[0009] The intelligent centrifugal pump cavitation detector of the present invention will be described in detail below using the accompanying drawings and embodiments
[0010] refer to figure 1 , the intelligent centrifugal pump cavitation detector of the present invention includes a signal acquisition system 2 composed of a pressure transmitter arranged at the pressure taking port at the inlet of the centrifugal pump 1, which is connected with the signal processing system 3, and the signal processing system 3 is connected with the detection system 4. , the detection system 4 is connected with the display control system 5 . The accurate acquisition of the pressure pulsation signal at the inlet of the centrifugal pump 1 by the signal acquisition system 2 is a key link in fault diagnosis. In order to obtain an accurate pressure pulsation waveform and reduce the detection error caused by waveform distortion, the pressure-taking port at the inlet of centrifugal pump 1 must be small enough; and since the low-frequency pulsation only changes in the axial direction of the tube, the pressure-taking port is selected in the centrifugal pump. 1. The static pressure measurement position of the inlet, so that the instantaneous value of the inlet pressure change can be obtained, and the annular chamber pressure method is adopted. The pressure transmitter of the signal acquisition system 2 should take the form of a flat film package, that is, the strain diaphragm and the side wall around the pressure measuring point are in a flush state, so that the pressure of the pressure measuring point can be measured more accurately and the pressure at the pressure measuring point can be reduced. The waveform distortion caused by hysteresis; in addition, it is necessary to ensure that the pressure transmitter has a certain range of dynamic response to meet the requirements of dynamic measurement. Considering the actual situation, install the pressure taking port on the pipe section very close to the inlet of the central pump 1, and the pressure transmitter can be directly connected with the pressure taking port through the thread.
[0011] refer to figure 1 and 2 , the signal processing system 3 of the cavitation fault detector of the present invention comprises an AT89C51 single-chip microcomputer U1, a latch U2, a program memory U3, a decoder U5, an ADC0809 analog-to-digital converter U6, a D flip-flop U7, a NOT gate U8A, or NOT gates U9A and U9B. The analog-to-digital converter U6 adopts the successive approximation ADC0809, the conversion time is 100us, and the 8-channel analog channel selection switch IN0-IN7 realizes the selection of one channel from the 0~5V output analog quantity of the pressure transmitter connected to the pressure inlet of the centrifugal pump to the internal The comparator compares, and the latch U2 and the decoder U5 are used to latch the 3-bit address sent from the ADDA, ADDB, ADDC3 address lines when the ALE signal is valid. After decoding, the channel selection signal is generated. Select the current analog channel in the analog channel. When the START signal is valid, it starts to convert the analog quantity of the output current channel. After the conversion, the converted digital quantity is sent to the latch U2 through the program memory U3. The EOC pin sends out the conversion end signal, and the latch U2 saves the digital quantity converted by the current analog channel. When the 0E signal is valid, the conversion result is sent out through D0~D7. The single-chip U1 reads in the data, and passes through the algorithm program of feature extraction and classification recognition programmed in advance and written into the program memory U3. The detection system 4 includes the analysis of wavelet multi-scale statistical feature quantities, extracts features from the signal from the signal processing system, and uses the support vector machine to classify and identify the extracted feature parameters to distinguish different cavitation states of the centrifugal pump 1 during operation. : In the normal operation stage, the primary cavitation stage, the slight cavitation stage and the severe cavitation stage, the final detection result is sent to the display control system 5, which includes the liquid crystal display module, the speaker LS and the fluctuation switches S3-S6. Set the P1.0 port to make the speaker LS alarm, and send it to the LCD U4 through the data bus for online display. LCD U4 adopts LCD module driven by two SED1520, it has 16 row drive ports and 61 column drive ports, can be directly connected with 8-bit microcontroller U1, and can display 7*2 Chinese character blocks or 15*4 character blocks . Enough to hold all the displayed information. Switches S3 to S6 realize the switching among multiple pressure transmitters, and the conversion control of displaying Chinese characters and characters. The power supply of the whole system is provided by terminal P1, and Vcc is +5V voltage.
