain the result of automatically judging the hidden layer of abs braking performance.
The number of hidden layers is two, corresponding to 1200 input layers of neurons, the first hidden layer has 50 neurons, and the second hidden layer has 25 neurons. And taking the output of the first hidden layer as the input of the second hidden layer, and obtaining the final reconstructed base vector through layer-by-layer learning of the data.
The output layer is set to 2 neurons according to the number of classifications. The normal state is represented by (0, 1), the failure state is represented by (1, 0), and the braking force shortage state is represented by (0, 0).
2.2 occlusion handling of data
In order to enhance the robustness of the algorithm model, the input data needs to be subjected to ' shielding ' processing, wherein the shielding proportion is 0.25, namely, 25% of neuron node data of the input layer are set to 0 in uniform distribution (qD distribution), so that a new input sample x ' is obtained; and then, the x' after shielding treatment is used as the input of a model, the sample x before shielding is expressed through deep learning of a multi-layer neural network, and the reconstruction of the original data is realized, so that the characteristic self-expression with high robustness is realized, and the noise resistance of an algorithm is improved.
2.3 fine tuning
The unsupervised self-learning process lacks enough labels, so that the classification accuracy is not high, and the classification accuracy requirement of performance judgment cannot be met. The fine tuning process adopts the source data with the labels, adjusts the whole network parameters including the classification layer based on the back propagation algorithm, reduces the output residual error of the neurons, and further optimizes the weight and the bias parameters of the network model, so that the discrimination precision of the network reaches more than 99%.
As described above, several examples of ABS detection parameters and detection methods thereof that can be employed in the present invention are described, but the ABS detection parameters and detection methods thereof that can be employed in the present invention are not limited to the above examples, and a designer may freely design according to the type of vehicle, ABS operation model, indoor detection platform, measurement and control system, and the structure type of the related simulation mechanism.
It is easy to find that compared with the traditional ABS detection method, the method utilizes the cloud database to accumulate real vehicle measurement and corresponding indoor whole vehicle test data, and can more comprehensively detect the braking execution conditions of vehicles assembled with ABS on various pavements, thereby ensuring the safety of vehicle braking, reducing traffic accidents and having higher practical value.
The embodiments of the present invention described above do not limit the scope of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of the present invention as set forth in the appended claims.
Claims (6)
1. An abs result automatic judging algorithm is characterized in that: comprises the steps of,
s1: acquiring measurement data and additional information related to the detection of the ABS braking performance of the automobile in the braking process;
s2, carrying out data preprocessing on the acquired measurement data and the additional information;
s3: making sample data with labels;
s4: training an unsupervised deep learning network model;
the unsupervised deep learning network model in the step S4 includes a primary network and a secondary network; the main network is provided with a feedforward connection structure and a feedback connection structure and has local memory capacity, and the main network is sequentially composed of an input layer, a bearing layer, an intermediate layer and an output layer; the secondary network is used for initializing the primary network and comprises a visual layer and an implicit layer;
the training process in the step S4 is divided into two stages: firstly, adopting a bottom-up unsupervised training mode; secondly, adopting a top-down supervised learning mode; the bottom-up training is layered training with no calibration data or with calibration data; firstly, inputting training samples, learning weights of a first layer of a network until a model n-1 layer, wherein the output of the training samples is used as the model n-layer input, so that neuron parameters of each layer are obtained, and the training process is unsupervised; the top-down learning is to further train the network through the data with the labels after the first training process is finished, so that errors are transmitted layer by layer from top to bottom, fine adjustment is carried out on the network parameters obtained by pre-training, and finally, network parameters of each layer of the model are determined, wherein the training process is supervised;
aiming at the vehicle type and or ABS braking system which accumulate abundant fault data, under the condition of limited or lacking training sample label information, fault characteristics are comprehensively and effectively extracted from measured vehicle measurement data and additional information of different vehicle types, vehicle numbers, vehicle ages, road positions, numerical values, characters and pictures, and the measured vehicle measurement data and the additional information with multiple structures, so that the method is suitable for automatic judgment of the ABS working state of the vehicle under the condition of class label deficiency; the method comprises the following specific steps: A. unstructured data is used as input of a convolutional neural network, and structured data is used as input of a deep neural network; B. the method comprises the steps that through a feature fusion layer comprising a plurality of hidden layers, a full-connection layer of CNN and neurons in a hidden layer of the last layer of DNN are fully connected with neurons in a hidden layer of the first layer of the feature fusion layer, and fault features extracted from unstructured data by CNN and fault features extracted from structured data by DNN are seamlessly integrated through full-connection operation; C. and inputting the output of the last hidden layer of the feature fusion layer into a softmax classifier to classify faults.
