CN105572492A - City rail train auxiliary inverter fault diagnosis device - Google Patents

Abstract

The invention provides a city rail train auxiliary inverter fault diagnosis device. Data collected by sensors is preprocessed by an auxiliary inverter system signal preprocessor; then, the preprocessed data is subjected to noise reduction through an extended Kalman filter, and effective diagnosis data is extracted through an arithmetic unit to remove device abnormity data due to unstable operation of a train; next, the extracted data is decomposed into the sum of a series of mode functions; the decomposed mode functions are subjected to energy torque extraction, and energy torques are input to a BP neural network; and finally, the running status of an auxiliary inverter of a city rail train is identified. The device can effectively identify the running status of the auxiliary inverter of the city rail train, thereby providing a quick and effective method for fault monitoring and diagnosis of the auxiliary inverter of the city rail train.

Description

A kind of municipal rail train subordinate inverter trouble-shooter

Technical field

The invention belongs to municipal rail train security technology area, be specifically related to the feature extraction of a kind of municipal rail train subordinate inverter and state identification method and device.

Background technology

While urban track traffic fast development, the safety problem of its train operation also receives much concern.As the complicated operational system of integration, the state of train and performance can constantly reduce in operational process, thus form potential safety hazard gradually, if do not got rid of in time, possibility will be accidents caused.In actual operation process, the accident that train fault causes happens occasionally.And the equipment of China's rail transit train monitoring at present and technical development relatively lag behind, collection and the safeguards technique of failure message respectively become system, transversely independent of one another in the vertical, lack the train supervision of integration, fault diagnosis and early warning technology, the support of the personalized maintenance program of train can not be met, and overhaul efficiency is low, has a strong impact on train movement capacity.

In municipal rail train operation process, the fault that subordinate inverter is more serious will affect its normal operation, need the proteges of the powerful who stay with their benefactions like parasites or need to exit operation behind terminus, become one of important factor in order hindering the operation of municipal rail train normal safe.Inverter method for diagnosing faults of the prior art directly carries out analyzing and diagnosing for the raw data gathered mostly, and in fact, because train is often in irregular operation in operational process, this irregular operation often produces interference to contravariant equipment, and then cause the diagnostic result of equipment may be caused by train operation instability, instead of cause due to equipment faults itself, the Monitoring Data of the contravariant equipment therefore making sensor collect directly can not be used for diagnostic analysis, and this interference greatly reduces the precision of existing fault diagnosis technology.

Summary of the invention

The present invention is based on above problem, for the deficiency of existing municipal rail train subordinate inverter fault diagnosis technology, object gets rid of the unstable device data abnormal conditions caused of train operation, and the data that diagnostic device is obtained truly can reflect equipment state, improve diagnostic accuracy.The municipal rail train subordinate inverter trouble-shooter provided is provided, for municipal rail train normal operation provides technical guarantee, the following technical scheme of concrete employing:

