CN113821875B - Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation - Google Patents
Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation Download PDFInfo
- Publication number
- CN113821875B CN113821875B CN202111128029.4A CN202111128029A CN113821875B CN 113821875 B CN113821875 B CN 113821875B CN 202111128029 A CN202111128029 A CN 202111128029A CN 113821875 B CN113821875 B CN 113821875B
- Authority
- CN
- China
- Prior art keywords
- prediction model
- formula
- vehicle
- cnn
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G07—CHECKING-DEVICES
- G07C—TIME OR ATTENDANCE REGISTERS; REGISTERING OR INDICATING THE WORKING OF MACHINES; GENERATING RANDOM NUMBERS; VOTING OR LOTTERY APPARATUS; ARRANGEMENTS, SYSTEMS OR APPARATUS FOR CHECKING NOT PROVIDED FOR ELSEWHERE
- G07C5/00—Registering or indicating the working of vehicles
- G07C5/08—Registering or indicating performance data other than driving, working, idle, or waiting time, with or without registering driving, working, idle or waiting time
- G07C5/0808—Diagnosing performance data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
-
- 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
- Y02T10/00—Road transport of goods or passengers
- Y02T10/60—Other road transportation technologies with climate change mitigation effect
- Y02T10/70—Energy storage systems for electromobility, e.g. batteries
Abstract
The invention discloses an intelligent vehicle fault real-time prediction method and system based on end cloud cooperation, wherein the method comprises the following steps: step 1, collecting the original data of n systems of the intelligent vehicle in the running period; step 2, performing real-time prediction training through a CNN-LSTM prediction model deployed at a cloud end, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end; step 3, obtaining the predicted residual life of a kth system in real time by a vehicle end through a trained single system residual life prediction model, wherein k= … …, n; and 4, according to the predicted residual service lives of the k systems, obtaining the residual service life of the whole vehicle through a residual service life prediction model of the serial-parallel system. The method has the advantages of strong real-time performance, high precision, good applicability, continuous iteration upgrading and the like, and solves the problems that the traditional method only aims at the prediction of a single system or a single component of a vehicle, the prediction of the vehicle fault is not timely and accurate, and the like.
Description
Technical Field
The invention relates to the technical field of intelligent vehicle fault diagnosis, in particular to an intelligent vehicle fault real-time prediction method and system based on end cloud cooperation.
Background
With the continuous increase of the storage quantity of new energy vehicles and the improvement of the intelligent degree requirements of vehicles, the fault diagnosis of intelligent vehicles becomes a research hotspot in the field of data analysis. The intelligent vehicle integrates the functions of environment sensing, planning decision, early warning diagnosis and the like, so that vehicle fault data is explosively increased, and challenges are brought to a traditional manual fault diagnosis mode. The growing popularity of fifth generation communication systems (5G) makes it possible to predict the remaining useful life of a vehicle in real time. It can predict before the failure of the component or system occurs, avoiding further deterioration of the failure and thus avoiding more serious safety accidents.
Model-based residual life (Remaining Useful Life, RUL) prediction techniques often require the determination of accurate physical or mathematical models to describe the system degradation process. However, it is difficult to build accurate models for complex high-order nonlinear systems of intelligent vehicles, while being less robust to noise. The failure prediction method based on the experience degradation model is a degradation mechanism model summarized in the long-term use process of technicians, and has no popularization. Based on a mathematical statistics theory, equipment degradation data is processed by a common principal component analysis or partial least square method, statistics is established, and equipment health state evaluation is carried out.
Current methods of residual life prediction based on deep learning are focused mainly on individual components or systems of the engine, crankshaft and battery in the vehicle field. However, a vehicle is a complex electromechanical system and the life curves of individual components or subsystems do not accurately reflect the performance and life of the entire vehicle. And for intelligent network-connected vehicles with more complex software, the residual service life of the intelligent network-connected vehicles is more difficult to predict.
The traditional fault prediction is deployed on the intelligent vehicle, and the fault prediction based on data driving has the defects that the intelligent vehicle is small in calculation power, a complex algorithm model is difficult to deploy, and the calculation speed is low, so that time lag is high. Meanwhile, if the intelligent vehicle itself fails to predict the controller hardware damage, the failure prediction system will fail completely. In addition, the scheme of deploying the fault prediction on the cloud server can solve the problems of small calculation power, poor data storage capacity and the like of the vehicle, but is subject to the problems of unstable communication state of the vehicle and the cloud, large data transmission quantity of the whole vehicle fault prediction and the like, and cannot complete real-time fault prediction.
Disclosure of Invention
The invention aims to provide an intelligent vehicle fault real-time prediction method and system based on end cloud cooperation, which have the advantages of strong real-time performance, high precision, good applicability, continuous iteration upgrading and the like, and solve the problems that the traditional method only aims at the prediction of a single system or a single component of a vehicle, and the prediction of the vehicle fault is not timely and accurate and the like.
In order to achieve the above purpose, the invention provides an intelligent vehicle fault real-time prediction method based on end cloud cooperation, which comprises the following steps:
step 1, collecting the original data of n systems of the intelligent vehicle in the running period;
step 2, performing real-time prediction training through a CNN-LSTM prediction model deployed at a cloud end, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end;
step 3, obtaining the predicted residual life of a kth system in real time by a vehicle end through a trained single system residual life prediction model, wherein k= … …, n;
and 4, according to the predicted residual service lives of the k systems, obtaining the residual service life of the whole vehicle through a residual service life prediction model of the serial-parallel system.
