Generator fault diagnosis method based on FA-BP neural network model

Abstract

The embodiment of the invention discloses a generator fault diagnosis method based on an FA-BP neural network model, and the method comprises the steps: collecting the operation data and image data of a generator, and enabling the operation data to comprise an electric signal and a vibration signal; obtaining operation feature vectors according to the operation data, and obtaining image features according to the image data; inputting the operation feature vector into an FA-BP neural network model to obtain an initial fault probability; calculating the correlation strength between the fault type and the image feature; correcting the initial fault probability by adopting association strength to obtain a final fault probability; and outputting a generator fault state according to the final fault probability. And the initial fault diagnosis result is corrected by using the image data, so that the accuracy, real-time performance and false alarm rate of fault state judgment of the generator can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis methods, and in particular to a generator fault diagnosis method based on an FA-BP neural network model. Background Art

[0002] In modern power systems, generators are core equipment, and their operating status is directly related to the stability and reliability of power supply. As power systems continue to expand in size and complexity, the importance of generator fault diagnosis technology is becoming increasingly prominent. Accurate and timely diagnosis of generator faults is crucial for ensuring safe power system operation, reducing maintenance costs, and minimizing downtime.

[0003] Traditional generator fault diagnosis methods mainly rely on sensor data. Although these methods can reflect the operating status of the generator to a certain extent, since fault diagnosis relies solely on operating data collected by sensors, they are easily interfered by factors such as environmental noise and sensor accuracy, leading to misjudgment or missed diagnosis.

[0004] Especially in the early stages of a fault, changes in operating data can be very subtle, making it difficult to accurately identify the fault, resulting in insufficient fault diagnosis accuracy. This is especially true for insulation damage in key generator components, such as bearing seats and stator windings. Relying solely on raw vibration signals cannot detect early generator faults. Summary of the Invention

[0005] In view of this, the purpose of an embodiment of the present invention is to provide a generator fault diagnosis method based on the FA-BP neural network model. By fusing operating data and image data and making full use of image features, the technical problems of misjudgment or missed judgment caused by using only sensor data and the inability to detect early faults are solved.

[0006] An embodiment of the present invention provides a generator fault diagnosis method based on an FA-BP neural network model, comprising: Collecting operating data and image data of the generator, wherein the operating data includes electrical signals and vibration signals; Obtaining an operation feature vector according to the operation data, and obtaining an image feature according to the image data; Inputting the operation feature vector into the FA-BP neural network model to obtain an initial failure probability; calculating the correlation strength between the fault type and the image feature; Correcting the initial failure probability using the correlation strength to obtain a final failure probability; The generator fault state is output according to the final fault probability.

[0007] Further, the operating data further includes a rotational speed signal, the electrical signal includes an output current signal, and the vibration signal includes a vibration acceleration signal.

[0008] Further, obtaining an operating feature vector based on the operating data includes: Obtaining a current imbalance degree based on the output current signal; Calculating an amplitude ratio of a once-per-revolution amplitude and a twice-per-revolution amplitude based on the vibration signal and the rotational speed signal; Calculating a high-frequency energy proportion based on the vibration signal; The operating feature vector at least includes the current imbalance degree, the amplitude ratio, and the high-frequency energy proportion.

[0009] Further, the image data includes a bearing housing image and a stator winding surface image.

[0010] Further, obtaining image features based on the image data includes: Obtaining a bearing light spot density based on the bearing housing image; Obtaining a crack proportion based on the stator winding surface image; The image features at least include the light spot density and the crack proportion.

[0011] Further, calculating the correlation strength between a fault type and image features includes: Constructing a correlation matrix between the fault type and the image features; Calculating an image feature saliency of the image features, where the image feature saliency represents the degree to which the image features exceed an image feature threshold; Calculating an image correction weight based on the correlation matrix and the image feature saliency.

[0012] Further, constructing the correlation matrix between the fault type and the image features according to prior knowledge.

[0013] Further, calculating the image correction weight based on the correlation matrix and the image feature saliency is as follows: ; where is the image feature saliency of the image feature , is the th fault type and the th correlation matrix of the image features, is the total number of fault types, , is the total number of image features, .

[0014] Further, the initial failure probability is corrected by the correlation strength to obtain the final failure probability as follows: ; wherein, is the final failure probability of the th type of failure, is the initial failure probability of the th type of failure, and are the correction weight coefficients, is the positive correction threshold, is the negative correction threshold.

