RU150919U1 - PERFORMANCE FORECASTING DEVICE FOR MULTI-PARAMETER ELECTROMECHANICAL SYSTEMS - Google Patents

Abstract

A device for predicting the operability of multi-parameter electromechanical systems (EMCs), comprising a measuring unit consisting of an information processing unit for normalizing input data from sensors and extracting a reference value for each measured parameter by obtaining temporary signals of the output parameters of the EMC and reference arrays of output signals for each monitored parameter, and a software calculator made of series-connected processor Matora, quad, integration unit, a computing unit for finding the coefficients neural nonlinear operators and their transfer to the neural network for their formation testing and the maximum potential corresponding to the predicted maintenance of EMC-specific parameter to be measured.

Description

The proposed technical solution relates to the field of diagnostic equipment and is a device for predicting the health of multi-parameter electromechanical systems (EMC).

There is a method of diagnosing an object consisting of series-connected functional blocks covered by feedbacks, and a device for its implementation according to the copyright certificate SU No. 1667013 MPK6 G05B 23/02, publ. 07/30/1991. A device for implementing this method comprises functionally connected function blocks, in particular a group of tolerance control sensors, a maximum calculation unit, a delay element, a summing unit, a conversion and storage unit for the values of the object state parameters being monitored, a key block, an OR element.

The disadvantage of this device is that it does not contain the appropriate blocks for predicting the state parameters of the object.

Also known is a method for diagnosing objects and a device for its implementation (RF patent No. 2239869, IPC 6 G07C 11/00, publ. 2004). To implement the described method, a device for diagnosing an object is used, consisting of a summing unit, a conversion and storage unit for monitoring the state parameters of an object, a block of keys, an OR element, which contains, in addition, a unit for recording operating times of functional blocks of a controlled object, a unit for setting functional resource times blocks, clock generator, block for specifying the period of following clock pulses, two time delay elements, element for selecting the minimum value eniya resource difference and the operating time, the block prediction state parameters calculation block prediction time calculating unit current operating time, the determination unit outputs the controlled parameters for valid values and probability calculating unit uptime functional blocks controlled object.

The main disadvantage of this device is the lack of a measuring system, that is, a set of sensors characterizing the object of diagnosis.

The closest technical solution is the device for vibro-acoustic diagnostics of cyclically functioning objects (RF patent No. 2239869, IPC 7 G07C 13/00, publ. 20.12.2005). The device for vibro-acoustic diagnostics of cyclically functioning objects includes one or several channels for extracting measurement information, each of which contains a measuring path for controlling a vibro-acoustic signal, consisting of a series-connected vibration sensor and a first adjustable amplifier, a speed monitoring path, consisting of a series-connected speed sensor and the second adjustable amplifier, part of the path for the formation of an electrical signal, including digital a log converter, a program-controlled selective filter based on resonant circuits, a third adjustable amplifier and controller, a programmable calculator, an indicator and a recording device, inputs that are associated with the corresponding outputs of a programmable computer, characterized in that it contains a predictive device, inputs of the conversion and storage unit the values of each of the monitored parameters, the state of the object of which is connected to the corresponding outputs of the paths of the monitored parameters , the inputs of the unit for recording the operating time of the functional blocks of the controlled object by the time the device and the controlled object are turned on are connected to the corresponding outputs of the controller, the inputs of the unit for setting the time of the resource of the functional blocks of the controlled object are connected to the corresponding outputs of the controller, the input of the block for setting the period of following clock pulses is connected to the corresponding output of the controller , and the outputs of the unit for calculating the current operating hours of the functional blocks of the controlled object connected to the corresponding inputs of the controller, and in each channel for the extraction of measurement information, an additional noise measurement path is introduced functionally complete and spatially located in one place of the blocks of the monitored object, consisting of a noise sensor and a fourth adjustable amplifier connected in series, and the electrical signal generation path is supplemented by a proximity switch and an adjustable digital filter, while the switch with its inputs is connected to the outputs of the control path in the acoustic signal, the speed control path, the noise measurement path of part of the object blocks and the corresponding controller output, the switch output is connected to the first input of the adjustable digital filter, the second filter input is connected to the output of the third adjustable amplifier, and its third input is connected to the corresponding controller output, output An adjustable digital filter is connected to the second input of the controller.

The disadvantage of the closest technical solution is the low accuracy of forecasting, which can occur when predicting EMC in the face of uncertainty, describing their state of parameters and in the presence of a lack of a priori information.