Further, the raw data of the step 1 includes electronic control system related data, battery system related data and motor system related data, and the residual life prediction model of the serial-parallel system of the step 4 is described as formula (17):
f＝min{(α*max{X,Y}),βZ} (17)
wherein f represents the residual service life of the whole vehicle, alpha and beta are weights obtained through continuous iterative updating in the step 2, X represents the predicted residual life of the electronic control system, Y represents the predicted residual life of the battery system, and Z represents the predicted residual life of the motor system.
Further, the method for continuously and iteratively updating the obtained weight in the step 2 specifically includes:
judging whether the current evaluation index is better than the evaluation index obtained by the single-system residual service life prediction model trained last time, if not, updating the weight used by the single-system residual service life prediction model trained at present; if so, step 4 is entered.
Further, the evaluation index includes a score value fire calculated by a sum formula (4) of root mean square error RMSE calculated by formula (2):
in the formula delta i Representing the deviation between the calculated true value and the predicted value at the moment, N is the total number of test samples, s i Represented by formula (5):
further, the method for obtaining the corresponding single-system residual service life prediction model through the CNN-LSTM prediction model in the step 2 specifically comprises the following steps:
step 21, inputting original data described by a feature map of n-dimensional feature dimensions;
step 22, extracting features of the original data sequence through a filter, activating the original data sequence through a Relu function, and outputting a first feature matrix C;
step 23, compressing the number of the extracted parameters of C through a maximum pooling function, and outputting a second feature matrix P;
step 24, according to the preset probability, randomly discarding the connection of part of neurons in P by a Dropout method, and outputting an influence factor characteristic sequence D;
step 25, inputting D into a CNN-LSTM prediction model for prediction, and finally converting the data into a one-dimensional structure through a full connection layer to obtain the predicted residual life of the kth system, namely: single system remaining life prediction model.
The invention also provides an intelligent vehicle fault real-time prediction system based on end cloud cooperation, which comprises:
the information acquisition unit is used for acquiring the original data of n systems of the intelligent vehicle during operation;
the fault diagnosis cloud service unit is used for carrying out real-time prediction training through a CNN-LSTM prediction model deployed at the cloud, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end;
the vehicle-mounted controller is used for preprocessing original data, obtaining the predicted residual life of a kth system through a trained single-system residual life prediction model, and obtaining the residual life of the whole vehicle through a serial-parallel system residual life prediction model according to the predicted residual life corresponding to each of n systems, wherein k=1 … … and n.
Further, the raw data includes electronic control system-related data, battery system-related data, and motor system-related data, and the series-parallel system remaining life prediction model is described as formula (17):
f＝min{(α*max{X,Y}),βZ} (17)
wherein f represents the residual service life of the whole vehicle, alpha and beta are weights obtained by continuous iterative updating of the fault diagnosis cloud service unit, X represents the predicted residual life of the electronic control system, Y represents the predicted residual life of the battery system, and Z represents the predicted residual life of the motor system.
Further, the method for continuously and iteratively updating the obtained weight by the fault diagnosis cloud service unit specifically comprises the following steps:
judging whether the current evaluation index is better than the evaluation index obtained by the single-system residual service life prediction model trained last time, if not, updating the weight of the single-system residual service life prediction model trained at present; if so, using a single system residual service life prediction model which is trained currently.
The data of the electronic control system, the battery system and the motor system are transmitted to the cloud end in real time through the communication network (WiFi, 5G) of the intelligent vehicle, and serve as the original data of intelligent vehicle fault prediction. Modeling is carried out by adopting a CNN-LSTM model, and life curves of all series-parallel systems are fused, so that the life curve of the whole vehicle can be obtained. The fault prediction model is deployed in the cloud with high computing capability for training, and the trained model is deployed to the vehicle end for predicting faults in real time.
Drawings
Fig. 1 and fig. 2 are schematic frame structures of an intelligent vehicle fault real-time prediction system based on end cloud cooperation according to an embodiment of the present invention.
Fig. 3 is a training flow chart of the CNN-LSTM model in fig. 1 and 2.
Detailed Description
In the drawings, the same or similar reference numerals are used to denote the same or similar elements or elements having the same or similar functions. Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
As shown in fig. 1 and fig. 2, the method for predicting the fault of the intelligent vehicle based on the end cloud cooperation in real time provided by the embodiment of the invention includes:
and 1, collecting the original data of n systems of the intelligent vehicle during operation. Preferably, the original data is subjected to preprocessing such as deletion processing, text data quantization processing and data normalization processing to obtain an original data sequence.
And 2, performing real-time prediction training through a CNN-LSTM prediction model deployed at the cloud according to the original data, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end.
And 3, obtaining the predicted residual life of the kth system by the vehicle end through a trained single-system residual life prediction model, wherein k=1 … … and n.
And 4, according to the predicted residual service lives of the n systems, obtaining the residual service life of the whole vehicle through a residual service life prediction model of the series-parallel system.
According to the method, the original data of different systems of the intelligent vehicle are preprocessed through the vehicle-mounted controller of the intelligent vehicle, and are transmitted to the cloud end in real time through the communication network (WiFi and 5G) to serve as the original data of intelligent vehicle fault prediction. Modeling is carried out by adopting a CNN-LSTM prediction model, and the life curves of the series-parallel systems are fused by using a residual life prediction model of the series-parallel systems, so that the life curve of the whole vehicle can be obtained. The CNN-LSTM prediction model is deployed on a cloud with high computing capacity, and the trained single-system residual service life prediction model is deployed on a vehicle end to predict faults in real time. Compared with the traditional model, the CNN-LSTM prediction model has the advantages of strong real-time performance, good applicability, high precision, continuous iteration upgrade and the like.
