CN100468263C - Continuous miner remote real-time failure forecast and diagnosis method and device - Google Patents

Abstract

The utility model relates to a method and a device for long-range real-time failure prediction and diagnosis of continuous miner, wherein the method comprises following steps: collecting data, processing data, forming a detector, learning and training the immune neural network, predicting and diagnosing failure and displaying the result. While the device comprises a host computer, a data communication device, a lower computer and a local sensing transmitter, wherein the local sensing transmitter comprises an acceleration transducer, a vibration pickup, a flow sensor, a pressure sensor and a temperature sensor; the data communication device is connected with the host computer and the lower computer respectively and the lower computer is connected with the local sensing transmitter. The local sensing transmitter device sends the signals collected by the sensors to the lower computer, while the collected signals undergo A/D transformation in the lower computer at certain time and then are sent to the host computer via the data communication device in order to predict and diagnose the failure. The utility model has the advantages of high accuracy and small sample size, and can not only predict and diagnose the failure, but also absorb and utilize new technology.

Description

Continuous miner remote real-time failure forecast and diagnostic method and device

Technical field

The invention belongs to failure prediction and diagnostic techniques field, particularly a kind of continuous miner remote real-time failure forecast and diagnostic method and device.

Background technology

Machine is adopted by cutting unit, tractive unit, motor and four most of compositions of auxiliary equipment by the most widely used company at present.Cutting unit adopts gear drive, and tractive unit adopts hydrostatic transmission.Most of faults of Lian Caiji are mechanical part fault and hydraulic system fault.

Comprise that the diagnosis of (1) temperature monitoring, (2) utilization analytic ferrographic method adopt the machine fault and carry out monitoring, diagnosing, (3) and connect and adopt the machine fault diagnosis expert system and (4) adopt the machine fault diagnosis system based on the company of artificial neural network connecting and connect the machine of adopting method for diagnosing faults commonly used at present.

Above-mentioned (1), (2) traditional diagnosis method do not have initiatively the ability of " thinking study ", it can only make judgement according to the field data that is input to the diagnostic system the inside, if the data of input are incomplete or conflicting, it will can not get conclusion even get the wrong sow by the ear, and the traditional diagnosis method poor efficiency of relatively requiring great effort.Same system is adopted (1), (2) two kinds of conventional fault diagnosis method, may obtain the diagnosis of mutual contradiction sometimes, at this moment still can only lean on experienced maintenance personal manually to diagnose.Connect and to adopt the machine fault diagnosis expert system there are the following problems: 1. " bottleneck " of knowledge acquisition.The management of present most coal mine machinery equipments, maintenance work are very unsound, lack various raw readings and theoretical property and sum up; Fault history, diagnosis and the service experience etc. of equipment do not form system documentation.2. the fragility of system, promptly knowledge and experience is not comprehensive, and it is just powerless to run into the problem that did not solve.Have only superficiality and empirical knowledge, lack essence and rational knowledge, ignored the deep layer effects such as understanding to knowledge, in case the situation that heuristic rule is not considered occurs, system performance sharply descends.3. knowledge base management difficulty.When knowledge base was too huge, because expert system adopts serial mode, inference method was simple and control strategy is dumb, caused knowledge shot array problems such as search speed descends, rule contradicts each other.Because system does not possess association and self-learning capability, to the suitable trouble of renewal, modification and maintenance of knowledge base.4. inferential capability is poor, and intelligence is low.5. versatility is poor, and particular device needs specific expert system.(defective of Artificial Neural Networks-ANN) is: 1. Learning Algorithms is a nonlinear optimal problem to artificial neural network, has the local minimum problem inevitably.Although 2. many improvement error back propagations (BP) algorithm is arranged, yet its studying convergence speed is still very slow.3. network must be through more time-consuming off-line training, and will obtain comprehensively and abundant training sample is difficult.

Summary of the invention

At above-mentioned problems of the prior art, the present invention combines immune algorithm with neural network, and a kind of continuous miner remote real-time failure forecast and diagnostic method and device based on immune neural net is provided.

Apparatus of the present invention comprise that host computer, data communication apparatus, slave computer, on-the-spot sensing become and send device.On-the-spot sensing becomes send device to comprise acceleration, vibration, flow, pressure, temperature sensor.Slave computer comprises one or more Programmable Logic Controllers.What data communication apparatus adopted is the CAN bus.Wherein data communication apparatus links to each other with host computer, slave computer respectively, and slave computer send device to link to each other with on-the-spot sensing change.Adopt the sensing change of the on-the-spot erecting stage of machine in company and send device, on-the-spot sensing becomes the sensor that send in the device signal of gathering is delivered to slave computer, slave computer regularly is sent to host computer with acquired signal by the CAN bus, and host computer carries out analyzing and processing to data again, carries out failure prediction and diagnosis then.

The each several part function of apparatus of the present invention:

(1) device is sent in on-the-spot sensing change: comprise acceleration, vibration, flow, pressure, temperature sensor, be responsible for parameter acquisition and transmission;

(2) slave computer: the signal A/D of being responsible for gathering is changed, and signal is sent to host computer;

(3) data communication apparatus: the signal that slave computer is sent on the CAN bus is sent to host computer through the USB-CAN module converts;

(4) host computer: receive the slave computer data, carry out analyzing and processing, provide the failure condition of Lian Caiji by the inventive method;

The inventive method mainly contains following characteristics:

(1) the immune system Negative Selection Algorithm is improved, and sure selection algorithm and improved Negative Selection Algorithm have been carried out effective combination, improved the efficient that detecting device generates by eliminating redundant detecting device;

(2) simultaneously by the adjustment of matching threshold, the detecting device that can obtain promptly will take place He break down;

(3) the fault sample of the detecting device that obtains, make the both measurable fault that will take place of immune neural net, the diagnosable again fault that has taken place as neural network training.

Attainable failure prediction of the present invention and diagnosis comprise adopts machine reduction-gear housing failure prediction and diagnosis, main electrical fault prediction and diagnosis, Main Hydraulic Pump failure prediction and diagnosis and hydraulic system fault prediction and diagnosis to connecting.

Failure prediction of the present invention and diagnostic method may further comprise the steps:

The collection of step 1, data

Gather acceleration, vibration, flow, pressure, temperature data; And through the A/D conversion;

The processing of step 2, data

Because the nerve of BP neural network adopts the Sigmoid function, if crossing conference, the input quantity absolute value make neuron output saturated, make the weights adjustment enter the flat region of error curved surface then, will make this moment error not too responsive to the variation of weights.So, before neural network training, need make normalized to the input data, the scope that makes it between [0,1], the present invention adopt suc as formula:

$x_i^{\′}=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}(\mathrm{HI}-\mathrm{LO})+\mathrm{LO}---\left(1\right)$

Method is carried out normalized to data.X wherein _iBe i data of measuring before the normalization, x _MaxBe { x _i, i=1,2 ..., the maximal value among the n; x _MinBe { x _i, i=1,2 ..., the minimum value among the n, x _i' be the data after the normalization, HI=0.9, LO=0.1, here add that two of HI, LO mainly are in order to eliminate input complete 0 in data and complete 1, occur 0 or at 1 o'clock because can find to input or output from weights are adjusted formula, can make to adjust and stagnate, error no longer reduces.In addition in normalized, because of the data differences of considering various patterns little, all on same quantity rank, so in normalized, minimum and maximum value is the value in all sample datas.If the data differences in each fault mode is bigger, basic on different data ranks, so in normalized, minimum and maximum value is got the value of sample data in the various fault modes respectively, rather than the value in all sample datas.