[0015] Further, the generator failure state is output according to the final failure probability, including: When the final failure probability of the kth type of failure is greater than or equal to the failure threshold, the kth type of failure is output; when the final failure probability of the kth type of failure is less than the failure threshold but greater than or equal to the warning threshold, a warning prompt is output; if all failure probabilities are less than the warning threshold, normal is output.

[0016] The beneficial effects of the embodiments of the present invention are as follows: By using the operation data collected by the sensor, the present invention generates the initial failure probability by using the FA-BP neural network model. The initial failure probability obtained at this time can reflect the possible failure probability. Then, the image features are synchronously collected, the correlation strength between the image features of the generator and the failure types is established, and the initial failure probability is corrected by the information related to the image features, which can improve the accuracy of generator failure diagnosis and avoid misjudgment or missed judgment. And when the image features include the spot density and the crack ratio, the technical solution of the present invention can have the early warning ability for the bearing seat failure and the stator winding insulation failure through the image features, improving the comprehensiveness, accuracy and timeliness of failure diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 is a schematic flowchart of the generator failure diagnosis method based on the FA-BP neural network model of the present invention; Figure 2 is a schematic flowchart of using the FA to optimize the BP neural network model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated herein can generally be arranged and designed in a variety of different configurations.

[0020] Please refer to Figures 1 to 2 , a generator fault diagnosis method based on the FA-BP neural network model of the present invention, comprising the steps of: S1. Collect the operating data and image data of the generator.

[0021] Before establishing a fault diagnosis model and performing fault diagnosis, it is necessary to encode the possible fault types of the generator. These encodings will be used as the target outputs of the model, enabling the model to correctly identify and classify fault types according to the input characteristic parameters. Table 1 shows 10 common generator fault types.

[0022] Table 1 Generator fault types and encodings:

[0023] According to the fault types, the present invention collects the operating data corresponding to the fault types. The operating data includes at least electrical signals and vibration signals.

[0024] Among them, the electrical signals include three-phase output current signals , and the output current is detected by a high-frequency current transformer installed on the three-phase bus at the output end of the generator. According to the three-phase current signals , the three-phase current unbalance degree can be obtained through a conventional current unbalance degree algorithm, or the three-phase current unbalance degree can also be directly collected by a zero-sequence current transformer signal, so as to judge the electrical short-circuit fault caused by the electrical line fault; the electrical signals also include the voltage signals output by the generator, and the overvoltage / undervoltage fault of the generator is detected according to the specific voltage value of the voltage signal; since the output power of the generator will decrease when a fault occurs, it is preferably also used to collect the output power of the generator by a power transmitter, and the output power of the generator is used as an input feature during fault diagnosis to accurately judge the analysis of the fault type.

[0025] The faults of the generator are usually related to vibration signals. The present invention associates the vibration signals with the fault types by setting vibration velocity sensors at multiple positions to detect the vibration signals. Among them, the process of obtaining the rotor imbalance fault is as follows: The vibration signal is collected by an acoustic vibration sensor on the generator bearing pedestal ; Calculate the amplitude of the first harmonic rotational frequency based on the vibration signal , the amplitude of the first harmonic rotational frequency is the amplitude of the sinusoidal component synchronized with the rotor rotation frequency in the vibration signal, and the calculation formula is as follows: ; wherein, represents the rotational frequency, represents the rotational speed measured by the rotational speed sensor.

[0026] represents performing a Fourier transform on the vibration signal .

[0027] Then calculate the amplitude of the second harmonic rotational frequency based on the vibration signal , the amplitude of the second harmonic rotational frequency is the amplitude of the sinusoidal component with a frequency twice that of the rotor rotation frequency in the vibration signal, and the calculation formula is as follows: ; Finally, calculate the amplitude ratio of the amplitude of the first harmonic rotational frequency and the amplitude of the second harmonic rotational frequency , and the formula is as follows: ; When the amplitude ratio exceeds a certain range, the unbalance fault of the rotor can be analyzed.