In one embodiment, the remaining life prediction model of the series-parallel system of step 4 is described as formula (1):
wherein T represents the residual service life of the whole vehicle, T k Representing the predicted remaining life, ω, of the kth system obtained in step 3 k Representing the weight of the kth system obtained via step 2 continuous iterative updating.
In one embodiment, the method for continuously and iteratively updating the obtained weight in step 2 specifically includes:
judging whether the current evaluation index is better than the evaluation index obtained by the single-system residual service life prediction model trained last time, if not, updating the weight used by the single-system residual service life prediction model trained at present; if so, step 4 is entered.
In one embodiment, the evaluation index comprises a score value fire calculated by the sum of root mean square errors RMSE calculated by equation (2) and calculated by equation (4):
Δ i ＝RUL predicted -RUL actual (3)
in the formula delta i Representing the calculated real value RUL at the i-th moment actual And the predicted value RUL predicted Deviation between, N is the total number of test samples, s i Represented by formula (5):
in the above embodiment, the smaller the value of the root mean square error RMSE, the better the value, and the smaller the value of the score value fire, the better the value.
As shown in fig. 3, the method for obtaining the corresponding single-system residual service life prediction model by using the CNN-LSTM prediction model in step 2 specifically includes:
step 21, inputting the original data described by the feature map of the n-dimensional feature dimension.
In step 22, the original data sequence is subjected to feature extraction by a filter (filter), and then activated by a Relu function, and a first feature matrix C is output.
And step 23, compressing the number of the extracted parameters of C through a maximum pooling function, and outputting a second feature matrix P. Thus, the dimension of the data can be further reduced to extract obvious characteristics, the operation efficiency is improved, and the overfitting is prevented
Step 24, according to the preset probability, the connection of part of neurons in P is randomly discarded by using the equation (4) through a Dropout method, so that the phenomenon of overfitting can be avoided as much as possible, and the influence factor characteristic sequence D is output. The preset probability can be an empirical value, and the value is 0.2 for improving the real-time performance of fault prediction.
And step 25, taking the D and the historical data sequence as training sequences, inputting the training sequences into a CNN-LSTM prediction model, and finally converting the data into a one-dimensional structure through a full connection layer to obtain the predicted residual life of the kth system.
In one embodiment, the raw data of step 1 includes three, electronic control system related data, battery system related data, and motor system related data, respectively. Preferably, the original data is subjected to preprocessing such as deletion processing, text data quantization processing and data normalization processing to obtain an original data sequence. The electronic control system related data comprise steering control, power driving control, braking control and CAN management control of the electronic control system. The battery system-related data has a battery pack, a number of single batteries, a single battery voltage value, a battery pack temperature, a charge-discharge state, a total voltage, a total current, a state of charge, and an insulation resistance of the battery system. The motor system-related data has a temperature, a vibration, a current, a rotational speed, and a torque of the motor system.
In step 21, the raw data or raw data sequence of the three systems are respectively described as feature graphs of n-dimensional feature dimensions, and are input into the CNN-LSTM model, as follows t As shown.
Wherein t represents single system data of the t-th time period, m represents the m-th feature, x tm Mth dimension data representing the t-th time period.
Step 22, as shown in fig. 3, a convolution operation is performed by using a convolution check feature map of 3*3 through a filter in the CNN convolution layer, so as to extract significant features in the intelligent vehicle system, where the convolution operation is shown in formula (6):
wherein X is the input of the convolution layer, corresponding to the original data sequence X, C in FIG. 3 is the output of the convolution layer, W C The weight matrix is determined by the length and width channel number of the input parameters of the convolution layer, the weight matrix is initialized before convolution, and then a new weight matrix is generated by feedback after each layer of convolution c The bias is used for better fitting function and extracting features. The method comprises the steps of obtaining in a calculation process;for convolution operation, f is the ReLU activation function. Wherein the ReLU activation function is represented by formula (7):
of course, the feature extraction may be performed using other means such as mean, variance, root mean square, skewness, kurtosis, crest factor, entropy, fourier transform, and wavelet transform.
The extracted features of the embodiment specifically include: characteristic values of steering control, power driving control, braking control and CAN management control of the electronic control system, battery packs, number of single batteries, voltage value of single batteries, temperature of the battery packs, charge and discharge state, total voltage, total current, state of charge, insulation resistance and relevant characteristics of battery faults of the vehicle battery system, and relevant characteristics of temperature, vibration, current, rotating speed and torque of the vehicle motor system.
The pooling layer is used for extracting the significant features in the intelligent vehicle system and retaining more information, and the pooling layer selects the maximum pooling mode as shown in the formula (8).
P＝max pooling(C) (8)
Wherein P is the pooling layer output; maxpooling (C) is the maximum pooling function.
The size of the model can be reduced on the one hand, the calculation speed is improved, and the robustness of the extracted features is improved by maximizing the quantity of the compression parameters of the pooling function; on the other hand, the deviation of the estimated mean value caused by the parameter error of the convolution layer can be reduced, and more texture information is reserved. Of course, instead of the maximum pooling function, an average pooling function may be used, which reduces the increase of the variance of the estimated value due to the limited neighborhood size.
For example: the input is a 4*4 feature matrix, the sliding window of the actual filter takes 2 x 2, the step length is 2, and the final output is a 2 x 2 feature matrix, and the compression is 1/4 of the original feature matrix.