Step 3, generation detecting device

The most original Negative Selection fault detection method has a lot of defectives: at first be that negative selection algorithm is not considered the mutual coupling between the detecting device, may cause some detecting devices to flock together, can not cover all nonego spaces, perhaps need very big detectors set.Therefore, when generating detecting device, the detecting device with the detecting device generation coupling that has generated is removed, generate the detecting device collection affirming that selection algorithm and improved Negative Selection Algorithm effectively combine to calculate again;

(1) definition oneself string S be contain limited character, length is the set of the character string of L.Oneself string is that the information data of L standard sign equipment state of the data pattern during real coding equipment operate as normal links together as the normal mode string, and chooses h and go here and there as the normal mode string $R_{}^{}$ ${}_0^{\&prime;}=[{\mathrm{<figure-callout\; id="1"\; label="X"\; filenames="C200710012704E00252.png,C200710012704E00281.png"\; state="\{\{state\}\}">X</figure-callout>}}_1,{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="C200710012704E00252.png"\; state="\{\{state\}\}">X</figure-callout>}}_2,\&CenterDot;\&CenterDot;\&CenterDot;,X_h],h=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;.$ For certain fault mode i, i=1,2 ..., k (k is the kind of fault) links together the information data of the characterization device fault signature of the L of the fault mode of a real coding standard (but binding isotherm derivation) as the fault mode string

(being the detecting device individuality) chooses the h string as the fault mode string $R_i^{\&prime;}=[X_{\mathrm{hi}+1},X_{\mathrm{hi}+2},\&CenterDot;\&CenterDot;\&CenterDot;,X_{\mathrm{hi}+h}].$ Obtain oneself's string $S=({\mathrm{<figure-callout\; id="0"\; label="R"\; filenames="C200710012704E00292.png,C200710012704E00301.png"\; state="\{\{state\}\}">R</figure-callout>}}_0^{\&prime;},{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="C200710012704E00252.png,C200710012704E00281.png"\; state="\{\{state\}\}">R</figure-callout>}}_1^{\&prime;},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{\&prime;}).$

(2) set matching threshold z.The present invention obtains each detecting device collection, neural network training then by getting different matching threshold z respectively.By the graph of errors of neural metwork training process with select nominal situation at random and the fault diagnosis situation of the data detection immune neural net that various fault mode obtains, determine the size of more excellent matching threshold z then.Distance Matching principle that the present invention adopts European (Euclidean) characterizes the affinity size, as the formula (2), reaches necessary requirement (the more little then affinity of Euclidean distance is big more) if Euclidean distance d, shows affinity less than preset threshold z.

$d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z---\left(2\right)$

In the formula: d is an Euclidean distance; Ab _tCoordinate for antibody; Ag _iCoordinate for antigen; Z is matching threshold; L is the length of string.

(3) generate detecting device collection R _iGenerate detecting device collection D at random, to each detecting device among the D, with the normal mode string among the S

Carry out affinity according to formula (2) and calculate (the detected set D that antibody in the formula and antigen are respectively oneself string S and generate at random),, then put into detecting device collection R if d less than preset threshold z, illustrates that affinity is enough big ₀In, if can not satisfy condition, continue and first kind of fault mode string

Carry out affinity and calculate, if d, then puts into detecting device collection R less than preset threshold z ₁In, if can not mate, will continue and following a kind of fault mode string

Carry out affinity and calculate, if d less than preset threshold z, puts into detecting device collection R so ₂In, if can not mate, continue circulation, up to mating with last a kind of fault mode.Repeat this process till reaching needed detector number M.

(4) generation forecast detecting device collection.Getting matching threshold is z+a, calculates detecting device and first kind of fault mode string among the D

Affinity, if d less than preset threshold z+a and more than or equal to z, then puts into the detecting device collection

In, if can not mate, will continue and following a kind of fault mode string

Carry out affinity and calculate, if d less than preset threshold z+a and more than or equal to z, puts into the detecting device collection so

In, if can not mate, continue circulation, up to mating, obtain the detecting device collection with last a kind of fault mode $R_a=(R_1^a,R_2^a,\&CenterDot;\&CenterDot;\&CenterDot;,R_k^a).$ Repeat whole process till reaching needed detector number M.In like manner, get matching threshold and be respectively z+2a, z+3a, a ∈ R, the detecting device collection of each fault mode of different matching threshold correspondences is respectively $R_{2a}=({\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="C200710012704E00252.png,C200710012704E00281.png"\; state="\{\{state\}\}">R</figure-callout>}}_1^{2a},{\mathrm{<figure-callout\; id="2"\; label="R"\; filenames="C200710012704E00252.png"\; state="\{\{state\}\}">R</figure-callout>}}_2^{2a},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{2a})$ With $R_{3a}=({\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="C200710012704E00252.png,C200710012704E00281.png"\; state="\{\{state\}\}">R</figure-callout>}}_1^{3a},{\mathrm{<figure-callout\; id="2"\; label="R"\; filenames="C200710012704E00252.png"\; state="\{\{state\}\}">R</figure-callout>}}_2^{3a},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{3a}).$ The affinity of different matching threshold detecting devices is calculated by following European formula control

$z\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+a---\left(3\right)$

$z+a\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+2a---\left(4\right)$

$z+2a\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+3a---\left(5\right)$

In the formula: a is the regulated quantity of matching threshold; Z is matching threshold; D is an Euclidean distance.

The study of step 4, immune neural net and training

With the detecting device collection R that trains ₀Normal sample as neural network training; With the detecting device collection R that trains _i(i=1,2 ..., k), R _a, R _2aAnd R _3aAs the fault sample of neural network training, determine that output 1 is fault, 0 is normal.Because some signal can change before fault took place, what answer is contingent fault sample in the future so matching threshold is the detecting device set pair of z+a, z+2a, z+3a, the fail result of the output of its neural network is indicating the fault that equipment promptly will take place, and the confidence level that will break down is represented in each matching threshold.

