Disclosure of Invention
Aiming at the fault characteristics of the numerical control machine and the defects of the conventional fault prediction method, the invention aims to provide the fault prediction method for the numerical control machine, which can improve the fault prediction capability of the whole machine and a subsystem of the numerical control machine, enhance the prediction robustness and has good application prospect and is based on the matching of a fault precursor judgment model and a dynamic confidence coefficient.
In order to solve the technical problems, the invention adopts the technical scheme that:
the invention relates to a fault prediction method for a numerical control machine tool, which comprises the following steps:
dividing the numerical control machine tool into a plurality of core subsystems by adopting a hierarchical structure model, and analyzing typical gradual faults of the numerical control machine tool;
installing temperature, vibration and noise sensors in each core subsystem, sending collected machine tool running state data samples to a fault prediction upper computer to obtain each sensor parameter data set, and reducing the sensor parameter data sets by using a neighborhood rough set method to obtain a fault precursor parameter data set and the relative association degree of the parameters and the faults;
selecting each fault foreboding parameter historical data set according to the fault evolution speed by taking the fault occurrence point as a boundary, and dividing according to the time sequence to obtain a corresponding fault foreboding state sequence;
extracting fault foreboding feature vectors of historical data in different time intervals by adopting a wavelet analysis technology, and training the fault foreboding feature vectors by adopting a back propagation neural network offline to obtain a fault foreboding judgment model of each parameter;
initializing the confidence coefficient, the accumulative confidence coefficient and other temporary variables of each state according to the fault precursor state sequence of each parameter, performing wavelet analysis on machine tool data samples acquired by each sensor in real time, and performing online fault precursor model identification on feature vectors by adopting a back propagation neural network to obtain the current state of each parameter real-time data sample;
judging the matching condition of the current state and the failure precursor state sequence, if the current state is matched with a certain state in the failure precursor state sequence, comparing the matching serial number of the current state with the matching serial number of the previous state, if the matching serial number of the previous state is smaller than the matching serial number of the current state, judging whether the descending count is larger than 0, if the descending count is equal to 0, quickly increasing the accumulative confidence level, further judging whether the accumulative confidence level reaches a parameter alarm threshold value, if the accumulative confidence level is larger than 0, slowly increasing the accumulative confidence level, and further judging whether the accumulative confidence level reaches the parameter alarm threshold value;
judging whether the accumulative confidence reaches a parameter alarm threshold value, if so, sending an alarm by the parameter, acquiring the accumulative confidence of all the premonitory representation parameters of the fault at the moment, and calculating the occurrence probability of the fault by means of the relative association degree of each parameter;
and judging whether the probability reaches a fault alarm threshold value, if so, selecting a matching state of a fault precursor parameter closest to the fault occurrence in time sequence and a matching state of the fault precursor parameter with the maximum product of the accumulated confidence coefficient and the relative correlation degree from the aspects of reliability and accuracy, respectively predicting the time of the fault occurrence, and taking the average value of the two times as the predicted occurrence time of the final fault.
If the last state matching serial number is equal to the current state matching serial number, judging the position of the current serial number in the failure precursor state sequence, if the current serial number is in the front of the sequence, reducing the accumulative confidence level, further judging whether the accumulative confidence level reaches a parameter alarm threshold value, if the current serial number is in the tail of the sequence, increasing the accumulative confidence level, further judging whether the accumulative confidence level reaches the parameter alarm threshold value, if the current serial number is in the middle of the sequence, correspondingly changing the accumulative confidence level according to the rising and falling trend of the actual state, and further judging whether the accumulative confidence level reaches the parameter alarm threshold value.
If the last state matching serial number is larger than the current state matching serial number, judging whether the ascending count is larger than 0, if so, quickly reducing the accumulative confidence, further judging whether the accumulative confidence reaches a parameter alarm threshold, if so, slowly reducing the accumulative confidence, and further judging whether the accumulative confidence reaches the parameter alarm threshold.