Municipal rail train subordinate inverter trouble-shooter comprises: signal picker, auxiliary inversion system signal pretreater, fault diagnosis device, data-carrier store, display, radio transmitting device, and signal picker is for gathering subordinate inverter Monitoring Data; Auxiliary inversion system signal pretreater comprises prime amplifier, anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter and logic control circuit; The data of collection are passed to prime amplifier by signal picker, data after prime amplifier process are successively through anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter, and pretreated data are delivered to fault diagnosis device by A/D converter; Logic control circuit and prime amplifier, anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter, fault diagnosis device are all electrically connected; Fault diagnosis device comprises extended Kalman filter, verifier, arithmetical unit, corrector, fault memorizer, fail result maker; After fault diagnosis device receives pretreated data, carry out following steps successively: (i) extended Kalman filter carries out noise reduction to data, then the data after noise reduction are sent to arithmetical unit; (ii) arithmetical unit calculates the data after noise reduction, calculates the change amplitude in its predetermined amount of time, is recorded by the timing node that change amplitude is greater than first threshold, and is stored in fault memorizer by data corresponding for described timing node; (iii) distribution of data in arithmetical unit statistical computation fault memorizer, and the data deviate of the intermediate value relative to distribution being greater than Second Threshold extract, and the data extracted are sent to fail result maker; (iv) first the data of the arithmetical unit received are decomposed into a series of mode function sum by fail result maker, step is as follows: (1) determines the Local Extremum that data are all, then take cubic spline line all local maximums, minimum point to be coupled together, form coenvelope line and lower envelope line; (2) average of coenvelope line and lower envelope line is designated as m ₁t (), then by signal x (t) after noise reduction and average m ₁t () obtains h ₁(t), h ₁(t)=x (t)-m ₁(t); (3) if h ₁t () does not meet mode function condition, h ₁t () as raw data, and repeats step (1) ~ (2), obtain the average m of lower envelope ₁₁(t), then h ₁₁(t)=h ₁(t)-m ₁₁(t); If h ₁₁t () does not still meet mode function condition, then repetitive cycling k time, until h _1kt () satisfies condition, first of x (t) modal components is designated as c ₁(t), c ₁(t)=h _1k(t); (4) by c ₁t () is separated and obtains r from x (t) ₁(t): r ₁(t)=x (t)-c ₁(t); By r ₁t () repeats above step as new raw data, obtain the 2nd the modal components c of x (t) ₂t (), repeats to obtain r n time _nt () is that a monotonic quantity terminates when therefrom can not extract modal components again; Signal x (t) then after noise reduction is finally decomposed into a series of mode function sum: then extract energy square to the mode function decomposed, step is as follows: 1) calculate corresponding energy square E according to following formula ₁, E ₂... E _m: wherein m is total sampling number, and k is sampled point, and Δ t is the sampling period; 2) structural attitude vector T: T=[E ₁e ₂e _m], T is normalized, after normalization, individual features vector T ' be: T'=[E ₁/ EE ₂/ E ... Em/E], wherein, finally by energy square input BP neural network, finally draw fault diagnosis result.

Preferably, if diagnostic result is for there being fault, diagnostic result is on the one hand transferred to data-carrier store by display screen Realtime Alerts by fault maker, on the other hand failure message is reached vehicle-mounted display terminal and center, ground by Ethernet.

Preferably, if diagnostic result is not for having fault, and verifier detects that fault has appearred in inverter, then corrector is based on the storage data correction first threshold in fault memorizer and Second Threshold; If diagnostic result is not for having fault, and verifier does not detect that inverter occurs that then corrector does not carry out the correction of first threshold and Second Threshold extremely.

Preferably, radio transmitting device comprises database, modulator and each radio node, and diagnostic result data are sent to database by modulator by radio node.

The present invention has technique effect useful as follows:

(1) process of the arithmetical unit in pretreater and fault diagnosis device, eliminates the interference that train operation causes the Monitoring Data of auxiliary inversion equipment to fluctuate, improves fault diagnosis precision.

(2) first adopt is mode function by data decomposition, then extracts energy square, finally energy square is imported the diagnostic method of neural network, improves fault diagnosis precision.

(3) when occurring that diagnostic result does not conform to the actual conditions, Rule of judgment is revised immediately until diagnostic result conforms to actual, for diagnostic process provides reliable guarantee.

(4) on the one hand by display screen Realtime Alerts; Raw data and diagnostic result reach ground by onboard wireless transmission platform regularly on the other hand; Convenient ground center analysis layer can under reach ground raw data carry out specifically analysis.

Accompanying drawing explanation

Fig. 1 is the structure composition diagram of apparatus of the present invention.

Fig. 2 is data pre-processor structure composition diagram of the present invention.

Fig. 3 is fault diagnosis system process flow diagram of the present invention.

Fig. 4 is mode function decomposition process figure.

Fig. 5 is neural network failure diagnostic flow chart.