The neural network that the present invention adopts is three layers of BP network model.The BP network model is exactly the training algorithm that adopts error back propagation, i.e. the feedforward network be made up of non-linear unit of BP network.Can adopt a kind of learning algorithm of supervising formula, its basic thought is to make the error sum of squares of network output layer reach minimum by the weights of regulating networks and threshold value, just makes output valve as far as possible near expectation value.Neural network learning, diagnosis and prediction detailed process are as follows in the inventive method:

(1) determining of network structure and learning parameter: determine that according to input vector the input layer of three-layer neural network counts n, determines that according to the fault mode kind output layer node counts q, rule of thumb and by considering the study number of times and the discrimination of network behind the training study, comprehensively relatively the back determines that hidden layer node counts p and weight convergence factor ζ and error convergence factor delta.

(2) network status initialization: give each connection weight W _Ij, V _JtAnd threshold value θ _j, γ _tGive the random value between [1 ,+1].

(3) provide the neural network learning sample right: selected characteristic sample data vector $A_k=[a_1^k,a_2^k,\&CenterDot;\&CenterDot;\&CenterDot;,a_n^k]$ And object vector $Y_k=[y_1^k,y_2^k,\&CenterDot;\&CenterDot;\&CenterDot;,y_q^k]$ Offer neural network.

(4) begin study, each sample carried out following operation:

1. each neuronic output of computational grid hidden layer and output layer

Hidden layer output $\begin{array}{cc}{b}_j=f(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{W}_{\mathrm{ij}}\&CenterDot;a_i-{\mathrm{\&theta;}}_j)& (j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,p)\end{array}---\left(7\right)$

Output layer output $\begin{array}{cc}{c}_t=f(\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{V}_{\mathrm{jt}}\&CenterDot;b_j-{\mathrm{\&gamma;}}_t)& (t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,q)\end{array}---\left(8\right)$

W in the formula _IjThe connection weight of-input layer i unit to hidden layer j unit;

θ _jThe threshold value of-hidden layer j unit;

V _Jt-hidden layer j unit to output layer t unit connection weight;

γ _t-output layer t cell threshode;

F-S type activation function $f\left(x\right)=\frac{1}{1+\exp (-x)};$

2. calculation training error

The hidden layer training error $\begin{array}{cc}{e}_j^k=[\underset{t=1}{\overset{q}{\mathrm{\&Sigma;}}}{d}_t^k\&CenterDot;V_{\mathrm{jt}}]\&CenterDot;b_j\&CenterDot;(1-b_j)& (j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,p)\end{array}---\left(9\right)$

The output layer training error $\begin{array}{cc}{d}_t^k=(y_t^k-c_t)\&CenterDot;c_t\&CenterDot;(1-c_t)& (t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,q)\end{array}---\left(10\right)$

In the formula

-desired output;

c _t-actual output;

b _jThe output of-hidden layer j unit;

3. revise weights

Input layer is to the new connection weight between the hidden layer:

$W_{\mathrm{ij}}(N+1)=W_{\mathrm{ij}}\left(N\right)+\mathrm{\&beta;}\&CenterDot;e_j^k\&CenterDot;a_i^k---\left(11\right)$

Hidden layer is to the new connection weight between the output layer:

$V_{\mathrm{jt}}(N+1)=V_{\mathrm{jt}}\left(N\right)+\mathrm{\&alpha;}\&CenterDot;d_t^k\&CenterDot;b_j---\left(12\right)$

In the formula

Respectively from formula (9) and (10);

N-study number of times;

α-learning coefficient, 0＜α＜1;

β-learning coefficient, 0＜β＜1;

(5) whether satisfy | b _j(N+1)-b _j(N) | 2＜ζ, carried out for (6) step if satisfy, otherwise turned back to for (4) step.

(6) whether satisfy | b _j(N)-Y _k|＜δ goes on foot if satisfy then carry out (7), otherwise turned back to for (2) step.

(7) weights and the threshold data that obtains deposited in the knowledge base of being set up.

Step 5, failure prediction and fault diagnosis

The neural network that trains is used for the failure prediction and the diagnosis of equipment, can judges according to output vector which kind of fault has taken place.For example: setting desired output and each modes relationships is: normal (0,0,0,0,0,0); Fault 1 (1,0,0,0,0,0); Fault 2 (0,1,0,0,0,0); Fault 3 (0,0,1,0,0,0); Prediction fault 1 (0,0,0,1,0,0) (confidence level 80%); Prediction fault 2 (0,0,0,0,1,0) (confidence level 80%); Prediction fault 3 (0,0,0,0,0,1) (confidence level 80%).Detect vector data to one group and input to the immune neural net system, after the prediction and diagnosis through immune neural net, which kind of pattern is actual output vector just can be defined as during near which kind of pattern desired output vector.When output vector is (0.2,0.1,0.2,0.1,0.1 , 0.8) time, with the prediction fault 3 (0,0,0,0,0, 1) more approaching, so determine fault for this reason.For the fault mode that increases newly, the process that generates detecting device by improved Negative Selection Algorithm obtains the detecting device collection, and neural network training gets final product then, uses very convenient.

Step 6, demonstration output result

The present invention is applied to Negative Selection Algorithm in the middle of the acquisition of neural network training fault sample according to immune characteristics, has improved the reliability of fault sample.Because some signal can change before fault took place, what choose fault sample correspondence that different matching threshold obtains is contingent fault sample in the future, Xun Lian neural network thus, the fail result of its output is indicating the fault that equipment promptly will take place, so the present invention can to or occurent equipment failure diagnose, can predict following imminent fault again.

The present invention---show as based on the remote real-time failure forecast of the Lian Caiji of immune neural net and the superiority of diagnostic method and device:

(1) is fit to the occasion that fault sample lacks,, comes neural network training according to the immune algorithm sample set that just can extensively be out of order as long as know a small amount of typical sample of fault; For the equipment of non-fault sample,, come neural network training according to the immune algorithm sample set that also can extensively be out of order if can reason out a spot of fault sample according to theory.

(2) for imminent fault, be difficult to extract feature, this method can be extracted the feature that is about to break down easily according to the adjustment of matching threshold.The neural network of training, both diagnosable fault, measurable again imminent fault.

Description of drawings

Fig. 1 is apparatus of the present invention structural representation;

Fig. 2 is example middle controller and controller CANOPEN wiring diagram;

Fig. 3 is example middle controller and host computer wiring diagram;

Fig. 4 chooses synoptic diagram for cut speed reduction unit measuring point;

Fig. 5 is for generating the process flow diagram of detecting device by improved Negative Selection Algorithm;

Fig. 6 adopts machine failure prediction and diagnosis immune neural net structure for connecting in the example;

Fig. 7 is an immune neural net training module process flow diagram in the example;

Fig. 8 reverse transmittance nerve network training process error curve diagram;

Fig. 9 immune neural net training process error curve diagram;

The verify error curve map of Figure 10 reverse transmittance nerve network;

The verify error curve map of Figure 11 immune neural net;

Figure 12 is a fault display interface in the example;

Figure 13 is a primary control program process flow diagram in the example;

Figure 14 is the output display interface of the immune neural net of cut reduction-gear housing;

Wherein: 1-measuring point, 1,2-measuring point, 2,3-measuring point, 3,4-measuring point, 4,5-motor, 6-Machine guarding, 7-Machine guarding, 8-motor, 9-left cutting mechanism, 10-reduction-gear housing, 11-connecting shaft, 12-right cutting mechanism.