[0012] All the devices involved in the present invention are commercially available commodities, among which the single-chip microcomputer, A/D and its liquid crystal display module can be replaced with higher-grade microprocessors or chips according to actual working conditions, but the detection principle and its functional structure remain unchanged. . The measuring range of the pressure transmitter is -0.1Mpa~+0.1MPa, the accuracy is 0.25%, the repeatability is 1%, and the frequency response is 1000Hz. The signal acquisition system 2 of the centrifugal pump 1 is composed of a plurality of pressure transmitters. The signal acquisition system 2 of the centrifugal pump 1 in this specific embodiment is composed of 8 pressure transmitters, and one pressure transmitter is installed on one centrifugal pump 1 , that is, the present invention can perform cavitation detection on a plurality of centrifugal pumps 1 at the same time.
[0013] refer to figure 1 and 2 , the detection system 4 includes the following specific contents:
[0014] ①Wavelet multi-scale statistic feature extraction of the original signal:
[0015] The wavelet transform retains the main features of the continuous wavelet transform in a very compact form, and no information is lost in the process. In order to realize the wavelet fast algorithm, Mallat proposed a tower algorithm based on orthogonal mirror filter, which can be described by linear filtering and matrix operation respectively. Let the time series X t length is N=2 J , where the subscript t is the label of the time series, and J is any positive integer. The first step of the tower algorithm is to evaluate {X=X i : t=0,1,...N-1} is orthogonally decomposed into two new sequences of length N/2: {W 1 =W 1,t : t=0,1,...N/2-1} and {V 1 =V 1,t : t=0, 1,...N/2-1} That is:
[0016] W = W 1 V 1 = A 1 B 1 X - - - ( 1 )
[0017] where W 1 , V 1 , A 1 , B 1 are all N/2×N order matrices, and satisfy:
[0018] A 1 A 1 T = B 1 B 1 T = I N / 2 - - - ( 2 )
[0019] A 1 is determined by the wavelet filter {h 1} After periodization, the translation is obtained, B 1 by the scale filter {g 1} After periodization, the translation is obtained; A 1 T , B 1 T respectively A 1 and B 1 The transpose matrix of , I N/2 is the identity matrix.
[0020] The linear filtering method can be expressed as follows: Let {h l : l=0,1,...L-1} is a wavelet filter of even length L, which satisfies:
[0021] Σ l = 0 L - 1 h l = 0 , Σ l = 0 L - 1 h l 2 = 1 , Σ l = 0 L - 1 h l h l + 2 n = 0 - - - ( 3 )
[0022] with {h 1} to X t Periodic filtering is performed, and the filtered coefficients are kept every two, and the following results are obtained:
[0023] W 1 = W 1 , t = Σ l = 0 L - 1 h l X 2 t + 1 - l mod N = Σ l = 0 N - 1 h l o X 2 t + 1 - l mod N - - - ( 4 )
[0024] in the formula is called a periodic filter, which is a combination of {h l} is obtained by extending with N as the period.
[0025] Let {g l : l=0,1,...L-1} is a scale filter of even length L, which satisfies:
[0026] Σ l = 0 L - 1 g l = 2 , Σ l = 0 L - 1 g l 2 = 1 , Σ l = 0 L - 1 g l g l + 2 n = 0 , Σ l = 0 L - 1 g l g l + 2 n = 0 - - - ( 5 )
[0027] with {g 1} to X t Periodic filtering is performed, and the filtered coefficients are kept every two, and the following results are obtained:
[0028] V 1 = V 1 , t = Σ l = 0 L - 1 g l X 2 t + 1 - l mod N = Σ l = 0 N - 1 g l o X 2 t + 1 - l mod N - - - ( 6 )
[0029] in the formula is called a periodic filter, which is a combination of {g l} is obtained by extending with N as the period.