Embodiment

Municipal rail train subordinate inverter trouble-shooter comprises sensor, auxiliary inversion system signal pretreater, fault diagnosis device, data-carrier store, display, radio transmitting device, and sensor is for gathering subordinate inverter Monitoring Data.Subordinate inverter trouble-shooter by gathering voltage, current sensor acquisition subordinate inverter Monitoring Data, and carries out diagnostic analysis to data, realizes fault identification, fault level classification, draws diagnostic result.Concrete structure as shown in Figure 1.

Quantity installed by sensor be the total number of probes of every train is 16.Concrete inventory is as following table:

Table 1 Equipments Setting schedule of quantities

Sequence number	Device name	Unit	Configuration quantity
				1	Voltage sensor	Row/only	8
2	Current sensor	Row/only	8
				3	Diagnosis main frame	Row/platform	2

Note: row/platform represents every train installed equipment number of units.

Increase each one of 110V DC control supply voltage and current sensor at each A car low-voltage cabinet, gather electric current and the voltage data of 110V DC control supply, to analyze the transport properties of 110V DC control supply voltage.

Increase each three of 380V three-phase alternating-current supply voltage and current sensor at each A car low-voltage cabinet, gather three-phase current and the voltage data of 380V three-phase alternating-current supply, to analyze the transport properties of 380V three-phase alternating-current supply.

Other hardware main performance index are as follows:

Table 2 Specifeca tion speeification

(2) power supply: 110VDC

(3) cabinet: shell white, the inside fixed mount of PXI bus and base

(4) touch-screen, main performance index is as follows:

Table 3 Specifeca tion speeification

Auxiliary inversion system signal preprocessing part is made up of prime amplifier, anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter and logic control circuit, as shown in Figure 2.

Subordinate inverter signal preprocessor main performance index is as follows:

Table 4 Specifeca tion speeification

Fault diagnosis device comprises extended Kalman filter, verifier, arithmetical unit, corrector, fault memorizer, fail result maker, and its diagnostic analysis flow process as shown in Figure 3.Method step is as follows:

I () extended Kalman filter carries out noise reduction to data, then the data after noise reduction are sent to arithmetical unit.Often be subject to the impact of neighbouring vibration equipment and other extraneous factors in equipment running process, in actual applications, denoiser needs to carry out noise reduction process to signal, removes interfere information.

(ii) arithmetical unit calculates the data after noise reduction, calculates the change amplitude in its predetermined amount of time, is recorded by the timing node that change amplitude is greater than first threshold, and data corresponding for described timing node is stored in fault memorizer.

(iii) distribution of data in arithmetical unit statistical computation fault memorizer, and the data deviate of the intermediate value relative to distribution being greater than Second Threshold extract, and the data extracted are sent to fail result maker.

Eliminate train interference to equipment under the state of irregular operation by above two steps, operation result remains the data of the true projection equipment running status of energy.

(iv) first the data of the arithmetical unit received are decomposed into a series of mode function sum by fail result maker, and mode decomposition is a kind of nonstationary random response method, and are the adaptive decomposition methods based on data itself.A sophisticated signal is decomposed into some mode functions by it, based on following hypothesis: any sophisticated signal is all made up of a series of mode function, each mode function, its extreme point and zero crossing number must equal or difference at most can not more than one, and up and down envelope for time shaft Local Symmetric.As shown in Figure 4, concrete steps are its screening process:

(1) determine the Local Extremum that data are all, then take cubic spline line all local maximums, minimum point to be coupled together, form coenvelope line and lower envelope line;

(2) average of coenvelope line and lower envelope line is designated as m ₁t (), then by signal x (t) after noise reduction and average m ₁t () obtains h ₁(t), h ₁(t)=x (t)-m ₁(t);

(3) if h ₁t () does not meet mode function condition, h ₁t () as raw data, and repeats step (1) ~ (2), obtain the average m of lower envelope ₁₁(t), then h ₁₁(t)=h ₁(t)-m ₁₁(t);

If h ₁₁t () does not still meet mode function condition, then repetitive cycling k time, until h _1kt () satisfies condition, first of x (t) modal components is designated as c ₁(t), c ₁(t)=h _1k(t);

(4) by c ₁t () is separated and obtains r from x (t) ₁(t): r ₁(t)=x (t)-c ₁(t);