[0028] The proportion of high-frequency energy can reflect the bearing wear fault. Therefore, based on the vibration signal collected by the acoustic vibration sensor on the generator bearing pedestal , further calculate the proportion of high-frequency energy. The high-frequency frequency range is the proportion of high-frequency energy from the frequency to the upper limit of the frequency bandwidth , and the calculation formula is as follows: ; wherein, is the power spectral density, is the upper limit of the analysis bandwidth. The power spectral density is obtained by dividing the vibration signal into segments, windowing each segment (such as a Hanning window), calculating the square of the FFT amplitude of each segment, and then taking the average. The specific calculation process is as follows: Perform segmented windowing on the vibration signal : ; wherein, is the number of segments into which the signal is divided, is the Time period is the Hanning window function representing the vibration signal at the time period after being processed by the Hanning window function, the windowed signal

[0029] For the windowed signal after segmentation perform FFT calculation: ; wherein is the windowed signal at the frequency the Fourier transform value at that point

[0030] Calculate the window function energy normalization factor U: ; where N is the number of sampling points of a single-segment signal

[0031] The calculation formula for the power spectral density P(f) is as follows: : wherein is the sampling frequency, and the power spectral density P(f) represents the distribution of vibration energy in the frequency domain

[0032] In order to detect insufficient cooling water circulation or cooling system failure, an acoustic sensor, such as a condenser microphone, is installed on the cooling water pipe wall. After the condenser microphone collects the analog voltage signal V(t), the cavitation sound pressure level is calculated based on the analog voltage signal V(t) to characterize the intensity of the cavitation phenomenon in the cooling water pipeline. The specific calculation process is to convert the voltage signal into a sound pressure signal , the sound pressure signal i.e., the change in air pressure caused by the propagation of sound waves, and the calculation formula is as follows: ; is the microphone sensitivity, unit: V / Pa, typical value 50 mV / Pa

[0033] From the sound pressure signal calculate the cavitation sound pressure level , and the calculation process is as follows: First, calculate the root mean square value of the sound pressure signal : ; where N is the number of sampling points of the sound pressure signal

[0034] Then, from the root mean square value of the sound pressure signal Obtain the cavitation sound pressure level as follows: ; wherein represents the reference sound pressure.

[0035] When it is also necessary to analyze the overheating fault of the generator, temperature sensors can also be respectively arranged inside the generator and in the cooling water system to detect the temperature inside the generator and the temperature of the generator cooling water, and the information collected by the temperature sensors is used as the input feature of the model for analyzing the overheating fault.

[0036] While collecting the operation data, the present invention synchronously collects the image data of the generator. Based on some microscopic faults in the generator, such as bearing wear or insulation damage, these fault characteristics will not cause obvious changes in the electrical signal or vibration signal in the early stage. By capturing the characteristics related to such faults through the image data, the faults can be detected in advance and the early diagnosis of the faults can be realized. For relatively macroscopic and less influential fault types such as fan faults, there is no need to additionally set up image acquisition equipment.

[0037] An ultraviolet imager is arranged near the outer surface of the generator bearing seat to collect the ultraviolet image data of the bearing. When the generator bearing wears and causes the oil film to rupture, leading to micro-discharge, that is, the phenomenon of electrical erosion, this phenomenon shows light spots in the ultraviolet image data, and the wear state of the bearing can be reflected by the light spot density. Therefore, in the present invention, the image data of the bearing seat is collected for the early diagnosis of faults and to make up for the bearing seat fault information that cannot be reflected by the vibration signal.

[0038] An industrial visible light high-definition camera is arranged near the stator winding of the generator to collect the image data of the stator winding. The insulation aging of the generator stator winding causes cracks on the surface of the winding. However, the cracks in the stator winding are relatively small in the early stage, and it is difficult for other sensor signals to accurately identify the insulation fault. The present invention analyzes the insulation surface crack characteristics in the stator winding image data and fuses it with the sensor data to timely discover the insulation layer breakage fault and deal with it, preventing serious faults such as short circuit caused by the expansion of the crack.

[0039] S2. Respectively obtain the operation feature vectors according to the operation data in step S1, and obtain the image features according to the image data in step S1.

[0040] In addition to including the calculated current unbalance degree, amplitude ratio, high-frequency energy ratio, and cavitation sound pressure level, the operation feature vector also includes all the collected original sensor data signals to comprehensively reflect the fault data of the generator, such as temperature signals, etc.

[0041] After the ultraviolet image data collected by the ultraviolet imager are subjected to Gaussian filtering and binary segmentation, the number of discharge spots is obtained by connected component analysis to mark the discharge spots, and the spot density is calculated. As follows: ; Among them, the spot density is the number of discharge spots per unit area (unit: pieces / cm²), representing the degree of bearing wear.