[0030] The second step of the tower algorithm is to convert V 1 As X in the first step, repeat the above decomposition process, namely:
[0031] W 2 = W 2 , t = Σ l = 0 L - 1 h l V 1,2 t + 1 - l mod N / 2 = Σ l = 0 N - 1 h l o V 1,2 t + 1 - l mod N / 2 - - - ( 7 )
[0032] V 2 = V 2 , t = Σ l = 0 L - 1 g l V 1,2 t + 1 - l mod N / 2 = Σ l = 0 N - 1 g l o V 1,2 t + 1 - l mod N / 2 - - - ( 8 )
[0033] In this way, the coefficient W after the normalized orthogonal discrete wavelet transform of J times can be decomposed into J+1 sub-vectors, namely:
[0034] W = W 1 V 1 = W 1 W 2 V 2 = . . . W 1 W 2 . . W J V J - - - ( 9 )
[0035] Among them, W j (j=1,2,3,...,J) is length N/2j column vector of . W j contains all about the scale λ j The discrete wavelet transform coefficients of . Here, db4 wavelet is used for multi-scale decomposition. After wavelet transformation, it can be expressed as:
[0036] X t = Σ j = 1 J W j , t + V J , t - - - ( 10 )
[0037] For unified representation, use W J+1,t instead of V J,t Then there are:
[0038] X t = Σ j = 1 J + 1 W j , t - - - ( 11 )
[0039] Assume (j=1,2,...J,J+1), then E j,t That is, the energy sequence, that is, the multi-scale representation of the signal energy. In the expression of wavelet energy entropy, represents the sum of the wavelet energy at the j scale, p j,i =E j,i /E j.
[0040] Feature extraction is a key link in fault diagnosis. Selecting and extracting optimal fault features can improve the efficiency and accuracy of diagnosis. Multi-scale energy statistics are general quantities with universal significance and representativeness. A common method in engineering is to calculate some statistical characteristics of wavelet coefficients at each scale. Here we will extract the statistics in Table 1 as feature vectors. The results of feature vector extraction are shown in Table 2.
[0041] Table 1 Multiscale energy statistics
[0042]
[0043] Table 2 Multi-scale eigenvectors of signals under different effective NPSH
[0044]
[0045]
[0046] ② Support vector machine classification and recognition:
[0047] A support vector machine is a learning system using linear function assumptions in a high-dimensional feature space, trained by a learning algorithm from optimization theory that implements a learning bias derived from statistical learning theory.
[0048] For linearly separable sample sets (x i , y i ), where i=1,2,...n; x i ∈R d , y i ∈{-1,1}, the general form of the linear discriminant function in d-dimensional space is: g(x)=w·x+b, where w·x+b=0 is the classification surface equation, w is the normal phase of the classification surface quantity. Normalize the discriminant function so that all samples of the two classes satisfy |g(x)|≥1, so that the classification interval is equal to 2/||w||, so that the classification hyperplane can correctly classify all samples, which is the requirement It satisfies:
[0049] y i [(w x i )+b]-1≥0, i=1, 2,...n (12)
[0050] The problem of constructing the optimal hyperplane is transformed into finding the minimum value of the following formula under the constraints of formula (12):
[0051] Φ(w)=||w|| 2 /2=(w·w)/2 (13)
[0052] The optimal solution for this optimization is the saddle point of the following Lagrange function:
[0053] L ( w , b , α ) = 1 2 | | w | | 2 - Σ i = 1 n α i [ y i ( w · x i + b ) - 1 ] - - - ( 14 )
[0054] In the formula, α i0, it is the Lagrange coefficient. This is a convex quadratic programming problem. There is a unique optimal solution. At the same time, if the optimal solution satisfies the KT condition, the original problem is transformed into a relatively simple quadratic programming, as follows:
[0055] max W ( α ) = Σ i = 1 n α i - 1 2 Σ i , j n α i α j y i y j ( x i · x j ) (15)
[0056] st Σ i = 1 n α i y i = 0 , α i ≥ 0 , i = 1,2 , . . . n
[0057] Solve the above problem and get the optimal solution and b, the optimal hyperplane can be determined.
[0058] When the sample set is linearly inseparable, the sample is mapped to a high-dimensional space Z using the kernel function, and then it is regarded as a linearly separable case in Z, and the original linear method is used to solve it.
[0059] The feature vectors in Table 2 are used for classification, recognition and detection with support vector machines. The test results are shown in Table 3.
[0060] The detection system 4 uses C language or assembly language to write programs.
[0061] Table 3 Recognition results of support vector machines
[0062]