2. An abs outcome automatic determination algorithm according to claim 1 wherein: the measured data in the step S1 comprise data of vehicle speed, wheel speed, pedal force, abs braking time, interval and brake pipeline pressure; the additional information comprises vehicle type, vehicle number, vehicle age, road position and vehicle type technical parameters.
3. An abs outcome automatic determination algorithm according to claim 2 wherein: the step S1 further comprises dividing the acquired measurement data and additional information into structured data and unstructured data; the structured data comprises numerical data and a database; the unstructured data includes text-type data or time domain waveform diagrams.
4. An abs outcome automatic determination algorithm according to claim 1 wherein: the step S2 specifically comprises the following steps of A1: intercepting data of the length related to abs braking action in the waveform time curve of the measured data as original waveform data; a2: removing baseline drift noise by adopting a high-pass filter; a3: confirming whether the noise is too high or not based on a standard variance and a threshold method, and removing noise interference by using a low-pass Butterworth filter when the noise is too high; a4: and carrying out normalization processing on all data according to the characteristics of multi-source and heterogeneous data sources.
5. An abs outcome automatic determination algorithm according to claim 1 wherein: in step S3, the measured data of the Abs braking system running under different working conditions have a certain similarity, and the time domain distribution map of the fault data has a certain similarity, that is, the source field and the target field have a common part.
6. The abs outcome automatic determination algorithm of claim 5 wherein: in the step S3, adding tags to the processed data under different working conditions and marking the data as a source field; the evaluation conclusion of the integral state of the ABS is that the label comprises normal, fault and deficiency; the preprocessed data acquired in real time is recorded as the target field.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110161343.6A CN112985830B (en) | 2021-02-05 | 2021-02-05 | Automatic abs result judging algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110161343.6A CN112985830B (en) | 2021-02-05 | 2021-02-05 | Automatic abs result judging algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112985830A CN112985830A (en) | 2021-06-18 |
| CN112985830B true CN112985830B (en) | 2023-06-23 |
Family
ID=76348107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110161343.6A Active CN112985830B (en) | 2021-02-05 | 2021-02-05 | Automatic abs result judging algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112985830B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113188807B (en) * | 2021-02-05 | 2024-05-03 | 深圳大雷汽车检测股份有限公司 | Automatic abs result judging algorithm |
| CN113156913B (en) * | 2021-02-05 | 2024-08-06 | 深圳大雷汽车检测股份有限公司 | ABS fault diagnosis system and method |
| CN114155476B (en) * | 2022-02-07 | 2022-07-05 | 天津所托瑞安汽车科技有限公司 | AEB (automatic Emergency bank) accident scene identification method, device, equipment and medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108956153A (en) * | 2018-04-27 | 2018-12-07 | 东华大学 | A kind of automobile anti-lock braking detection method based on RBF radial base neural net |
| CN110751633A (en) * | 2019-10-11 | 2020-02-04 | 上海眼控科技股份有限公司 | Multi-axis cart braking detection method, device and system based on deep learning |
| CN111238825A (en) * | 2020-01-10 | 2020-06-05 | 东南大学 | Intelligent driving automatic emergency braking performance testing method for combined test pavement |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001344590A (en) * | 2000-05-31 | 2001-12-14 | Fuji Electric Co Ltd | Neural network and method for learning the same and method for analyzing the same and method for judging abnormality |