If the current state is not matched with any state in the failure foreboding state sequence, whether the previous state is matched with the state in the sequence or not is continuously judged, if not, the accumulative confidence coefficient is reduced, whether the accumulative confidence coefficient reaches a parameter alarm threshold value or not is further judged, if the previous state is matched with a certain state in the failure foreboding state sequence, the accumulative confidence coefficient is kept unchanged, and whether the accumulative confidence coefficient reaches the parameter alarm threshold value or not is further judged.
And if the accumulated confidence coefficient does not reach the parameter alarm threshold value, continuously acquiring the machine tool data sample in real time, and repeating the matching process.
And if the occurrence probability of the fault does not reach the fault alarm threshold value, continuously acquiring a machine tool data sample in real time, and repeating the matching process.
The hierarchical structure model is specifically as follows: the numerical control machine tool is used for constructing a structural model in a step-by-step manner according to six different levels, namely a machine tool complete machine level, a subsystem level, a fault level, a parameter level, a characteristic level and a state level, wherein the subsystem level is a core subsystem of the numerical control machine tool, the fault level is a typical gradual-change fault of each subsystem, the parameter level is a fault precursor parameter capable of representing the fault, the characteristic level is data characteristics of each fault precursor parameter before the fault occurs, and the state level is a fault precursor state corresponding to each data characteristic in the fault evolution/recovery process.
Selecting each fault foreboding parameter historical data set according to the fault evolution speed, dividing the fault foreboding parameter historical data sets into fault foreboding parameter data sets within a certain time range before the fault occurs according to the fault evolution speed, and dividing the data sets according to equal time intervals; one state in the failure precursor state sequence is used for representing a sampling data value contained in a corresponding interval; the back propagation neural network takes each component of the failure precursor feature vector as input, and the corresponding failure precursor state as output.
The initial confidence of each state in the failure precursor state sequence can be expressed as
Wherein p is
tFor the tth failure precursor parameter dataset of the failure,
for dividing the data set by time interval, corresponding to failure precursor state sequenceThe number of states in the column is,
is the state number of the match; states in the front of the failure precursor state sequence have lower initial confidence, and states in the tail of the failure precursor state sequence have higher initial confidence; the initial confidence degree increasing amplitude gradually increases along with the sequential matching of the failure foreboding state and then tends to be flat; the cumulative confidence of the fault evolution process after all fault precursor states are sequentially experienced is 1.
The failure occurrence probability is calculated by
Wherein p is
tAnd P is the number of failure foreboding parameters of the tth failure foreboding parameter data set of the failure.
The invention has the following beneficial effects and advantages:
1. the fault prediction capability of the whole machine and the subsystems of the machine tool is effectively improved. The method of the invention aims at the history records of the primary gradual change faults of the complete machine and the subsystem of the machine tool, comprehensively utilizes the time sequence, the wavelet analysis, the BP neural network and the information fusion technology, monitors the dynamic matching of the real-time state of the multi-fault foreboding parameters and the foreboding state sequence, predicts the occurrence time of the faults according to the accumulated confidence coefficient, and has the characteristics of high prediction accuracy, small prediction time difference and low prediction false alarm rate, thereby effectively improving the fault prediction capability of the complete machine and the subsystem of the machine tool.
2. And the robustness is strong. The method of the invention carries out fusion prediction aiming at multiple fault parameters of the complete machine and the subsystem of the machine tool, can adapt to the changes of the working environment and the system parameters of the machine tool to a certain extent, avoids the defect that single fault parameter loses information, and simultaneously adopts a multiple-precursor state sequence to represent the fault evolution process and is sensitive to micro state change, thereby having good robustness.
3. Has wide application prospect. The method of the invention needs less prior knowledge and has low requirement on original data, and can obtain the failure foreboding judgment model only aiming at the history of the gradual change failure of the whole machine and the subsystem of the machine tool, and the conditions of the working environment, the system parameters and the like of the machine tool are the same when the same failure occurs, and the model generated by sampling data is more accurate, thereby having good application prospect.