In step 24, the cnn-LSTM model approximates the local features of the training data during the training process, and an overfitting phenomenon occurs. To solve this problem, a Dropout mechanism is added to the CNN-LSTM model. The Dropout mechanism may randomly discard some neurons and their connections according to a probability that the output value of the discarded neurons is set to 0, preventing the forward and backward propagation of these neurons. The Dropout mechanism can effectively reduce the dependence of the model on certain local characteristics, and avoid the occurrence of the overfitting phenomenon, so that the generalization capability of the prediction model is enhanced. For comparison of a standard neural network with a Dropout network structure, equation (9) of the Dropout method is as follows:
wherein D is an influence factor feature sequence of the output of the Dropout layer: m is a random number between 0 and 1, p is a discarding probability, and p <1 is not less than 0. For example: the value of p is 0.2, i.e., the output value of 20% neurons is randomly set to 0 during model training.
The core component of the CNN-LSTM model is a 3-gate unit structure, namely a forgetting gate f t Input gate i t Output gate o t 。
At time t, there are three inputs to the memory cell of the CNN-LSTM model: input value x of network at present moment t Output value h of memory cell at last moment t -1, cell state C at the previous moment t -1; the outputs of the memory cells of the CNN-LSTM model are two: output value h of memory cell at current time t And cell state C at the current time t 。
Forgetting door f t Is responsible for controlling whether to continue to save the long-term state C or not, and is controlled by the input value x at the moment t t And the output value h of the hidden layer at the previous moment t -1. Forgetting door f t The calculation of (2) is shown in formula (10):
f t ＝σ(W f ·[h t-1 ,x t ])+b f (10)
wherein W is f Forgetting door f for t moment t Weight matrix of b) f For the offset, σ employs a Sigmoid activation function.
Input gate i t It is decided what new information to save in the memory unit. By inputting the input x of the current state t And output h of the last state t -generating a value i between 0 and 1 in a Sigmoid function t To determine how much information the memory cell needs to retain. At the same time, a tanh layer will pass the output h of the previous state t -1 and input x of current state t To determine candidate memory state input values i t And candidate statesAs shown in the formula (11) and the formula (12):
i t ＝σ(W t ·[h t-1 ,x t ])+b i (11)
obtain input i t And a current candidate memory stateMultiplying to obtain the updated information which is really added into the memory unit. Memory cell state value C t The update of (2) is as shown in equation (13):
the output gate determines what information is to be output from the memory cell, and, like the input gate, uses a Sigmoid function to determine how much information C in the memory cell needs to be output t . His calculation is as in equation (14):
o t ＝σ(W o ·[h t-1 ,x t ]+b o ) (14)
information C of memory cell t And then with o t Multiplying and activating through a tanh layer to obtain output information h of the current LSTM block t ,h t As shown in equation (15):
h t ＝o t *tanh(C t ) (15)
the mean square error loss function Z is adopted for optimization in training, as shown in a formula (16):
wherein Z is a loss function result, n is the length of training data, the number of parameters of the sensor on the intelligent vehicle is required to be determined, d is the dimension of the sliding window, the number is determined according to a prediction result and a parameter,for the predicted lifetime of the cycle d+j times, x (d+j) is the actual lifetime.
The present embodiment uses Adam optimizer for parameter training in back propagation to update weights. The final model weights are refreshed until a predefined small penalty is reached.
For evaluating the effectiveness of the model, two commonly used evaluation indexes are adopted for evaluation, namely a root mean square error RMSE calculated by the formula (2) and a score value fire calculated by the formula (4), and are used for evaluating the effectiveness of the model.
The biggest characteristic of the scoring function (4) is that the penalty for excessively large predicted values is serious. This is desirable in real engineering applications, because in the field of carrier vehicles, etc., predicted failure times later than real failure times may lead to erroneous maintenance decisions and thus very serious losses.
In one embodiment, the predicted remaining life for three large systems of the intelligent vehicle are respectively defined as: x represents the predicted remaining life of the electronic control system, Y represents the predicted remaining life of the battery system, and Z represents the predicted remaining life of the motor system.
Calculating the residual service life f of the whole vehicle by using a residual service life prediction model of the serial-parallel system described by the formula (17):
f＝min{(α*max{X,Y}),βZ}(17)
x and Y are parallel systems, Z and X and Y are serial systems, alpha and beta are weights obtained through continuous iterative updating in the step 2, and the numerical value range of the two is (0, 1). The determination principle considers the degree of fault influence of each system. The prediction of the residual service life of the whole intelligent vehicle is promoted based on the service life of the series-parallel system. .
The intelligent vehicle fault real-time prediction system based on end cloud cooperation provided by the embodiment of the invention comprises an information acquisition unit, a fault diagnosis cloud service unit and a vehicle-mounted controller, wherein:
the information acquisition unit is used for acquiring the original data of n systems of the intelligent vehicle during operation and transmitting the original data to the vehicle-mounted controller.
The vehicle-mounted controller preprocesses the acquired data and transmits the data to the fault diagnosis cloud service unit through the vehicle communication unit.
The fault diagnosis cloud service unit is used for carrying out real-time prediction training through a CNN-LSTM prediction model deployed at the cloud, obtaining corresponding single-system residual service life prediction models corresponding to the systems, and issuing the single-system residual service life prediction models to a vehicle end.
The vehicle-mounted controller is used for preprocessing original data, obtaining predicted residual life of a kth system through a trained fault prediction model, and fusing the life of each series-parallel system according to the respective predicted residual life of n systems to obtain the residual service life of the whole vehicle, wherein k= … … and n.