By r ₁t () repeats above step as new raw data, obtain the 2nd the modal components c of x (t) ₂t (), repeats to obtain r n time _nt () is that a monotonic quantity terminates when therefrom can not extract modal components again; Signal x (t) then after noise reduction is finally decomposed into a series of mode function sum:

$x\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{c}_i\left(t\right)+r_n\left(t\right).$

The decomposable process of EMD is the process of " screening " in fact, but the envelope mean value of the mode function separated from actual signal can not be zero, and this just needs to arrange a condition and stops screening.Usually, whether decide screening according to the size of standard deviation SD and stop, SD is by the h sieving out _{1 (k-1)}and h _1kobtain, namely

$SD=\underset{t=0}{\overset{T}{\mathrm{\Σ}}}\[\frac{{|h_{1(k-1)}\left(t\right)-h_{1k}\left(t\right)|}^2}{h_{1(k-1)}^2\left(t\right)}\]$

Wherein T is the total length of discrete signal sequence, and the value of SD is between 0.2 to 0.3.

Rilling improves on the basis of the above, proposes

$\mathrm{\δ}\left(t\right)=\frac{\left|e_{\max }+e_{\min }\right|}{\left|e_{\max }-e_{\min }\right|}$

Wherein e _maxfor the coenvelope of signal, e _minfor the lower envelope of signal.

The screening stop condition that Rilling proposes more can reflect the average characteristics of mode function, therefore the present invention's screening stop condition of adopting Rilling to propose.

Then energy square is extracted to the mode function decomposed, 1) original signal is analyzed, draw n respective mode state function c _i(t), i=1,2,3 ..., n;

2) from analysis result, select corresponding mode function component to calculate energy square, the formula of IMF energy square is

$E_i={\∫}_{-\mathrm{\∞}}^{+\mathrm{\∞}}t{\left|c_i\left(t\right)\right|}^{2}dt$

If discrete signal, then corresponding energy square is:

$E_i=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}(k\·\mathrm{\Δ}t){\left|c_i(k\·\mathrm{\Δ}t)\right|}^2$

Wherein m is total sampling number, and k is sampled point, and Δ t is the sampling period.Corresponding energy square E is calculated to selected IMF ₁, E ₂...

3) structural attitude vector T:

T＝[E ₁E ₂…E _n]

If when corresponding energy square value is larger, need be normalized T, after normalization, individual features vector T ' be:

T'＝[E ₁/EE ₂/E…E _n/E]

Wherein

$E={\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\left|E_i\right|}^2\right)}^{\frac{1}{2}}$

Energy square not only reflects the size of IMF energy, also reflects energy distribution situation in time.

Finally by energy square input BP neural network, finally draw fault diagnosis result, diagnostic process as shown in Figure 5.Neural network is used for auxiliary when being fault diagnosis, generally many samples in each Pattern Class will be provided as training sample to neural network.Neural network, through study, can not only identify trained sample, and can identify the sample do not occurred.This ability of neural network is called popularization power or generalization ability.Neural network is better than extrapolation point for the generalization ability of interpolative data point, thus general with a large amount of training sample neural network trainings to obtain good generalization ability.

When by neural fusion fault diagnosis, be equivalent to set up in neural network and inquiry fault dictionary.Owing to employing neural network, the foundation of fault dictionary is completed by the study of neural network to fault dictionary characteristic feature; The inquiry of dictionary is then realized the association of fault signature by neural network.As long as can meet the fault of fault isolation condition in dictionary method, its feature is inevitable independent of other faults, just can use the diagnosis of neural fusion fault.Therefore, the diagnosis of neural network is adopted all quite to have superiority in the foundation of dictionary and inquiry.

Auxiliary inversion system diagnosability algorithm based on BP neural network comprises failure modes identification three processes of the structure of BP neural network, the training of BP neural network and BP neural network.When application BP neural network carries out failure modes, need the auxiliary inversion system failure characteristic parameter of extraction to be divided into two groups, one group as training sample, one group as treating diagnostic sample.Wherein training sample is used for training BP neural network, treats that diagnostic sample is used for testing the BP neural network trained, with the classification capacity of supervising network.BP neural network is utilized to carry out the step of fault diagnosis as shown in Figure 5 to auxiliary inversion system.