Detailed Description
The following describes an embodiment of a fault prediction method for a numerically controlled machine tool according to the present invention in detail with reference to the accompanying drawings and examples.
Fig. 1 is a block diagram of a failure prediction system applied in the method of the present invention. The temperature, noise and vibration sensors are used for acquiring machine tool running state data of a complete machine and a core subsystem of the numerical control machine in the graph, the temperature data are transmitted to the data acquisition main control box through preliminary processing of the temperature acquisition board, the noise data are transmitted to the data acquisition main control box through preliminary processing of the signal conditioner, the vibration data are directly transmitted to the data acquisition main control box, and the data acquisition main control box carries out unified denoising and filtering on the data of each sensor and then transmits the data to the fault prediction upper computer.
FIG. 2 shows a flow chart of the method of the present invention. Step S1, dividing the numerical control machine tool into a plurality of core subsystems by adopting a hierarchical structure model and analyzing typical gradual faults of the numerical control machine tool; step S2, acquiring machine tool running state data by using a sensor, reducing a sensor parameter data set, and obtaining a failure foreboding parameter data set and the relative association degree of the parameters and the failure; step S3, with the fault occurrence point as a boundary, selecting each fault foreboding parameter historical data set according to the fault evolution speed and dividing the data sets according to the time sequence to obtain a corresponding fault foreboding state sequence; step S4, extracting failure foreboding feature vectors of historical data in different time intervals by adopting a wavelet analysis technology, and training the failure foreboding feature vectors by adopting a back propagation neural network offline to obtain a failure foreboding judgment model of each parameter; step S5, the cumulative confidence of each failure foreboding parameter is monitored on line by adopting a dynamic confidence matching algorithm, the state dynamic matching result of each failure foreboding parameter is fused from the aspects of reliability and correctness, and the probability and time of failure occurrence are predicted on line.
Fig. 3 is a schematic diagram illustrating a process of constructing a failure precursor determination model in the method of the present invention. The specific implementation method comprises the following steps:
constructing a step-by-step hierarchical structure model of the numerical control machine according to the complete machine level, the subsystem level, the fault level, the parameter level, the characteristic level and the state level of the machine tool, wherein the subsystem level comprises U core subsystems, and the fault level analyzes typical gradual faults of each subsystem
Wherein k ∈ [1, U ]];
Aiming at the history of the gradual change fault f, a neighborhood Rough Set (RS) method is adopted to reduce a sensor parameter data Set to obtain a fault foreboding parameter data Set { p }t|t∈[1,P]And the relative degree of association of each parameter with the fault
Analyzing the fault evolution speed, selecting a fault foreboding parameter historical data set p in a certain time range by taking a fault occurrence point as a boundary
tAt equal time intervals
Sub-division into intervals
A sampling value, then
The sampling value contained in each interval corresponds to one state in the failure precursor state sequence;
for xi,jPerforming wavelet decomposition to obtain m-dimensional failure foreboding feature vector ei,j=(a1,a2,…,am);
Training and testing by using BP neural network to obtain failure foreboding judgment model
Wherein,
are inputs to the neural network,
is the connection weight of the neural network, failure foreboding state
Is the output of the neural network and is,
is composed of
An input sample.