[0042] After the original image collected by the Basler industrial visible light high-definition camera is denoised by Gaussian filtering, the binary edge map is output by edge detection through the Canny operator, and then the spurious lines are removed through morphological processing to obtain the crack pixels. The calculation of the proportion of stator winding crack pixels is as follows: ; Among them, the crack proportion represents the degree of insulation damage of the stator winding.

[0043] The crack proportion and the spot density are both used as image features to correct the initial fault probability.

[0044] In addition, an electrical short circuit fault will cause the temperature of the generator rotor lead joint to rise, which will in turn cause the growth of hot spots. To further optimize the judgment of such faults, preferably, the image data of the rotor lead joint are also collected to obtain the image feature reflecting the hot spot growth rate. .

[0045] S3. Input the operation feature vector into the FA-BP neural network model to obtain the initial fault probability.

[0046] The present invention adopts the FA-BP neural network model, that is, the Firefly Algorithm (FA) is used to optimize the BP neural network (Back Propagation Neural Network, feedforward neural network) model. In the FA-BP neural network model of the present invention, the Firefly Algorithm FA and the BP neural network are integrated with each other, aiming to overcome problems such as the local optimal solution trap and slow convergence speed faced by the traditional BP neural network during the training process. Based on the powerful global search ability of the Firefly Algorithm, by optimizing the weights and thresholds of the BP neural network, the model has a more superior parameter initialization setting at the beginning of training. This algorithm strategy can not only significantly improve the convergence efficiency of the model, but also enhance the global search ability and diagnostic accuracy of the network in complex fault diagnosis tasks to a certain extent, thereby improving the robustness and stability of the model. This method effectively combines the global exploration characteristics of FA with the powerful learning ability of BP, providing a more reliable and efficient solution for the generator fault diagnosis of complex systems. According to the optimization process as shown in Figure 2 The BP neural network is optimized by the Firefly Algorithm to obtain the optimal initial weights W1, W2 and biases b1, b2.

[0047] The operation feature vector related to the original sensor data is input into the FA-BP neural network model to obtain the initial fault probability : ; wherein, is the activation function, W1, W2 are the weights of the BP neural network optimized by FA, and b1, b2 are the biases of the BP neural network optimized by FA.

[0048] S4. Calculating the association strength between the fault type and the image features includes: S4-1. Constructing the association matrix between the fault type and the image features.

[0049] The association matrix is a p×q matrix, indicating the association strength between the k( )th fault type and ( ) kinds of image features. Taking 10 fault types and 2 image features as an example, a 10×2 association matrix is defined in advance according to physical knowledge or other prior knowledge.

[0050] S4-2. Calculating the image feature saliency of the image features; For each image feature, calculate the image feature saliency as follows: ; For example, the proportion of cracks in the image features significance is calculated as follows: ; Among them, is the crack index threshold, and the value can be set to 0.05%. The crack proportion significance represents the ratio of the actual crack proportion to the threshold.

[0051] For example, the light spot density in the image features significance is calculated as follows: ; Among them, is the light spot density threshold, and the value can be set to 5 per cm². The light spot density significance represents the ratio of the actual light spot density to the threshold.

[0052] S4-3. Calculate the image correction weight according to the correlation matrix and the image feature significance.

[0053] For each fault type k (k = 1~p), calculate the image correction weight as follows: ; Among them, is the image feature image feature significance, is the th fault type and th image feature, is the total number of fault types, , is the total number of image features, .

[0054] Preferably, the acquisition and processing of image data are processed in parallel with the acquisition and processing of sensor data, which can achieve fast diagnosis of generator faults.

[0055] S5. Obtain the final fault probability by correcting the initial fault probability with the correlation strength .

[0056] The present invention corrects the initial fault probability through a contradiction / support double correction mechanism, where the support correction means that when the image feature image feature significance exceeds the positive correction threshold When it is, it indicates that the fault type exists through image feature confirmation. At this time, increasing the probability of this fault through supportive correction can effectively reduce the missed detection caused by sensor diagnosis and improve the early warning ability. When the positive correction threshold is set to 0.8, it is determined that When it is, supportive correction is performed; contradictory correction means that when the image feature The saliency of the image feature is less than the negative correction threshold When it is, it indicates that the image feature believes that the fault type does not exist. At this time, reducing the probability of this fault through contradictory correction can suppress the false alarm of the sensor diagnosis result. When the negative correction threshold is set to 0.5, it is determined that When it is, contradictory correction is performed.