The objective of the proposed technical solution is to increase the reliability and accuracy of predicting performance, reducing the error in the conditions of uncertainty of the EMC and environmental parameters and insensitivity to the lack of a priori information.

The task is achieved in that the device for predicting the health of multi-parameter electromechanical systems (EMC), contains a measuring unit, consisting of an information processing unit designed to normalize the input data from the sensors and allocate a reference value for each measured parameter by obtaining temporary signals of output parameters EMC and reference arrays of output signals for each controlled parameter, and a software calculator made from the last sequences coupled processor, adder, quad, integration unit, a computing unit for finding the coefficients neural nonlinear operators and their transfer to the neural network for their formation testing and the maximum potential corresponding to the predicted maintenance of EMC-specific parameter to be measured.

In FIG. depicts a structural diagram for predicting the health of multi-parameter electromechanical systems.

The device for predicting the health of multi-parameter electromechanical systems 1 contains a measuring unit 2, consisting of N sensors for measuring the values of the controlled parameters of the EMC, the outputs of which are connected to the inputs of the information processing unit 3, the outputs of which are connected to the input of a software computer 4, consisting of a series-connected processor 5, performing the decomposition function of the input signal by Legendre polynomials, adder 6, subtracting from the received N signals N standards for each a measured parameter, a quadrator 7, which executes the squaring of the obtained difference, integration unit 8, which integrates the forecast of each of the N differences on the observation interval, and a computing unit 9, which finds the corresponding coefficients, the outputs of which are connected to the inputs of the Kohonen neural network 10, consisting of from the cascade connection of two layers of the distribution layer and the competitive layer.

A device for predicting the operability of multi-parameter electromechanical systems works as follows. First, the input array of N measured EMC signals 1 is processed by Legendre operators by expanding it in a series according to the number of input (measured) parameters, then the N standards, previously also transformed by Legendre operators, are subtracted from the received N signals, followed by squaring and integration over the interval observing the forecast of each of the N differences, and then, the final processing of the obtained functionals is carried out by a neural network (NS) representing the Kohonen network 10 with a cascade distribution and a layer of competitive layer, by providing the closest to a given metric signal indicative of predictive maintenance state EMC. Prediction of EMC operability in this technical solution occurs by assigning a temporary array of output (measured) EMC signals to a certain class of technical conditions characterizing one or another EMC behavior in the future, by neural network processing of the totality of EMC parameters measured at different intervals of operation.

It is possible to imagine the output parameters of the EMC Z (t) as a vector in an N-dimensional Euclidean space, while the Cartesian coordinates of the end of the vector are real numbers z ₁ , z ₂ , ..., z _{n that} are signs of future EMC technical states and an input for a predictor.

The solutions to the “operational - non-operational” problem involves the introduction for each class of forecast states of the standard (the reference array of output signals Z _H (t)), which can be understood as some average nominal array of output signals of the corresponding class. Then, comparing the current array of signals Z (t) with the reference one will determine the degree of similarity between the vector of the obtained predicted state of the EMC and the vector of the reference (nominal) state of the EMC in the future.

, где Z_Hji - i-я компонента вектора эталона j-го класса; Z_i - i-я компонента вектора Z классифицируемого состояния.As a distinguishing norm of the distances between two points (coordinates of the ends of the vectors), which display the various predicted technical states of the EMC, the Euclidean norm is introduced:

where Z _Hji is the i-th component of the standard vector of the j-th class; Z _i is the ith component of the vector Z of the classified state.

,And with neural network transformation we get:

,

moreover, all arrays (including the reference one) are subjected to the same NS nonlinear transformation. The type of this transformation for each class is determined in the process of minimizing the distance to the standard within a given class and simultaneously maximizing the distance to the same standard for arrays of other classes.

But if the observed signals at the base of the forecast are not distinguishable, but lead to different forecasts, then it is necessary to first introduce a transformation that separates them, and then the NS can recognize them.

The condition of complete distinguishability (observability) of signals of two classes (“norm - not norm”) can be written in the following form:

According to the functional J, a nonlinear transformation is carried out, which is then intended for clustering by a Kohonen layer. Without preliminary transformation, the classification itself is impossible, forecasts will be indistinguishable, because arrays of signals of both classes (“norm - not norm”) will be mixed and selection of one of any class will be impossible.