Fig. 4 is a schematic diagram of a dynamic confidence matching algorithm in the method of the present invention. Wherein, the confidence degree refers to the credibility degree of the occurrence of the predicted fault;
sequence of finger failure precursor states including a pre-failure parameter p
tCorresponding time series based failure precursor states, the number of the states E is
A plurality of;
finger-shaped
Initial confidence of (2);
p in finger state matching process
tDynamic cumulative confidence of (2); CTH refers to a parameter such as p
tAn alarm threshold of (d);
finger p
tReal-time data samples of (a);
finger-shaped
The feature vector of (2);
finger p
tThe failure precursor determination model of (1);
finger-shaped
Warp beam
A recognized state; pre and now refer to the state sequence numbers of the previous matching and the current matching; up refers to the rising trend count of the premonitory state, i.e. the ascending count; down refers to the count of the descending trend of the premonitory state, i.e. descending count; the equivalent refers to the repeated matching times of the premonitory state, namely, the counting of the sequence; rho means
The reduction coefficient is used for the repeated matching situation of the failure foreboding state; e
1And E
2Between finger-divided zones
A boundary value of (d);
finger p
tRelative degree of association of; PR refers to the probability of failure occurrence; PRTH refers to an alarm threshold for predicting the occurrence of a fault f.
The method specifically comprises the following steps:
obtaining failure foreboding state sequence
The amount of data required to initialize the algorithm:
up,down,equal,E
1,E
2,ρ,CTH,PRTH,
etc.;
real-time acquisition of machine tool data samples
Wavelet decomposition to obtain characteristic vector
Based on
Identifying neural network online failure foreboding model and acquiring state of real-time data sample
Judging the current state
With failure precursor state sequences
If there is a match of
And
if the current state matching sequence number now is matched with the previous state matching sequence number pre, if pre is less than now, up + +, down-max (down-up, 0), and equal-0, according to whether down is greater than 0, the accumulated confidence change is as shown in formula (1),
if pre is now, up and down remain the same, equal + +, and the cumulative confidence change, depending on where now is in the sequence, is as shown in equation (2),
if pre > now, down + +, up max (up-down, 0), equal 0, the cumulative confidence change is as shown in equation (3) depending on whether up is greater than 0,
if it is not
And
if any state in the state table is not matched, the current state is the normal state
Continuously judging the previous state
Whether or not to cooperate with
The states in (1) match, if not, up and down remain unchanged, equal + +,
if it is not
And
if one state in the data is matched, keeping the accumulative confidence coefficient unchanged;
recording the current state
Judgment of
Whether the parameter alarm threshold value CTH is reached, if not, the step is continuedCollecting machine tool data samples continuously in real time, and repeating the matching process;
if it is not
When CTH is reached, see p
tSending out an alarm to obtain the accumulated confidence of all the premonitory parameters of the fault at the moment
By relative degree of correlation of all precursor parameters of the fault
Calculating the probability of occurrence of a fault
Judging whether the probability PR reaches a fault alarm threshold PRTH, if not, continuously acquiring a machine tool data sample in real time, and repeating the matching process;
if PR reaches PRTH, from the viewpoint of reliability, selecting the matching state of the failure precursor parameter closest to the failure occurrence in time sequence to predict the time T from the failure occurrence
1Then from the viewpoint of accuracy, the selection of the product satisfies
The matching state of the failure precursor parameters of the condition is used for predicting the time T from the occurrence of the failure
2And taking the average value of the two times as the predicted occurrence time of the final fault.
FIG. 5 is a graph illustrating initial confidence and cumulative confidence curves in the method of the present invention. The specific implementation method comprises the following steps:
in the process of fault evolution, the possibility that the state in the front of the fault precursor state sequence is misjudged is high, and a low initial confidence coefficient is given; the higher the possibility of future failure of the state at the tail of the failure precursor state sequence is, the higher initial confidence is given; in order to improve the efficiency of fault prediction, the initial confidence degree increasing amplitude is gradually increased and then becomes gentle along with the sequential matching of the fault precursor states. Therefore, the initial confidence is defined as shown in equation (4),
ideally, the fault evolution process goes through the fault precursor states according to the sequence of the fault precursor states, and the relationship between the cumulative confidence and the initial confidence at this time is shown in formula (5).