Embodiment

Be used for connecting to adopt the failure prediction of machine cut reduction-gear housing and be diagnosed as example further specifying with the present invention below.

Apparatus of the present invention comprise that host computer, slave computer, data communication apparatus, on-the-spot sensing become and send device that wherein on-the-spot sensing becomes send device to comprise acceleration, vibration, flow, pressure, temperature sensor.As shown in Figure 1.

Slave computer adopts the third generation CAN controller of Finland Epec Oy company exploitation in this example---and EPECOY.E3002024 (being called for short EPEC2024), EPEC2024 is embedded CAN bus communication module carries CANopen and CAN2.0 bus interface.Only need the pairing stitch of CAN bus access controller is provided with its node number, just can realize the mutual communication of controller on the CAN bus, as shown in Figure 2.

Host computer links to each other by the embedded CAN bus communication module of USB interface and EPEC2024, CAN_GND, the CAN_L of the CAN bus communication module that EPEC2024 is embedded, CAN_H pin link to each other with XM4.3, XM4.6, the XM4.2 pin of controller respectively, as shown in Figure 3.Hop adopts twisted-pair feeder to realize host computer and the long-range carrier communication of slave computer, and communications protocol adopts the senior communications protocol of CANOPEN.

On-the-spot sensing becomes the signal that send device to adopt EPEC2024 controller pick-up transducers to be collected, and the transmission storage space of the data storage of being gathered to the EPEC2024 controller, by program data are sent on the CAN bus, transmission interface receives and sends to the information on the CAN bus and send into host computer.

Cutting mechanism is one of vitals of Lian Caiji, and it is made up of 2 motor, two cover Machine guardings, speed reduction unit, left and right sides cutting drum, cutter chain and a cutting arm etc.Two motors split in the cutting arm left and right sides with vertical machine y direction symmetry, send power to speed reduction unit by Machine guarding separately, and the shaft of speed reduction unit drives left and right cutting drum and the broken coal seam of cutter chain simultaneously.Because motor is axially vertical with the big direction of principal axis of intermediate calender rolls, so speed reduction unit adopted one-level spur gear, one-level cone gear (commutation) and three grades of decelerations of primary planet pinion, and reduction case is T-shaped.Cutting mechanism acts directly on the coal petrography, be cantilever position during cut, the stressed complexity of cylinder, the cut load change is bigger, cause mechanical disorder easily, in the course of the work, speed reduction unit bears high varying load, becomes high pulling torque, rate of breakdown is higher, most of faults of Lian Caiji appear at cutting unit speed reduction unit this part.In this example to five kinds of diagnosing malfunctions of distortion of its broken teeth, flank of tooth fatigue, deformation of tooth surface, tooth surface abrasion, axle.

Failure prediction and diagnostic method are carried out according to the following steps in this example:

The collection of step 1, data

When carrying out data acquisition, reasonably selecting measuring point is online prerequisite of carrying out failure prediction and diagnosis.According to the needs of analyzing and the design feature of equipment, according to the structure of reduction-gear housing and the selection principle of vibration signal measuring point, the measuring point arrangement as shown in Figure 4, measuring point 1 (3) and measuring point 2 (4) differ 180 °, 4 measuring points all are placed on the output shaft end cap.The frequency of 1～5f scope is divided into one section every 1f, measures extreme value in each segment, these 16 extreme values are joined together as the description of this data vector as the representative of this section.Adopt the HG-3518 data acquisition unit to carry out signal data acquisition, the instrument sampling number selects N=1024, maximum analysis frequency f=2000Hz.

The processing of step 2, data

Obtained 900 groups of data through experiment, before neural network training, the input data have been made normalized, the scope that makes it is between [0,1], and after employing was carried out normalized suc as formula (1) method to data, it was as shown in table 1 to have chosen 18 groups of data wherein.

The training sample data of table 1 neural network

Step 3, generation detecting device

(1) the self-string length of definition is the limited string assemble S of L=16.Choose three strings as the normal mode string $R_{}^{}$ ${}_0^{\&prime;}=[{\mathrm{<figure-callout\; id="1"\; label="X"\; filenames="C200710012704E00252.png,C200710012704E00281.png"\; state="\{\{state\}\}">X</figure-callout>}}_1,{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="C200710012704E00252.png"\; state="\{\{state\}\}">X</figure-callout>}}_2,X_3];$ For certain fault mode i, (i=1,2 ..., 5), choose three strings as every kind of fault mode string $R_i^{\&prime;}=[X_{3i+1},X_{3i+2},X_{3i+3}].$

(2) generate detecting device collection R _iSet matching threshold z=0.500 earlier, generate detecting device collection D at random, for each detecting device among the D, with the normal mode among the S

Carry out affinity and calculate, if d less than preset threshold 0.500, then puts into detecting device collection R suc as formula (2) ₀In, if can not satisfy condition, continue and broken teeth fault mode string

Carry out affinity and calculate, if d, then puts into detecting device collection R less than preset threshold 0.500 ₁In, if can not mate, will continue and following a kind of flank of tooth wear-out failure pattern string

Carry out affinity and calculate, if d less than preset threshold 0.500, puts into detecting device collection R so ₂In, if can not mate, continue circulation, up to mating, obtain detecting device collection R with last a kind of bending shaft fault mode ₅Repeat this process till every kind of pattern reaches needed detector number 400.So just generated the normal mode of operation R of cut reduction-gear housing ₀, and the detecting device collection of each fault mode: broken teeth R ₁, the tired R of the flank of tooth ₂, deformation of tooth surface R ₃, tooth surface abrasion R ₄With bending shaft R ₅

Ideal matching threshold value z=0.420 has been determined in the training of the immune neural net by three kinds of matching thresholds, then generation forecast detecting device collection.The setting matching threshold is z+a=0.420+0.020=0.440, promptly gets a=0.020.(the detecting device collection is R for detecting device among the calculating D and the data pattern among the S ₀) carry out affinity calculating suc as formula (2), if d less than preset threshold 0.500, then deletes,, continue and broken teeth fault mode string if can not satisfy condition

Carry out affinity and calculate, if d less than preset threshold z+a=0.440 and more than or equal to z=0.420, then puts into the detecting device collection

In, if can not mate, will continue and following a kind of fault mode string

Carry out affinity and calculate, if d less than preset threshold z+a=0.440 and more than or equal to z=0.420, puts into the detecting device collection so

In, if can not mate, continue circulation, up to mating, obtain the detecting device collection with last a kind of fault mode

Repeat whole process till reaching needed detector number M=400.In like manner, get matching threshold and be respectively z+2a, z+3a (a=0.020), the detecting device collection of each fault mode of different matching threshold correspondences is respectively

With

The affinity of different matching threshold detecting devices is calculated by formula (3), (4), (5) and is calculated.