[0057] ; Among them, Is the final fault probability of the th type of fault, Is the initial fault probability of the th type of fault. Among them, And Are correction weight coefficients, obtained through experiments The value is 0.25, When the value is 0.30, the false alarm rate can be significantly reduced and the early detection rate can be improved.

[0058] To prevent the final fault probability From exceeding the maximum value of 1, the final fault probability Is constrained as follows: .

[0059] S6. Output the generator fault status according to the final fault probability .

[0060] After image correction, a more accurate final fault probability Is obtained. If the corrected probability Of the kth type of fault is greater than or equal to the fault threshold (for example, 0.85), it is determined that the kth type of fault occurs and the kth type of fault is output; if the corrected probability Of the kth type of fault is less than the fault threshold but greater than or equal to the early warning threshold (for example, 0.7), it is determined that the kth type of fault is about to occur and an early warning prompt is output; if all fault probabilities are less than the early warning threshold, the generator is in a healthy state and "normal" is output.

[0061] The present invention can improve the accuracy, real-time performance and false alarm rate of the generator fault status judgment by collecting image data of some key equipment of the generator using an image acquisition device and correcting the initial fault diagnosis result using the image data.

[0062] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any technician familiar with the technical field of the present invention can still modify the technical solutions described in the foregoing embodiments or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A generator fault diagnosis method based on the FA-BP neural network model, characterized in that, Including: Collecting the operation data and image data of the generator, where the operation data includes electrical signals and vibration signals; Obtaining an operation feature vector based on the operation data, and obtaining an image feature based on the image data; Inputting the operation feature vector into the FA-BP neural network model to obtain an initial fault probability; Calculating the correlation intensity between the fault type and the image feature; Using the correlation intensity to correct the initial fault probability to obtain a final fault probability; Outputting the fault state of the generator according to the final fault probability.

2. The method according to claim 1, characterized in that, The operation data further includes a rotational speed signal, the electrical signal includes an output current signal, and the vibration signal includes a vibration acceleration signal.

3. The method according to claim 2, characterized in that, Obtaining an operation feature vector based on the operation data includes: Obtaining a current imbalance degree based on the output current signal; Calculating the amplitude ratio of the fundamental frequency amplitude and the double frequency amplitude according to the vibration signal and the rotational speed signal; Calculating the high-frequency energy ratio according to the vibration signal; The operation feature vector at least includes the current imbalance degree, the amplitude ratio, and the high-frequency energy ratio.

4. The method according to claim 1, characterized in that The image data includes bearing housing images and stator winding surface images.

5. The method according to claim 4, characterized in that Obtaining an image feature based on the image data includes: Obtaining a bearing light spot density based on the bearing housing image; Obtaining a crack ratio based on the stator winding surface image; The image feature at least includes the light spot density and the crack ratio.

6. The method according to claim 1, characterized in that, Calculating the correlation intensity between the fault type and the image feature includes: Constructing a correlation matrix between the fault type and the image feature; Calculating the image feature saliency of the image feature, where the image feature saliency represents the degree to which the image feature exceeds the image feature threshold; Calculating an image correction weight according to the correlation matrix and the image feature saliency.

7. The method according to claim 6, characterized in that, Constructing a correlation matrix between the fault type and the image feature according to prior knowledge.

8. The method according to claim 6, wherein Calculate the image correction weight according to the association matrix and the image feature saliency as follows: ； Among them, is the image feature saliency of the image feature, is the th type of fault and the association matrix of the image features, is the total number of fault types, , is the total number of image features, .

9. The method according to claim 8, characterized in that, The formula for using the correlation intensity to correct the initial fault probability to obtain the final fault probability is as follows: ； Among them, is the final failure probability of the type of failure, is the initial failure probability of the type of failure, and are the correction weight coefficients, is the positive correction threshold, is the negative correction threshold.

10. The method according to claim 7, wherein Outputting the fault state of the generator according to the final fault probability includes: When the final fault probability of the kth type of fault is greater than or equal to the fault threshold, output the kth type of fault; when the final fault probability of the kth type of fault is less than the fault threshold but greater than or equal to the warning threshold, output a warning prompt; if all fault probabilities are less than the warning threshold, output normal.

Images