Fig. 6 is a schematic diagram showing the accidental occurrence of the normal state in the method of the present invention. When in use
Is in a normal state
When the fault state is a fault state, the state identification of the fault precursor is describedOtherwise, the accidental normal state occurs. This situation is generally due to two reasons: the system enters a normal state through maintenance, and the accumulative confidence coefficient is reduced at the moment; in the fault evolution process of the system, the system is obtained due to factors such as electromagnetic interference, network delay, mode identification errors and the like
Inaccurate, the cumulative confidence should not change at this time. The two reasons are combined, and in order to prevent the early warning from being lost, the accumulative confidence coefficient is not changed under the condition.
Fig. 7 is a schematic diagram illustrating forward evolution of failure precursor state in the method of the present invention. When pre < now, the forward evolution of the failure precursor state is shown in the failure precursor state identification process. There are generally three reasons for this: the system is in the process of fault evolution, and the accumulated confidence coefficient should be increased at the moment; the system is in the fault recovery process, but the state is in convolution matching due to factors such as electromagnetic interference, network delay, mode identification errors and the like, and the subsequent state indicates that the system is still in the fault recovery process, and the accumulated confidence coefficient is reduced or unchanged at the moment; the system is in the fault recovery process before, and then enters the fault evolution process, and the accumulated confidence coefficient should be increased at the moment. Accumulating the change in confidence requires further judgment of the above three reasons.
Fig. 8 is a schematic diagram illustrating the smooth evolution of the failure precursor state in the method of the present invention. When pre is not, the situation that the failure precursor state evolves smoothly in the failure precursor state identification process is explained. Cumulative confidence change and
the located intervals are related: if at [1, E
1]If the system is stable, the possibility of the system is high, and the accumulative confidence coefficient should be reduced; if at
The probability of system failure is high, and the accumulated confidence should be increased; if at (E)
1,E
2) And if so, correspondingly changing the accumulated confidence according to the actual state ascending and descending trend.
Fig. 9 is a schematic diagram illustrating reverse evolution of failure precursor states in the method of the present invention. When pre is greater than now, the situation that the failure precursor state reversely evolves in the failure precursor state identification process is shown. There are generally three reasons for this: the system is in the fault recovery process, and the accumulative confidence coefficient is reduced at the moment; the system is in the fault evolution process, but the state is in convolution matching due to factors such as electromagnetic interference, network delay, mode identification errors and the like, and the subsequent state indicates that the system is still in the fault evolution process, and the accumulated confidence coefficient is required to be increased or unchanged at the moment; the system is in the fault evolution process before, and then enters the fault recovery process, and the accumulated confidence coefficient is reduced at the moment. Accumulating the change in confidence requires further judgment of the above three reasons.
In order to verify the effectiveness of the method, the failure prediction method is researched through a main shaft transmission subsystem poor lubrication failure experiment in a vertical machining center, and the failure prediction accuracy rate AR, the failure prediction time difference TE and the failure prediction false alarm rate FAR of the failure prediction method are analyzed. Wherein, AR refers to the probability of failure occurrence accurately predicted by the prediction method; TE refers to the error between the time when the fault occurs and the actual time when the fault occurs in the prediction method, and the mean square error is
FAR refers to the probability that a prediction method predicts that a fault occurs but does not actually occur.
FIG. 10 further illustrates the method of the present invention using a spindle drive subsystem as an example. Wherein, VIBi(i-1, 2) is an eddy current type displacement vibration sensor, NOIi(i-1, 2) is a free field noise sensor, TEMPi(i ═ 1, 2, …, 12) is a patch type temperature sensor. TEMP1、VIB1And NOI1Mounted on spindle motors, TEMP2、VIB2And NOI2Mounted in gearboxes, TEMP3Is arranged in the middle of the bottom plate of the main shaft box,TEMP4Mounted on spindle sleeve, TEMP5Mounted on machine-tool tables, TEMP6Mounted at the lower end of the column, TEMP7Mounted on the upper end of a column, TEMP8Mounted on Z-axis ball-screw bearings, TEMP9And TEMP10Mounted on side plates of main spindle box, TEMP11Mounted on the front bearing of the spindle, TEMP12And is arranged on the rear bearing of the main shaft.