| CN106482938B (en) * | 2016-10-14 | 2018-08-03 | 温州大学 | Brake fluid system multi-source fusion failure predicting method based on GA-BP networks |
| CN109000935B (en) * | 2018-07-12 | 2020-07-28 | 清华大学深圳研究生院 | Method for judging performance of new energy automobile brake system |
| JP2022523730A (en) * | 2019-01-30 | 2022-04-26 | パーセプティブ オートマタ インコーポレイテッド | Neural network-based navigation of autonomous vehicles sewn between traffic entities |
| CN111814966A (en) * | 2020-08-24 | 2020-10-23 | 国网浙江省电力有限公司 | Neural network architecture searching method, neural network application method, device and storage medium |
-
2021
- 2021-02-05 CN CN202110161343.6A patent/CN112985830B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108956153A (en) * | 2018-04-27 | 2018-12-07 | 东华大学 | A kind of automobile anti-lock braking detection method based on RBF radial base neural net |
| CN110751633A (en) * | 2019-10-11 | 2020-02-04 | 上海眼控科技股份有限公司 | Multi-axis cart braking detection method, device and system based on deep learning |
| CN111238825A (en) * | 2020-01-10 | 2020-06-05 | 东南大学 | Intelligent driving automatic emergency braking performance testing method for combined test pavement |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112985830A (en) | 2021-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112985830B (en) | Automatic abs result judging algorithm | |
| CN108960426B (en) | Road slope comprehensive estimation system based on BP neural network | |
| CN113188807B (en) | Automatic abs result judging algorithm | |
| Li et al. | A deep transfer nonnegativity-constraint sparse autoencoder for rolling bearing fault diagnosis with few labeled data | |
| CN111915059B (en) | Attention mechanism-based Seq2Seq berth occupancy prediction method | |
| Liang et al. | Rolling bearing fault diagnosis based on one-dimensional dilated convolution network with residual connection | |
| CN111325112A (en) | Cutter wear state monitoring method based on depth gate control circulation unit neural network | |
| WO2019037557A1 (en) | Method for learning time sequence characteristics of locomotive operation | |
| CN111080597A (en) | Track fastener defect identification algorithm based on deep learning | |
| CN113239720B (en) | Subway vehicle running gear fault diagnosis method based on deep migration learning | |
| CN112000084B (en) | Intelligent BIT design method of controller module based on 1D-CNN and GRU-SVM | |
| CN108491859A (en) | The recognition methods of driving behavior heterogeneity feature based on automatic coding machine | |
| CN108734958B (en) | Traffic speed prediction method | |
| CN112966853A (en) | Urban road network short-term traffic flow prediction method based on space-time residual error mixed model | |
| CN112381027B (en) | Wheel polygon wave depth estimation method based on train axle box vertical acceleration signal | |
| CN115204035A (en) | Generator set operation parameter prediction method and device based on multi-scale time sequence data fusion model and storage medium | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| CN114626625B (en) | Bearing life prediction method based on multi-module U-BiLSTM network | |
| CN115186594A (en) | Energy-saving speed optimization method under influence of man-vehicle-road coupling | |
| CN115774942A (en) | Driving style identification model modeling and statistical method based on Internet of vehicles real vehicle data and SVM | |
| CN117312962A (en) | Long-service elevator guide rail fault diagnosis method based on transfer learning and data driving | |
| CN113156913B (en) | ABS fault diagnosis system and method | |
| CN114357372A (en) | Aircraft fault diagnosis model generation method based on multi-sensor data driving | |
| CN118312879A (en) | Proportion servo valve fault diagnosis method based on attention convolution capsule network | |
| CN111863153A (en) | Method for predicting total amount of suspended solids in wastewater 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 |