In one embodiment, the series-parallel system remaining life prediction model is described as equation (1).
In one embodiment, the raw data includes electronic control system related data, battery system related data, and motor system related data, and the series-parallel system remaining life prediction model is described as equation (17).
In one embodiment, the method for continuously and iteratively updating the obtained weight by the fault diagnosis cloud service unit specifically includes:
judging whether the current evaluation index is better than the evaluation index obtained by the last trained fault prediction model, if not, updating the weight used by the current trained fault prediction model; if so, outputting the residual service life of the whole vehicle by the currently trained fault prediction model.
The fault prediction system collects data of three important systems of the intelligent vehicle by way of example, and comprises the following steps: relevant data of steering control, power driving control, braking control and CAN management control of a vehicle electronic control system, relevant data of battery packs, single batteries, single battery voltage values, battery pack temperature, charge and discharge states, total voltage, total current, charge states, insulation resistance and battery faults of a vehicle battery system, and relevant data of temperature, vibration, current, rotating speed and torque of a vehicle motor system. The system can be popularized to a fault prediction system of the whole intelligent vehicle system.
Finally, it should be pointed out that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting. Those of ordinary skill in the art will appreciate that: the technical schemes described in the foregoing embodiments may be modified or some of the technical features may be replaced equivalently; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (8)
1. An intelligent vehicle fault real-time prediction method based on end cloud cooperation is characterized by comprising the following steps:
step 1, collecting the original data of n systems of the intelligent vehicle in the running period;
step 2, performing real-time prediction training through a CNN-LSTM prediction model deployed at a cloud end, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end;
step 3, obtaining the predicted residual life of a kth system in real time by a vehicle end through a trained single system residual life prediction model, wherein k= … …, n;
step 4, according to the predicted residual service lives of the k systems, the residual service life of the whole vehicle is obtained through a residual service life prediction model of the serial-parallel system; the raw data of the step 1 includes electronic control system related data, battery system related data and motor system related data, and the residual life prediction model of the serial-parallel system of the step 4 is described as formula (17):
f＝min{(α*max{X，Y})，βZ} (17)
wherein f represents the residual service life of the whole vehicle, alpha and beta are weights obtained through continuous iterative updating in the step 2, X represents the predicted residual life of the electronic control system, Y represents the predicted residual life of the battery system, and Z represents the predicted residual life of the motor system; the method for continuously and iteratively updating the obtained weight in the step 2 specifically comprises the following steps:
judging whether the current evaluation index is better than the evaluation index obtained by the single-system residual service life prediction model trained last time, if not, updating the weight used by the single-system residual service life prediction model trained at present; if yes, enter step 4; the evaluation index includes a score value fire calculated by a sum formula (4) of root mean square error RMSE calculated by formula (2):
in the formula delta i Representing the deviation between the calculated true value and the predicted value at the i-th moment, N is the total number of test samples, s i Represented by formula (5):
the method for obtaining the corresponding single-system residual service life prediction model through the CNN-LSTM prediction model in the step 2 specifically comprises the following steps:
step 21, inputting original data described by a feature map of n-dimensional feature dimensions;
step 22, extracting features of the original data sequence through a filter, activating the original data sequence through a Relu function, and outputting a first feature matrix C;
step 23, compressing the number of the extracted parameters of C through a maximum pooling function, and outputting a second feature matrix P;
step 24, according to the preset probability, randomly discarding the connection of part of neurons in P by a Dropout method, and outputting an influence factor characteristic sequence D;
and 25, inputting the D into a CNN-LSTM prediction model for prediction, and finally converting the data into a one-dimensional structure through a full connection layer to obtain the predicted residual life of the kth system.
2. The intelligent vehicle fault real-time prediction method based on end cloud coordination according to claim 1, wherein the electronic control system related data comprises steering control, power driving control, braking control and CAN management control of the electronic control system, the battery system related data comprises battery packs, number of single batteries, voltage value of single batteries, temperature of the battery packs, charge and discharge states, total voltage, total current, charge state and insulation resistance of the battery system, and the motor system related data comprises temperature, vibration, current, rotation speed and torque of the motor system.
3. The intelligent vehicle fault real-time prediction method based on end cloud cooperation as claimed in claim 2, wherein in step 21, the raw data or raw data sequence of the three systems are respectively described as feature graphs of n-dimensional feature dimensions, and are input into a CNN-LSTM model, as follows X t Shown;
wherein t represents single system data of the t-th time period, m represents the m-th feature, x tm Mth dimension data representing a t-th period;
in step 22, a convolution operation is performed by using a convolution check feature map of 3*3 through a filter in the CNN convolution layer, so as to extract features in the intelligent vehicle system, where the extracted features specifically include: characteristic values of steering control, power driving control, braking control and CAN management control of the electronic control system, battery packs, the number of single batteries, voltage values of the single batteries, battery pack temperature, charge and discharge states, total voltage, total current, charge states, insulation resistance and relevant characteristics of battery faults of the vehicle battery system, and relevant characteristics of temperature, vibration, current, rotating speed and torque of the vehicle motor system;
the convolution operation is shown in equation (6):
wherein X is the input of the convolution layer, C is the output of the convolution layer corresponding to the original data sequence X, W C The weight matrix is determined by the length and width channel number of the input parameters of the convolution layer, and the weight matrix is subjected to convolutionInitializing, and then feeding back each layer of convolution to generate a new weight matrix, b c As the amount of the offset to be used,for convolution operation, f is a ReLU activation function, expressed as equation (7):
the pooling layer is used for extracting the remarkable characteristics in the intelligent vehicle system and reserving more information, and the pooling layer selects the maximum pooling mode as shown in a formula (8);
P＝max pooling(C) (8)
wherein P is the pooling layer output; max pooling (C) is the maximum pooling function;
in step 24, a Dropout mechanism is added to the CNN-LSTM model, and the equation (9) of the Dropout method is as follows:
wherein D is an influence factor feature sequence of the output of the Dropout layer: m is a random number between 0 and 1, p is a discarding probability, and p <1 is not less than 0.