Firstly, aiming at the history record 300min before the occurrence of the main shaft lubrication fault, a field rough set method is used for reducing each sensor parameter data set to obtain a fault precursor parameter data set { NOI2, NOI3, VIB2, VIB3, TEMP2, TEMP11 and TEMP12} and the relative association degree of each parameter and the fault {0.21, 0.18, 0.08, 0.10, 0.16, 0.13 and 0.14 }. Dividing the parameter data set by the failure precursor state number E to obtain a time interval of (300/E) min, and corresponding to the failure precursor state sequence { s }1,s2,…,sEAnd selecting a sampling value every (300/E) s, obtaining 60 training original samples at each interval, training a fault foreboding judgment model, and finally identifying the running state of the spindle on line based on the fault foreboding judgment model and a dynamic confidence coefficient matching algorithm to monitor the evolution of the fault. The spindle rotation speed in the history record is 500rpm, the db3 wavelet is adopted to carry out four-layer decomposition on the original data to obtain 16-dimensional feature vectors, the BP neural network adopts a topological structure of '16 input-8 hidden-1 output', the rho in the dynamic confidence coefficient matching algorithm is 0.01, the CTH is 0.7, the PRTH is 0.5, and the E is1Is E/3, E2Is 2E/3.
In order to illustrate the influence of different spindle rotation speeds V and failure precursor state numbers E on AR and TE, historical failure data sets are divided by the failure precursor state numbers of 5, 10, 15, 20 and 25 respectively, corresponding failure precursor judgment models are constructed, and then spindle rotation speeds of 300rpm, 500rpm, 700rpm, 1000rpm and 1500rpm are respectively used for operation under poor lubrication conditions. And aiming at a failure precursor parameter data set 300min before the actual occurrence of the failure, selecting a group of sampling values by taking (300/E) min as an interval, obtaining 60 groups of sampling values by taking (300/E) s as delay, respectively carrying out state identification and dynamic confidence matching based on failure precursor judgment models with different state numbers, recording the accurate prediction group number, and determining the ratio of the AR to the accurate prediction group number and the sampling group number G.
Fig. 11 is a diagram showing the result of the failure prediction accuracy obtained by the method of the present invention. As the spindle speed increases, AR generally decreases. As the number of states increases, AR increases and then decreases. The maximum AR can reach 93.33% when the spindle speed is 300rpm and the state number is 15.
As shown in fig. 12 and 13, the failure prediction time difference result graph and the failure prediction time difference mean square error result graph obtained by the method of the present invention are shown. TE of each state number prediction model is concentrated in [ -20min, 20min under different spindle rotating speeds]In the above-mentioned manner,
less than 25min, acceptable for gradual failure and accurate time prediction.
To illustrate the effect of different spindle speeds V and the number of failure precursor states E on FAR, the lubrication was run at spindle speeds of 300rpm, 500rpm, 700rpm, 1000rpm, 1500rpm for 300min under normal conditions. And selecting a group of sampling values at the interval of (300/E) min, obtaining 60 groups of sampling values by taking (300/E) s as delay, respectively carrying out state identification and dynamic confidence matching based on the fault foreboding judgment models with different state numbers, and recording the false alarm group number, wherein FAR is the ratio of the false alarm group number to the sampling group number G.
As shown in table 1, the results of the fault prediction false alarm rate obtained by the method of the present invention are shown. The number of fault prediction false alarm sets increases with the increase of the rotating speed of the main shaft. When the spindle speed was 1500rpm and the number of states was 5, the FAR was 5% at the maximum. FAR is 0 when the spindle speed is low (V < 700 rpm).
TABLE 1
The results show that the method has the advantages of high failure prediction accuracy, small failure prediction time difference, low failure prediction false alarm rate and strong robustness, can effectively improve the failure prediction capability of the complete machine and the subsystem of the numerical control machine, reduces excessive detection and maintenance of the machine tool, ensures reliable operation of the machine tool, and has good application prospect.