The desired values of the elements of the neural network operator φ, i.e. coefficients {A _i } can be obtained by taking derivatives with respect to A _i and equating them to zero, since under this condition the elements of the set {A _i } correspond to the maximum distinguishability of the feature vectors of two classes (“norm is not norm”).

For this, assuming that all time arrays of the measured signals in the form of functions Z (t) after reduction to the normalization interval [0, 1] belong to the space L ² (integrable with a square), we represent them in the form of a linear combination (N + 1) orthonormal polynomials, in this case, Legendre polynomials. The array of nominal parameters is subject to the same decomposition.

Based on the foregoing, a technical solution was proposed for predicting the operability of multi-parameter EMC.

Using a measuring unit 2, including a number of sensors (D _i , i = 1, ... N), installed at various points of the controlled object - the electromechanical system 1, the controlled signals are received from the input of the information processing unit 3, which normalizes the input data and selects the reference values for each measured parameter. This is due to the fact that the measured parameters have different physical dimensions, therefore, in order for the neural network 10 to perform arithmetic and logical operations on them, they are normalized in the information processing unit 3, passing to dimensionless quantities represented by numbers in the range 0 ... 1 . Under the reference signal can be understood as some average nominal array of output signals.

From the output of the information processing unit 3 temporary signals of the output parameters of the EMC Z (t) reference arrays of output signals Z _H (t) for each controlled parameter are sent to the processor 5, where they are decomposed into a series of Legendre operators. Then, in adder 6, N standards are subtracted from the received N signals for each measured parameter, then the resulting difference is squared in squared block 7, and then each of the N differences is integrated over the forecast observation interval in integration block 8.

преобразования функций двух классов («норма - не норма») по каждому из N контролируемых параметров ЭМС. Далее полученный массив временных сигналов передают на нейронную сеть Кохонена 10 с каскадным подключением распределительного слоя и соревновательного слоя.In the computing unit 9, the coefficients {A _i } are found from equation (1), which is equivalent to setting the neural network nonlinear operators {φ _i },

transformations of the functions of two classes (“norm - not norm”) for each of the N controlled EMC parameters. Next, the resulting array of temporary signals is transmitted to the Kohonen neural network 10 with a cascade connection of the distribution layer and the competitive layer.

Classification of time series by groups or pattern recognition of a series is carried out using an artificial neural network. The task of pattern recognition is to assign the input data set representing the recognizable object to one of the previously known classes. The neural network approach has shown its effectiveness in solving poorly formalized recognition problems (see, for example, Korneev V.V., Gareev A.F., Vasyutin S.V., Raikh V.V. ″ Databases. Intelligent information processing ″ - M .: Publisher Molgacheva S.V., 2001). To achieve this goal, an artificial Kohonen NS was used (see, for example, C. Amza et. Al. TreadMarks: Shared Memory Computing on Networks of Workstation. Computer. Vol.29, No.2, 1996, pp. 18-23).

Practice shows that the neural network 10 is able to attribute similar images to one class. Testing the neural network 10 by supplying examples to its inputs allows us to establish: what examples the neural network 10 relates to each class, the number of classes and combinations of output values corresponding to each class. During operation, a trained neural network 10, upon supplying a next input vector to its inputs, produces a combination of output values corresponding to the class to which the neural network assigned the input vector.

The output of the Kohonen neural network 10 forms the maximum potential at that output - E, which corresponds to the predicted technical state of the EMC 1 for a particular measured parameter and determines whether the EMC is operational or inoperative through the forecast interval.

The proposed device can be implemented in software and hardware based on an industrial base.

The proposed device for predicting the operability of multi-parameter electromechanical systems is characterized by sufficient reliability and accuracy of forecasting, reducing errors in the conditions of uncertainty of the EMC parameters and the external environment, and insensitivity to the lack of a priori information.

Claims (1)

A device for predicting the operability of multi-parameter electromechanical systems (EMCs), comprising a measuring unit consisting of an information processing unit for normalizing input data from sensors and extracting a reference value for each measured parameter by obtaining temporary signals of the output parameters of the EMC and reference arrays of output signals for each monitored parameter, and a software calculator made of series-connected processor Matora, quad, integration unit, a computing unit for finding the coefficients neural nonlinear operators and their transfer to the neural network for their formation testing and the maximum potential corresponding to the predicted maintenance of EMC-specific parameter to be measured.

Images