4. The intelligent vehicle fault real-time prediction method based on end cloud cooperation as claimed in any one of claims 1 to 3, wherein the core components of the CNN-LSTM model are respectively forgetting doors f t Input gate i t Output gate o t ；
At time t, there are three inputs to the memory cell of the CNN-LSTM model: input value x of network at present moment t Output value h of memory cell at last moment t -1, cell state C at the previous moment t -1; the outputs of the memory cells of the CNN-LSTM model are two: output value h of memory cell at current time t And cell state C at the current time t ；
Forgetting door f t Is responsible for controlling whether to continue to save the long-term state C or not, and is controlled by the input value x at the moment t t And the output value h of the hidden layer at the previous moment t -1 co-determination, forget door f t The calculation of (2) is shown in formula (10):
f t ＝σ(W f ·[h t-1 ，x t ])+b f (10)
wherein W is f Weight matrix of forgetting gate ft at t moment, b f Sigmoid activation function is adopted for the offset sigma;
input gate i t By input of the current state input x t And output h of the last state t -generating a value i between 0 and 1 in a Sigmoid function t To determine how much information the memory cell needs to retain, and a tanh layer outputs h from the previous state t -1 and input x of current state t To determine candidate memory state input values i t And candidate statesAs shown in the formula (11) and the formula (12):
i t ＝σ(W t ·[h t-1 ，x t ])+b i (11)
obtain input i t And a current candidate memory stateMultiplying to obtain the updated information to be added to the memory cell, and obtaining the state value C of the memory cell t The update of (2) is as shown in equation (13):
the output gate uses a Sigmoid function to determine how many memory cells of information C need to be output t Calculation as formula (14):
o t ＝σ(W o ·[h t-1 ，x t ]+b o ) (14)
information C of memory cell t And then with o t Multiplying and activating through a tanh layer to obtain output information h of the current LSTM block t ，h t As shown in equation (15):
h t ＝o t *tanh(C t ) (15
the mean square error loss function Z is adopted for optimization in training, as shown in a formula (16):
wherein Z is a loss function result, n is the length of training data, which is determined according to the number of sensor parameters on the intelligent vehicle, d is the dimension of a sliding window,for the predicted lifetime of the cycle d+j times, x (d+j) is the actual lifetime.
5. An intelligent vehicle fault real-time prediction system based on end cloud cooperation is characterized by comprising:
the information acquisition unit is used for acquiring the original data of n systems of the intelligent vehicle during operation;
the fault diagnosis cloud service unit is used for carrying out real-time prediction training through a CNN-LSTM prediction model deployed at the cloud, obtaining a single-system residual service life prediction model corresponding to each system, and issuing the single-system residual service life prediction model to a vehicle end;
the vehicle-mounted controller is used for preprocessing the original data, obtaining the predicted residual life of the kth system through a trained single-system residual life prediction model, and obtaining the residual life of the whole vehicle through a serial-parallel system residual life prediction model according to the predicted residual life corresponding to each of the n systems, wherein k= … …, n; the raw data includes electronic control system related data, battery system related data, and motor system related data, and the series-parallel system remaining life prediction model is described as formula (17):
f＝min{(α*max{X，Y})，βZ} (17
wherein f represents the residual service life of the whole vehicle, alpha and beta are weights obtained by continuous iterative updating of a fault diagnosis cloud service unit, X represents the predicted residual life of an electronic control system, Y represents the predicted residual life of a battery system, and Z represents the predicted residual life of a motor system; the method for continuously and iteratively updating the obtained weight by the fault diagnosis cloud service unit specifically comprises the following steps:
judging whether the current evaluation index is better than the evaluation index obtained by the single-system residual service life prediction model trained last time, if not, updating the weight of the single-system residual service life prediction model trained at present; if yes, using a single system residual service life prediction model trained at present;
the evaluation index includes a score value fire calculated by a sum formula (4) of root mean square error RMSE calculated by formula (2):
in the formula delta i Representing the deviation between the calculated true value and the predicted value at the i-th moment, N is the total number of test samples, s i Represented by formula (5):
the method for obtaining the corresponding single-system residual service life prediction model through the CNN-LSTM prediction model specifically comprises the following steps:
step 21, inputting original data described by a feature map of n-dimensional feature dimensions;
step 22, extracting features of the original data sequence through a filter, activating the original data sequence through a Relu function, and outputting a first feature matrix C;
step 23, compressing the number of the extracted parameters of C through a maximum pooling function, and outputting a second feature matrix P;
step 24, according to the preset probability, randomly discarding the connection of part of neurons in P by a Dropout method, and outputting an influence factor characteristic sequence D;
and 25, inputting the D into a CNN-LSTM prediction model for prediction, and finally converting the data into a one-dimensional structure through a full connection layer to obtain the predicted residual life of the kth system.