The aspects such as the number of plies of network, every layer of neuronic nodes, initial value and learning rate should be considered when carrying out the design of BP neural network in the fault diagnosis of auxiliary inversion system.

(1) number of plies of network.Three layers of BP neural network can approach any continuous function, the mapping tieing up m dimension from n can be realized, increase the number of plies to improve precision further, reduce error, but also can increase the learning training time of network, and make network complicated, and suitably increase hidden layer node number and also can improve error precision.In the fault diagnosis of auxiliary inversion, the BP neural network number of plies generally chooses three layers just can obtain good Nonlinear Mapping effect.

(2) hidden layer node number.Hidden layer node number is difficult to reach training requirement very little, and number can increase the training time again too much, can with reference to following experimental formula:

l＝2n+1

$l=\sqrt{n+m}+a$

l＝log ₂n

Wherein l is the neuron node number of hidden layer, and n is the neuron node number of input layer, and m is the neuron node number of output layer, and a is the integer between 1 to 10.In specific design, an interval can be suitably set, hidden layer node number is chosen successively in interval range, carries out training and the error of calculation, then select the node number corresponding to minimum error as the node number of BP neural network hidden layer to BP neural network.

(3) initialization weights.Auxiliary inversion system is a nonlinear system, initial value choose for the training study time length, can network restrain and reach local minimum much relations.Initial value too small or cross big city pace of learning is had an impact, therefore the initial value of weights generally should select equally distributed decimal empirical value, can between-1 to 1 random selecting.

(4) learning rate.Weights variable quantity in each circuit training of BP neural network is determined by learning rate, and learning rate is larger, larger to the amendment of weights, and e-learning speed is faster.But excessive learning rate will make weights learning process produce concussion, and too small learning rate can make network convergence excessively slow, and weights are difficult to tend towards stability.Generally tend to choose less learning rate, selection range, between 0.01 to 0.8, is stablized with keeping system.

(5) anticipation error.In the training process of BP neural network, expected error value can by determining a suitable value after comparative training, suitable is determine relative to the nodes of required hidden layer, and less expected error value will obtain by increase training time and node in hidden layer.Generally as a comparison, can train the network of two different anticipation errors simultaneously, therefrom determine to adopt one of them network.

If diagnostic result is for there being fault, diagnostic result is on the one hand transferred to data-carrier store by display screen Realtime Alerts by trouble-shooter, on the other hand failure message is reached vehicle-mounted display terminal and center, ground by Ethernet.

Deviation is there is in order to prevent diagnostic result, then need when diagnostic result and device physical status be inconsistent, judge condition in continuous correction diagnostic process, therefore, be provided with the verifier for inspection machine state in device, and revise the truing device that fault diagnosis passes judgment on condition, if diagnostic result is not for having fault, and verifier detects that fault has appearred in inverter, then corrector is based on the storage data correction first threshold in fault memorizer and Second Threshold; If diagnostic result is not for having fault, and verifier does not detect that inverter occurs that then corrector does not carry out the correction of first threshold and Second Threshold extremely.

Claims (4)

1. a municipal rail train subordinate inverter trouble-shooter, it comprises: signal picker, auxiliary inversion system signal pretreater, fault diagnosis device, data-carrier store, display, radio transmitting device, is characterized in that:

Signal picker is for gathering subordinate inverter Monitoring Data;

Auxiliary inversion system signal pretreater comprises prime amplifier, anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter and logic control circuit;

The data of collection are passed to prime amplifier by signal picker, data after prime amplifier process are successively through anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter, and pretreated data are delivered to fault diagnosis device by A/D converter;

Logic control circuit and prime amplifier, anti-aliasing low-pass filter, sample/hold circuit and multiway analog switch, programmable amplifier, A/D converter, fault diagnosis device are all electrically connected;