Be the detecting device R that z=0.420 and a=0.020 generate with matching threshold like this ₀, (R ₁, R ₂..., R ₅),

Sample training neural network as neural network.The immune neural net of training both can carry out failure prediction, can carry out fault diagnosis again.

The study of step 4, immune neural net and training

According to input vector determine input layer count n=16, according to fault mode kind output layer node count q=21, the hidden layer node number is a key of finding the solution network problem, it directly has influence on the recognition capability of network model to test sample book, rule of thumb and by the node that this immune neural net behind the training study is chosen hidden layer counts p=15.The final network structure of determining is 16-15-21, as shown in Figure 6.Selected weight convergence factor ζ=0.001 and error convergence factor delta=0.1, training objective 0.001, maximum training batches 2000 times.

The corresponding relation of table 2 cut reduction-gear housing typical fault and failure symptom (different threshold value)

Determine the network model and the training data of neural network this moment, then determined the encoding relation of the ideal output of pattern to be discriminated and neural network.Provided the corresponding relation of cut reduction-gear housing typical fault to be discriminated with failure symptom (different threshold value) in table 2, it is that equipment operation is normal that output is at 0 o'clock.

Comparison example

(1) contrast of immune neural net and BP neural metwork training graph of errors

Train the BP neural network to obtain the graph of errors of BP neural metwork training process shown in Figure 8 with the data that 900 groups of experiments obtain, as can be seen from the figure after having passed through 1398 training, just reached the target error of setting.

During matching threshold z=0.420 above, the detecting device collection R of acquisition ₀(R ₁, R ₂..., R ₅) train immune neural net as the training sample of immune neural net.Obtain the graph of errors of immune neural net training process shown in Figure 9, as can be seen from the figure after having passed through 1058 training, just reached the target error limit of setting.

By in the training error curve of Fig. 8 and Fig. 9 as can be seen this immune neural net only need 1058 training just can reach the training objective error of former BP neural network through training just to reach for 1398 times.So generally speaking immune neural net can be saved time of 1/4th than former BP neural network, so immune neural net error convergence speed is very fast.

(2) fault diagnosis examples comparative

With 60 groups of check data randomly drawing the BP neural network that trains is tested, in table 3～table 8, provided the actual output diagnostic result of the BP neural network after the check.With the error (∞-norm as shown in Equation 14) of ideal output with the ∞-norm mark BP neural network of actual output result's difference, calculate detection error as shown in table 9, for the ease of observing, this paper is according to the error result of the table 9 verify error curve of BP neural network that drawn accordingly, as shown in figure 10.

‖x—y‖ _∞＝max(|x ₁—y ₁|，|x ₂—y ₂|，…，|x ₂₁—y ₂₁|) (14)

Wherein＜x ₁, x ₂..., x ₂₁Be the real output value of BP neural network,＜y ₁, y ₂..., y ₂₁It is the idea output of BP neural network.

Table 3 nominal situation is the diagnostic result of corresponding BP neural network down

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.894	0.027	0.135	-0.391	0.033	0.073
0.652	0.087	-0.217	0.136	0.249	-0.276
						0.617	-0.156	0.107	0.312	-0.061	0.327
0.583	0.269	0.266	0.290	0.224	0.055
						0.877	0.156	0.301	-0.213	0.167	0.195

0.686	-0.030	0.177	0.292	-0.324	-0.122
						0.597	0.195	0.314	0.136	0.164	0.192
0.751	0.354	-0.281	0.272	0.246	0.255
						0.746	0.183	0.467	0.330	0.360	0.143
0.906	0.158	0.144	-0.281	0.278	0.185

The output of the BP neural network of table 4 broken teeth correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						-0.035	0.797	0.437	-0.212	0.215	-0.186
0.306	0.852	-0.213	0.192	0.374	0.360
						-0.074	0.601	0.314	0.266	-0.136	0.316
0.133	0.751	0.285	0.348	0.171	0.500
						0.115	0.615	-0.126	0.319	0.415	0.370
0.093	0.582	-0.475	0.309	0.414	0.345
						0.079	0.911	0.193	0.242	0.243	0.381
0.145	0.687	0.329	0.271	0.181	0.265
						-0.044	0.817	0.387	0.166	0.442	0.434
0.124	0.922	0.187	0.239	0.203	0.199

The output of the BP neural network of the tired correspondence of table 5 flank of tooth

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.103	0.384	0.684	0.243	0.275	0.168
0.082	0.262	0.710	0.304	0.304	-0.081
						-0.085	0.319	0.744	0.263	0.231	0.125
0.176	0.341	0.572	0.397	0.460	0.285
						0.184	0.218	0.843	0.179	-0.121	0.260
0.128	0.320	0.915	-0.148	0.254	0.211
						0.038	-0.166	0.738	0.391	0.183	0.181
0.242	0.208	0.605	0.322	0.411	-0.028
						0.145	0.210	0.757	0.231	0.196	0.222
-0.024	0.179	0.597	0.342	0.109	0.247

The output of the BP neural network of table 6 deformation of tooth surface correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						-0.101	0.180	0.262	0.546	0.475	0.168

0.067	-0.213	0.161	0.794	0.221	0.256
						-0.230	0.314	0.341	0.651	0.381	-0.279
0.129	0.236	0.189	0.814	0.214	0.186
						0.061	0.434	0.143	0.731	0.115	0.333
0.021	0.112	0.268	0.911	0.196	-0.162
						0.142	0.255	0.312	0.607	0.360	0.169
0.282	0.296	0.243	0.813	0.247	0.219
						0.110	0.220	-0.190	0.741	0.295	0.126
-0.012	0.190	0.178	0.605	0.301	0.243

The output of the BP neural network of table 7 tooth surface abrasion correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						-0.001	0.265	0.376	0.423	0.827	0.336
0.014	0.362	-0.188	-0.173	0.912	-0.188
						0.115	0.381	0.266	0.380	0.682	0.396
0.170	0.307	0.439	0.321	0.643	0.338
						0.093	-0.175	0.215	0.291	0.702	0.211
-0.128	0.371	0.272	0.263	0.592	0.228
						0.092	0.270	-0.320	0.216	0.709	-0.124
0.167	0.269	0.148	-0.312	0.617	0.301
						0.317	0.247	0.335	0.456	0.728	0.163
0.113	0.323	0.241	0.385	0.723	0.262

The output of the BP neural network of table 8 bending shaft correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						-0.103	-0.104	-0.213	0.140	0.257	0.901
0.092	0.170	0.326	0.296	-0.223	0.719
						-0.102	0.152	0.381	0.117	0.283	0.882
0.107	0.144	0.315	0.378	-0.176	0.724
						0.044	0.219	0.175	0.145	0.283	0.584
0.053	-0.355	0.142	0.402	0.262	0.827
						-0.030	0.298	0.292	0.246	-0.181	0.728
0.124	0.192	-0.125	-0.299	0.302	0.620
						0.092	-0.126	0.295	0.354	0.317	0.597
0.081	0.308	0.204	0.271	0.114	0.742

The detection deviation of table 9 BP neural network diagnostic result

Maximum 0.475 in 60 verify errors as can be seen from Table 9, minimum then is 0.186.Can find out after this network is through training it is to satisfy the fault diagnosis of even adopting machine cut reduction-gear housing substantially to require intuitively from the verify error curve.But the diagnostic result deviation is also bigger, and is most of between 0.350～0.460, forms erroneous judgement easily according to the output of BP neural network.