6. The intelligent vehicle fault real-time prediction system based on end-cloud coordination as claimed in claim 5, wherein the electronic control system related data has steering control, power driving control, braking control and CAN management control of the electronic control system, the battery system related data has battery pack, number of single batteries, voltage value of single batteries, temperature of battery pack, charge and discharge state, total voltage, total current, state of charge and insulation resistance of the battery system, and the motor system related data has temperature, vibration, current, rotation speed and torque of the motor system.
7. The intelligent vehicle fault real-time prediction system based on end cloud cooperation as claimed in claim 5, wherein in step 21, the raw data or raw data sequence of the three systems are respectively described as feature graphs of n-dimensional feature dimensions, and are input into a CNN-LSTM model, as follows X t Shown;
wherein t represents single system data of the t-th time period, m represents the m-th feature, x tm Mth dimension data representing a t-th period;
in step 22, a convolution operation is performed by using a convolution check feature map of 3*3 through a filter in the CNN convolution layer, so as to extract features in the intelligent vehicle system, where the extracted features specifically include: characteristic values of steering control, power driving control, braking control and CAN management control of the electronic control system, battery packs, the number of single batteries, voltage values of the single batteries, battery pack temperature, charge and discharge states, total voltage, total current, charge states, insulation resistance and relevant characteristics of battery faults of the vehicle battery system, and relevant characteristics of temperature, vibration, current, rotating speed and torque of the vehicle motor system;
the convolution operation is shown in equation (6):
wherein X is the input of the convolution layer, C is the output of the convolution layer corresponding to the original data sequence X, W C The weight matrix is determined by the length and width channel number of the input parameters of the convolution layer, the weight matrix is initialized before convolution, and then a new weight matrix is generated by feedback after each layer of convolution c As the amount of the offset to be used,for convolution operation, f is a ReLU activation function, expressed as equation (7):
the pooling layer is used for extracting the remarkable characteristics in the intelligent vehicle system and reserving more information, and the pooling layer selects the maximum pooling mode as shown in a formula (8);
P＝max pooling(C) (8)
wherein P is the pooling layer output; max pooling (C) is the maximum pooling function;
in step 24, a Dropout mechanism is added to the CNN-LSTM model, and the equation (9) of the Dropout method is as follows:
wherein D is an influence factor feature sequence of the output of the Dropout layer: m is a random number between 0 and 1, p is a discarding probability, and p <1 is not less than 0.
8. The intelligent vehicle fault real-time prediction system based on end-cloud collaboration as claimed in any one of claims 5-7, wherein the core components of the CNN-LSTM model are respectively forgetting doors f t Input gate i t Output gate o t ；
At time t, there are three inputs to the memory cell of the CNN-LSTM model: input value x of network at present moment t Output value h of memory cell at last moment t -1, cell state C at the previous moment t -1; the outputs of the memory cells of the CNN-LSTM model are two: output value h of memory cell at current time t And cell state C at the current time t ；
Forgetting door f t Is responsible for controlling whether to continue to save the long-term state C or not, and is controlled by the input value x at the moment t t And the output value h of the hidden layer at the previous moment t -1 co-determination, forget door f t The calculation of (2) is shown in formula (10):
f t ＝σ(W f ·[h t-1 ，x t ])+b f (10)
wherein W is f Forgetting door f for t moment t Weight matrix of b) f Sigmoid activation function is adopted for the offset sigma;
input gate i t By input of the current state input x t And output h of the last state t -generating a value i between 0 and 1 in a Sigmoid function t To determine how much information the memory cell needs to retain, and a tanh layer outputs h from the previous state t -1 and input x of current state t To determine candidate memory state input values i t And candidate statesAs shown in the formula (11) and the formula (12):
i t ＝σ(W t ·[h t-1 ，x t ])+b i (11)
obtain input i t And a current candidate memory stateMultiplying to obtain the updated information to be added to the memory cell, and obtaining the state value C of the memory cell t The update of (2) is as shown in equation (13):
the output gate uses a Sigmoid function to determine how many memory cells of information C need to be output t Calculation as formula (14):
o t ＝σ(W o ·[h t-1 ，x t ]+b o ) (14)
information C of memory cell t And then with o t Multiplying and activating through a tanh layer to obtain output information h of the current LSTM block t ，h t As shown in equation (15):
h t ＝o t *tanh(C t ) (15)
the mean square error loss function Z is adopted for optimization in training, as shown in a formula (16):
wherein Z is a loss function result, n is the length of training data, which is determined according to the number of sensor parameters on the intelligent vehicle, d is the dimension of a sliding window,for the predicted lifetime of the cycle d+j times, x (d+j) is the actual lifetime.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111128029.4A CN113821875B (en) | 2021-09-26 | 2021-09-26 | Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111128029.4A CN113821875B (en) | 2021-09-26 | 2021-09-26 | Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113821875A CN113821875A (en) | 2021-12-21 |
| CN113821875B true CN113821875B (en) | 2023-08-29 |
Family
ID=78921132
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111128029.4A Active CN113821875B (en) | 2021-09-26 | 2021-09-26 | Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113821875B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114742297A (en) * | 2022-04-11 | 2022-07-12 | 中国第一汽车股份有限公司 | Power battery treatment method |