Fault diagnosis device comprises extended Kalman filter, verifier, arithmetical unit, corrector, fault memorizer, fail result maker;

After fault diagnosis device receives pretreated data, carry out following steps successively:

I () extended Kalman filter carries out noise reduction to data, then the data after noise reduction are sent to arithmetical unit;

(ii) arithmetical unit calculates the data after noise reduction, calculates the change amplitude in its predetermined amount of time, is recorded by the timing node that change amplitude is greater than first threshold, and is stored in fault memorizer by data corresponding for described timing node;

(iii) distribution of data in arithmetical unit statistical computation fault memorizer, and the data deviate of the intermediate value relative to distribution being greater than Second Threshold extract, and the data extracted are sent to fail result maker;

(iv) first the data of the arithmetical unit received are decomposed into a series of mode function sum by fail result maker, and step is as follows:

(1) determine the Local Extremum that data are all, then take cubic spline line all local maximums, minimum point to be coupled together, form coenvelope line and lower envelope line;

(2) average of coenvelope line and lower envelope line is designated as m ₁t (), then by signal x (t) after noise reduction and average m ₁t () obtains h ₁(t), h ₁(t)=x (t)-m ₁(t);

(3) if h ₁t () does not meet mode function condition, h ₁t () as raw data, and repeats step (1) ~ (2), obtain the average m of lower envelope ₁₁(t), then h ₁₁(t)=h ₁(t)-m ₁₁(t);

If h ₁₁t () does not still meet mode function condition, then repetitive cycling k time, until h _1kt () satisfies condition, first of x (t) modal components is designated as c ₁(t), c ₁(t)=h _1k(t);

(4) by c ₁t () is separated and obtains r from x (t) ₁(t): r ₁(t)=x (t)-c ₁(t);

By r ₁t () repeats above step as new raw data, obtain the 2nd the modal components c of x (t) ₂t (), repeats to obtain r n time _nt () is that a monotonic quantity terminates when therefrom can not extract modal components again; Signal x (t) then after noise reduction is finally decomposed into a series of mode function sum:

$x\left(t\right)=\underset{i=1}{\overset{n}{\Σ}}{c}_i\left(t\right)+r_n\left(t\right);$

Then extract energy square to the mode function decomposed, step is as follows:

1) corresponding energy square E is calculated according to following formula ₁, E ₂... E _m:

$E_i=\underset{k=1}{\overset{m}{\Σ}}(k\·\mathrm{\Δ}t)|c_i(k\·\mathrm{\Δ}t){|}^2$

Wherein m is total sampling number, and k is sampled point, and Δ t is the sampling period;

2) structural attitude vector T:

T＝[E ₁E ₂…E _m]

T is normalized, after normalization, individual features vector T ' be:

T'＝[E ₁/EE ₂/E…Em/E]

Wherein, $E={\left(\underset{i=1}{\overset{m}{\Σ}}\right|E_i{|}^2)}^{\frac{1}{2}};$

Finally by energy square input BP neural network, finally draw fault diagnosis result.

2. municipal rail train subordinate inverter trouble-shooter as claimed in claim 1, it is characterized in that, if diagnostic result is for there being fault, diagnostic result is on the one hand transferred to data-carrier store by display screen Realtime Alerts by fault maker, on the other hand failure message is reached vehicle-mounted display terminal and center, ground by Ethernet.

3. municipal rail train subordinate inverter trouble-shooter as claimed in claim 1 or 2, it is characterized in that, if diagnostic result is not for having fault, and verifier detects that fault has appearred in inverter, then corrector is based on the storage data correction first threshold in fault memorizer and Second Threshold; If diagnostic result is not for having fault, and verifier does not detect that inverter occurs that then corrector does not carry out the correction of first threshold and Second Threshold extremely.

4. the municipal rail train subordinate inverter trouble-shooter as described in claim 1-3, is characterized in that, radio transmitting device comprises database, modulator and each radio node, and diagnostic result data are sent to database by modulator by radio node.