Then the immune neural net that trains is tested, in table 10～table 15, provided the actual output diagnostic result of the immune neural net after the check with 60 groups of check data randomly drawing.With ideal output and the actual error of the ∞ of difference-norm labelled immune neural network as a result of exporting, draw detection error result shown in table 16 at last, for the ease of observing, this paper is according to drawn the accordingly verify error curve of immune neural net of the error result of table 16, as shown in figure 11.

The output of the immune neural net of table 10 nominal situation correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.853	0.184	-0.014	0.209	0.122	0.205
0.981	-0.088	0.189	0.135	-0.139	0.191
						0.862	0.123	-0.109	-0.077	0.051	-0.166
0.986	-0.103	0.034	0.105	0.152	0.073
						0.862	0.086	0.103	0.132	0.038	0.152
0.845	0.171	0.097	0.067	0.102	0.201
						0.953	0.050	0.035	0.086	0.115	0.143
0.852	0.049	0.100	-0.212	0.215	0.169
						0.794	0.207	0.072	0.122	0.121	-0.148
0.902	0.142	0.104	0.201	0.077	0.056

The output of the immune neural net of table 11 broken teeth correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.082	0.998	-0.131	0.039	-0.123	0.064
-0.093	0.906	0.045	-0.184	0.202	-0.141
						0.176	0.876	0.091	0.204	-0.114	0.013
-0.161	0.891	0.158	-0.155	0.226	0.171
						0.050	0.798	-0.117	0.063	0.201	-0.149
0.107	0.848	0.130	0.180	0.150	0.099
						-0.124	0.923	-0.204	0.191	-0.138	0.072
0.185	0.884	-0.082	-0.097	0.209	-0.176
						-0.219	0.785	0.158	-0.076	-0.169	0.225
0.110	0.958	0.183	0.034	0.098	0.169

The output of the immune neural net of the tired correspondence of table 12 flank of tooth

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.089	0.213	0.984	0.143	0.073	0.175
0.123	-0.072	0.893	-0.117	0.148	-0.185
						0.159	0.113	0.839	0.026	-0.092	0.099
-0.105	0.156	0.877	-0.109	0.212	0.224
						0.046	-0.125	0.789	0.018	0.032	0.236
0.267	0.087	0.885	0.056	0.139	0.229
						0.228	0.172	0.966	-0.118	-0.137	-0.194
-0.134	0.243	0.887	0.239	0.144	0.156
						0.121	0.244	0.977	0.132	-0.024	0.117
0.165	0.153	0.874	0.218	0.123	0.237

The output of the immune neural net of table 13 deformation of tooth surface correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.170	0.072	0.104	0.994	0.141	-0.138
-0.222	-0.157	0.129	0.869	-0.211	0.078
						0.081	0.312	-0.055	0.792	0.127	-0.165
-0.089	-0.225	0.193	0.896	-0.160	0.088
						0.157	0.242	0.128	0.867	0.215	0.151

0.166	0.185	-0.186	0.982	0.174	0.155
						-0.076	-0.220	0.234	0.885	0.184	-0.217
0.216	0.133	-0.283	0.876	-0.088	0.133
						-0.098	-0.233	0.219	0.915	0.152	0.038
0.154	0.131	0.205	0.853	0.210	0.141

The output of the immune neural net of table 14 tooth surface abrasion correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion decreases	Bending shaft
						0.249	0.116	0.229	0.172	0.925	0.287
-0.072	-0.098	-0.209	-0.169	0.878	0.201
						0.040	0.087	0.130	0.240	0.876	0.193
-0.092	-0.015	-0.181	0.051	0.955	-0.056
						0.106	0.214	0.133	0.247	0.797	0.073
-0.090	0.125	0.217	0.212	0.851	0.221
						0.137	0.215	0.163	0.213	0.918	0.145
0.128	-0.244	0.230	0.176	0.877	-0.058
						0.223	0.173	-0.228	0.218	0.941	0.290
-0.085	0.065	0.158	-0.174	0.901	0.159

The output of the immune neural net of table 15 bending shaft correspondence

Nominal situation	Broken teeth	Flank of tooth fatigue	Deformation of tooth surface	Tooth surface abrasion	Bending shaft
						0.082	0.209	0.285	0.072	0.115	0.937
-0.150	-0.178	-0.172	0.158	0.220	0.873
						0.058	0.137	0.179	-0.223	-0.119	0.905
0.063	0.254	0.231	0.192	0.212	0.886
						0.185	0.042	0.134	0.241	0.094	0.793
0.187	0.108	0.162	0.196	0.092	0.963
						0.285	0.126	-0.253	0.257	0.109	0.875
0.223	-0.119	0.062	-0.241	-0.165	0.929
						-0.139	0.097	0.166	0.184	0.211	0.903
0.173	0.158	0.210	0.264	0.062	0.859

The detection deviation of table 16 immune neural net diagnostic result

Maximum in 60 verify errors as can be seen from Table 16 also have only 0.290, and minimum then has only 0.131.Be to satisfy to connect the fault diagnosis of adopting machine cut reduction-gear housing after the immune neural net process is trained as can be seen by verify error.Can compare error-checking size very intuitively by Figure 10 and Figure 11.And the error convergence speed of comparing immune neural net with the BP neural network is very fast, and output bias is less, and accuracy is higher.

It is faster than BP neural network speed of convergence to draw the immune neural net speed of convergence according to the comparison of training error curve; The diagnostic result that can draw immune neural net according to the comparison of verify error curve is more accurate than the diagnostic result of BP neural network.So the method for diagnosing faults ratio based on immune neural net that invention proposes is more superior based on the BP Neural Network Method for Fault Diagnosis.

Step 5, failure prediction and fault diagnosis

In order better to verify accuracy and the practicality that immune neural net carries out failure prediction and diagnosis, the application practice by the eastern coal mining of god scene has fully shown accuracy of the present invention and practicality.