| CN114879050A (en) * | 2022-06-14 | 2022-08-09 | 山东大学 | Intelligent and rapid power battery service life testing method and system based on cloud edge cooperation |
| CN115482665B (en) * | 2022-09-13 | 2023-09-15 | 重庆邮电大学 | Knowledge and data collaborative driving multi-granularity traffic accident prediction method and device |
| CN117374469B (en) * | 2023-12-07 | 2024-02-09 | 深圳市普裕时代新能源科技有限公司 | Control method based on immersed liquid cooling energy storage system |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102015218830A1 (en) * | 2015-09-30 | 2017-03-30 | Bayerische Motoren Werke Aktiengesellschaft | Method for a motor vehicle, for detecting poor road conditions and a related system and motor vehicle |
| CN109460921A (en) * | 2018-11-12 | 2019-03-12 | 北京航天智造科技发展有限公司 | A kind of enterprise intelligentization transformation general-purpose system |
| CN109814523A (en) * | 2018-12-04 | 2019-05-28 | 合肥工业大学 | Method for diagnosing faults based on CNN-LSTM deep learning method and more attribute time series datas |
| CN110738360A (en) * | 2019-09-27 | 2020-01-31 | 华中科技大学 | equipment residual life prediction method and system |
| CN110956651A (en) * | 2019-12-16 | 2020-04-03 | 哈尔滨工业大学 | Terrain semantic perception method based on fusion of vision and vibrotactile sense |
| CN111144588A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院电工研究所 | Management system of vehicle PHM system |
| CN112633317A (en) * | 2020-11-02 | 2021-04-09 | 国能信控互联技术有限公司 | CNN-LSTM fan fault prediction method and system based on attention mechanism |
| CN112884717A (en) * | 2021-01-29 | 2021-06-01 | 东莞市牛犇智能科技有限公司 | System and method for real-time workpiece surface detection and tool life prediction |
-
2021
- 2021-09-26 CN CN202111128029.4A patent/CN113821875B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102015218830A1 (en) * | 2015-09-30 | 2017-03-30 | Bayerische Motoren Werke Aktiengesellschaft | Method for a motor vehicle, for detecting poor road conditions and a related system and motor vehicle |
| CN109460921A (en) * | 2018-11-12 | 2019-03-12 | 北京航天智造科技发展有限公司 | A kind of enterprise intelligentization transformation general-purpose system |
| CN109814523A (en) * | 2018-12-04 | 2019-05-28 | 合肥工业大学 | Method for diagnosing faults based on CNN-LSTM deep learning method and more attribute time series datas |
| CN110738360A (en) * | 2019-09-27 | 2020-01-31 | 华中科技大学 | equipment residual life prediction method and system |
| CN110956651A (en) * | 2019-12-16 | 2020-04-03 | 哈尔滨工业大学 | Terrain semantic perception method based on fusion of vision and vibrotactile sense |
| CN111144588A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院电工研究所 | Management system of vehicle PHM system |
| CN112633317A (en) * | 2020-11-02 | 2021-04-09 | 国能信控互联技术有限公司 | CNN-LSTM fan fault prediction method and system based on attention mechanism |
| CN112884717A (en) * | 2021-01-29 | 2021-06-01 | 东莞市牛犇智能科技有限公司 | System and method for real-time workpiece surface detection and tool life prediction |
Non-Patent Citations (1)
| Title |
|---|
| An efficient forecasting approach for resource utilization in cloud data center using CNN-LSTM model;Soukaina Ouhame;Neural Computing and Applications;第33卷;10043–10055 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113821875A (en) | 2021-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113821875B (en) | Intelligent vehicle fault real-time prediction method and system based on end cloud cooperation | |
| CN111339712B (en) | Proton exchange membrane fuel cell residual life prediction method | |
| CN112241608A (en) | Lithium battery life prediction method based on LSTM network and transfer learning | |
| CN111680848A (en) | Battery life prediction method based on prediction model fusion and storage medium | |
| CN114372417A (en) | Electric vehicle battery health state and remaining life evaluation method based on charging network | |
| WO2021017416A1 (en) | Deep compression power lithium battery fault diagnosis method under perceptual adversarial generation | |
| CN112434848B (en) | Nonlinear weighted combination wind power prediction method based on deep belief network | |
| CN112986827A (en) | Fuel cell residual life prediction method based on deep learning | |
| CN115291116B (en) | Energy storage battery health state prediction method and device and intelligent terminal | |
| CN116010900A (en) | Multi-scale feature fusion gearbox fault diagnosis method based on self-attention mechanism | |
| CN114692827A (en) | Electric vehicle lithium battery SOH online prediction method facing edge federal learning | |
| CN112215434A (en) | LSTM model generation method, charging duration prediction method and medium | |
| CN110516813A (en) | A method of batteries of electric automobile RDR prediction is carried out based on big data machine learning | |
| CN116298936A (en) | Intelligent lithium ion battery health state prediction method in incomplete voltage range | |
| CN116523177A (en) | Vehicle energy consumption prediction method and device integrating mechanism and deep learning model | |
| CN114355199A (en) | Battery thermal runaway risk prediction method and system based on recurrent neural network | |
| CN115166534A (en) | Method and system for predicting remaining service life of lithium ion battery | |
| CN113791351B (en) | Lithium battery life prediction method based on transfer learning and difference probability distribution | |
| CN116774089A (en) | Convolutional neural network battery state of health estimation method and system based on feature fusion | |
| CN116665483A (en) | Novel method for predicting residual parking space | |
| CN116736130A (en) | Lithium battery residual service life prediction method and system | |
| CN113459897B (en) | Electric vehicle charging big data-based state of charge correction and prediction method | |
| CN113989066A (en) | Galvanic pile energy consumption analysis based on big data | |
| CN114839539A (en) | Lithium battery SOH estimation method based on multilevel sequence information self-adaptive fusion | |
| CN114115199A (en) | Monitoring and fault diagnosis system for new energy sanitation vehicle |
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 |