In the eastern coal mining production run of god, on Lian Caiji, used the inventive method and carried out failure prediction and diagnosis.The confidence level that can judge the fault that tooth surface abrasion takes place from the output of immune neural net on April 16 is 85%, because for suitable stage reason, the cut reduction-gear housing is not safeguarded this moment, still allows it proceed running.And, because construction environment is abominable, noise is bigger, to compare with other reduction-gear housing, this moment, imperceptible reduction-gear housing from the sense organ had fairly obvious variation.Find that reduction-gear housing sends big abnormal sound but spent two; body temperature is above 70 ℃; and can judge that the confidence level of the fault that flank of tooth fatigue takes place is 70% in the output of apparatus of the present invention at this moment, the confidence level that the fault of tooth surface abrasion takes place is 100%, should in time carry out shutdown maintenance.Find through tearing the reduction-gear housing maintenance open: occur a spot of spot corrosion, peeling phenomenon on the flank of tooth of one-level spur gear, and the primary planet pinion flank of tooth has a large amount of cuts, because this accident treatment is in time appropriate, avoided complete machine to be in and stagnated even state of paralysis, guaranteed working continuously of production.

The present invention---based on continuous miner remote real-time failure forecast and the diagnostic method and the device of immune neural net, can compare remote live judgement fast and accurately to connecting the operational situation of adopting machine cut reduction-gear housing, avoid the blindness of maintenance, fully guaranteed the normal operation of Lian Caiji.

Step 6, output is display interface as a result, as shown in figure 14

The major failure of cut reduction-gear housing has been listed in output as a result in the display interface: broken teeth, flank of tooth fatigue, deformation of tooth surface, tooth surface abrasion and bending shaft, correspondingly listed failure cause, output result, confidence level (for different threshold values) and historical record etc.It was fault greater than 0.65 o'clock that native system is set immune neural net output, whether the output result only shows fault, there is the result of fault output to adopt the pink colour font simultaneously, be convenient to observe the output result like this, click the historical record button, can see detailed historical record (date of breaking down, time, output result etc.).

For the software development part, select suitable developing instrument can reduce the difficulty of system development, accelerate development progress.The design of this monitoring software adopts Visual C++6.0 as developing instrument based on WindowsXP operating system.The design philosophy of this monitoring software is that the failure prediction and the diagnostic method that adopt the present invention to propose are monitored underground equipment, mainly comprises the generation of data acquisition, detecting device, the study of neural network and the realization of parts such as training, failure prediction and fault diagnosis, fault log management.

The main control process flow of host computer realization failure prediction and diagnosis is carried out according to the following steps, as shown in figure 13:

Step 1, beginning;

The setting of step 2, baud rate and controller equiment number;

Step 3, initialization USB-CAN modular converter;

Step 4, data acquisition;

The processing of step 5, data;

The generation of step 6, detecting device;

Step 7, neural network learning and training;

Whether step 8, failure judgement prediction and diagnosis have fault; If, execution in step 9; If not, return step 4;

The type of step 9, diagnostic process failure judgement;

Step 10, fault prompting are reported to the police, and send failure message to the down-hole, return step 4.

The element that Lian Caiji often breaks down also has main motor, hydraulic power unit.The fault of motor mainly shows as bolt looseness, bearing fault and resonance that mechanical vibration cause, make rotor eccentricity sweep that thorax, back end cover for motor eccentric wear, water cooling are bad to cause motor temperature rising etc.Generally speaking, when being out of order, hydraulic system all at first expects hydraulic pump, because pump is the power source of hydraulic system, hydraulic pump is hydraulic system " heart ", it converts mechanical energy to the hydraulic pressure energy, provide the fluid of enough flows and pressure to system, finish the needs of exercises to satisfy machinery, the action of all elements all will have relation with pump.The fault of in advance considering hydraulic pump is necessary, because can regularly trace fault rootstock by logic like this, helps fixing a breakdown.The hydraulic pump most common failure has: (1) not oil extraction; (2) vibration and noise; (3) pressure anomaly; (4) underfed or shakiness; (5) abnormal heating etc.

The present invention---can also predict and diagnose the fault of main motor, hydraulic power unit based on the method and apparatus of the continuous miner remote real-time failure forecast of immune neural net and diagnosis, can both well carry out remote live judgement fast and accurately to corresponding component, increased the Homes Using TV of Lian Caiji, greatly improve the reliability of Lian Caiji, guaranteed the efficient high yield of its workplace.

Claims (3)

1, a kind of continuous miner remote real-time failure forecast and diagnostic method, this method depends on host computer, slave computer, data communication apparatus and the change of on-the-spot sensing and send device, embedded the functional software of realizing failure prediction and diagnosis in the host computer, it is characterized in that this method may further comprise the steps:

Step 1, image data

Gather acceleration, vibration, flow, pressure, temperature data, and through the A/D conversion;

Step 2, deal with data

The input data are made normalized, and the scope that makes it adopts following formula between [0,1]:

$x_i^{\&prime;}=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}(\mathrm{HI}-\mathrm{LO})+\mathrm{LO}---\left(1\right)$

X wherein _iBe i data of measuring before the normalization, x _MaxBe { x _i, i=1,2 ..., the maximal value among the n; x _MinBe { x _i, i=1,2 ..., the minimum value among the n, Be the data after the normalization, HI=0.9, LO=0.1;

Step 3, generation detecting device carry out according to the following steps

(1) definition oneself string S is for containing limited character, and length is the set of the character string of L;

Oneself string is that the information data of L standard sign equipment state of the data pattern during real coding equipment operate as normal links together as the normal mode string, and chooses h and go here and there as the normal mode string $R_0^{\&prime;}=[X_1,X_2,\&CenterDot;\&CenterDot;\&CenterDot;,X_h],$ H=1,2, For certain fault mode i, i=1,2 ..., k links together the information data of the characterization device fault signature of the L of the fault mode of a real coding standard as the fault mode string

, choose the h string as the fault mode string $R_i^{\&prime;}=[X_{\mathrm{hi}+1},X_{\mathrm{hi}+2},\&CenterDot;\&CenterDot;\&CenterDot;,X_{\mathrm{hi}+h}],$ Obtain oneself's string $S=(R_0^{\&prime;},R_1^{\&prime;},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{\&prime;});$

(2) set matching threshold z

Obtain each detecting device collection, neural network training then by getting different matching threshold z respectively; By the graph of errors of neural metwork training process with select nominal situation at random and the fault diagnosis situation of the data detection immune neural net that various fault mode obtains, determine more excellent matching threshold z then; Adopt the Euclidean distance matching principle to characterize the affinity size, as the formula (2), if Euclidean distance d, shows that affinity reaches necessary requirement less than preset threshold z;

$d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z---\left(2\right)$

In the formula: d is an Euclidean distance; Ab _iCoordinate for antibody; Ag _iCoordinate for antigen; Z is matching threshold; L is the length of string;

(3) generate detecting device collection R _i

Generate detecting device collection D at random, to each detecting device among the D, with the normal mode string among the S

Carry out affinity according to formula (2) and calculate, wherein antibody in the formula and antigen are respectively the detected set D that the oneself goes here and there S and generates at random; If d less than preset threshold z, illustrates that affinity is enough big, then put into detecting device collection R ₀In, if can not satisfy condition, continue and first kind of fault mode string

Carry out affinity and calculate, if d, then puts into detecting device collection R less than preset threshold z ₁In, if can not mate, will continue and following a kind of fault mode string Carry out affinity and calculate, if d less than preset threshold z, puts into detecting device collection R so ₂In, if can not mate, continue circulation, up to mating with last a kind of fault mode; Repeat this process till reaching needed detector number M;

(4) generation forecast detecting device collection

Getting matching threshold is z+a, calculates detecting device and first kind of fault mode string among the D

Affinity, if d less than preset threshold z+a and more than or equal to z, then puts into the detecting device collection

In, if can not mate, will continue and following a kind of fault mode string

Carry out affinity and calculate, if d less than preset threshold z+a and more than or equal to z, puts into the detecting device collection so

In, if can not mate, continue circulation, up to mating, obtain the detecting device collection with last a kind of fault mode $R_a=(R_1^a,R_2^a,\&CenterDot;\&CenterDot;\&CenterDot;,R_k^a);$ Repeat whole process till reaching needed detector number M; In like manner, get matching threshold and be respectively z+2a, z+3a, a ∈ R, the detecting device collection of each fault mode of different matching threshold correspondences is respectively $R_{2a}=(R_1^{2a},R_2^{2a},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{2a})$ With $R_{3a}=(R_1^{3a},R_2^{3a},\&CenterDot;\&CenterDot;\&CenterDot;,R_k^{3a});$ The affinity of different matching threshold detecting devices is calculated by following European formula control:

$z\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+a---\left(3\right)$

$z+a\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+2a---\left(4\right)$

$z+2a\&le;d=\sqrt{\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{ab}}_i-{\mathrm{ag}}_i)}^2}<z+3a---\left(5\right)$

In the formula: a is the regulated quantity of matching threshold; Z is matching threshold; D is an Euclidean distance;

The study of step 4, immune neural net and training;

With the detecting device collection R that trains ₀Normal sample as neural network training; With the detecting device collection R that trains ₁, R _a, R _2aAnd R _3aAs the fault sample of neural network training, i=1 wherein, 2 ..., k determines that output 1 is fault, 0 is normal; Because some signal can change before fault took place, what answer is contingent fault sample in the future so matching threshold is the detecting device set pair of z+a, z+2a, z+3a, the fail result of the output of its neural network is indicating the fault that equipment promptly will take place, and the confidence level that will break down is represented in each matching threshold;

Step 5, failure prediction and fault diagnosis;

The neural network that trains is used for the failure prediction and the diagnosis of equipment, judges according to output vector which kind of fault has taken place;

Step 6, demonstration output result.

2, continuous miner remote real-time failure forecast according to claim 1 and diagnostic method is characterized in that the study of the immune neural net in the described step 4 and training detailed process are as follows:

(1) determining of network structure and learning parameter: determine that according to input vector the input layer of three-layer neural network counts n, determines that according to the fault mode kind output layer node counts q, rule of thumb and by considering the study number of times and the discrimination of network behind the training study, comprehensively relatively the back determines that hidden layer node counts p and weight convergence factor ζ and error convergence factor delta;

(2) network status initialization: give each connection weight W _Ij, V _JtAnd threshold value θ _j, γ _tGive the random value between [1 ,+1];

(3) provide the neural network learning sample right: selected characteristic sample data vector $A_k=[a_1^k,a_2^k,\&CenterDot;\&CenterDot;\&CenterDot;,a_n^k]$ And object vector $Y_k=[y_1^k,y_2^k,\&CenterDot;\&CenterDot;\&CenterDot;,y_q^k]$ Offer neural network;

(4) begin study, each sample carried out following operation:

1. each neuronic output of computational grid hidden layer and output layer

Hidden layer output $b_j=f(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{W}_{\mathrm{ij}}\&CenterDot;a_i-{\mathrm{\&theta;}}_j)---\left(7\right)$

Output layer output $c_t=f(\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{V}_{\mathrm{jt}}\&CenterDot;b_j-{\mathrm{\&gamma;}}_t)---\left(8\right)$

J=1 in the formula, 2 ..., p;

t＝1，2，…，q；

W _IjThe connection weight of-input layer i unit to hidden layer j unit;

θ _jThe threshold value of-hidden layer j unit;

V _Jt-hidden layer j unit to output layer t unit connection weight;

γ _t-output layer t cell threshode;

F-S type activation function $f\left(x\right)=\frac{1}{1+\exp (-x)};$

2. calculation training error

The hidden layer training error $e_j^k=[\underset{i=1}{\overset{q}{\mathrm{\&Sigma;}}}{d}_t^k\&CenterDot;V_{\mathrm{jt}}]\&CenterDot;b_j\&CenterDot;(1-b_j)---\left(9\right)$

The output layer training error $d_t^k=(y_t^k-c_t)\&CenterDot;c_t\&CenterDot;(1-c_t)---\left(10\right)$

In the formula

-desired output;

c _t-actual output;

b _jThe output of-hidden layer j unit;

3. revise weights

Input layer is to the new connection weight between the hidden layer:

$W_{\mathrm{ij}}(N+1)=W_{\mathrm{ij}}\left(N\right)+\mathrm{\&beta;}\&CenterDot;e_j^k\&CenterDot;a_i^k---\left(11\right)$

Hidden layer is to the new connection weight between the output layer:

$V_{\mathrm{jt}}(N+1)=V_{\mathrm{jt}}\left(N\right)+\mathrm{\&alpha;}\&CenterDot;d_t^k\&CenterDot;b_j---\left(12\right)$

In the formula

Respectively from formula (9) and (10);

N-study number of times;

α-learning coefficient, 0＜α＜1;

β-learning coefficient, 0＜β＜1;

(5) whether satisfy | b _j(N+1)-b _j(N) |＜ζ, carried out for (6) step if satisfy, otherwise turned back to for (4) step;

(6) whether satisfy | b _j(N)-Y _k|＜δ goes on foot if satisfy then carry out (7), otherwise turned back to for (2) step;

(7) weights and the threshold data that obtains deposited in the knowledge base of being set up.

3, realize the device of described continuous miner remote real-time failure forecast of claim 1 and diagnostic method, it is characterized in that this device comprises host computer, data communication apparatus, slave computer, on-the-spot sensing becomes send device, on-the-spot sensing becomes send device to comprise acceleration, vibration, flow, pressure, temperature sensor, slave computer comprises one or more Programmable Logic Controllers, what data communication apparatus adopted is the CAN bus, wherein data communication apparatus respectively with host computer, slave computer links to each other, slave computer send device to link to each other with on-the-spot sensing change, adopt the sensing change of the on-the-spot erecting stage of machine in company and send device, on-the-spot sensing becomes the sensor that send in the device signal of gathering is delivered to slave computer, slave computer regularly is sent to host computer with acquired signal by the CAN bus, host computer carries out analyzing and processing to data again, carries out failure prediction and diagnosis then.

Images