Online potential fault diagnosis method of intelligent release power supply circuit based on QPSO-SVM (Quantum Particle Swarm Optimization-Support Vector Machine)

Abstract

The invention discloses a QPSO-SVM-based potential fault on-line diagnosis method for an intelligent release power supply circuit, and the method comprises the steps: firstly carrying out the fault mode and impact analysis of the power supply circuit through combining with the working principle of the power supply circuit, determining the object of fault diagnosis, and setting the fault type of the power supply circuit; then, on the basis of charging and discharging characteristics of an energy storage capacitor of a power supply circuit and characteristics of a control loop, utilizing a Monte Carlo simulation experiment to extract a group of characteristic parameters capable of reflecting a circuit fault under non-constant excitation, and adopting a support vector machine recursive characteristic elimination method to carry out importance ranking on high-dimensional fault characteristics; and screening according to a sorting result to obtain an optimal fault feature set. And the parameters of the SVM model are optimized through the QPSO, so that the accuracy of fault diagnosis of the power supply circuit is improved. On the premise that a constant excitation signal is not applied and the circuit structure is not changed, on-line diagnosis of potential faults of the intelligent release power supply circuit is achieved, and the method has high engineering application value.

Description

Technical Field

[0001] The present invention relates to the field of circuit fault diagnosis, and in particular to an online diagnosis method for potential faults in a power supply circuit of an intelligent release based on QPSO-SVM. Background Art

[0002] As the level of informationization and intelligence of power systems continues to improve, a large number of intelligent circuit breakers have been put into power grid operation, playing a key role in ensuring the safe and stable operation of power distribution systems. Intelligent circuit breakers are the integration of circuit breaker bodies and intelligent releases, which can protect power systems from overload, residual current, under / overvoltage and other fault hazards. Intelligent releases are the core components of intelligent circuit breakers, mainly composed of power modules, signal detection and conditioning modules, central control modules, output drive modules and human-computer interaction modules, and are the key to realizing the basic protection functions of intelligent circuit breakers. Among them, the power module uses a self-generated power circuit to extract electrical energy from the line to provide stable energy for other modules. Once the power circuit fails, other modules will also be affected, which will eventually lead to the failure of the circuit breaker protection function. Therefore, timely diagnosis of potential faults in the power circuit of the intelligent release is of great significance to ensure the stable operation of the power system.

[0003] At present, the research on power circuit fault diagnosis methods is a hot topic. Ensuring the stable operation and efficient maintenance of power circuits has become a key link in ensuring the reliability of various electronic equipment and systems. However, there is a lack of research on online fault diagnosis methods for power circuits in the specific field of intelligent releases. Existing power circuit fault diagnosis strategies generally rely on injecting a constant excitation signal into the circuit under test, then analyzing the fault feature information contained in the circuit response, and realizing fault diagnosis by monitoring the changes in these features. However, this method is essentially limited to offline diagnosis, and it is difficult to monitor the circuit status in real time and dynamically, which limits its breadth and effectiveness in practical applications. In addition, considering that the power circuit of the intelligent release is integrated into the intelligent circuit breaker system, if a constant excitation signal is required to perform fault diagnosis, this process will undoubtedly add a lot of complexity and inconvenience in the actual operation and maintenance links, thereby restricting its application and maintenance in real scenarios. Therefore, it is particularly important and urgent to study the online fault diagnosis method of potential faults that can be directly applied to the power circuit inside the intelligent circuit breaker without constant excitation. Summary of the invention

[0004] In view of the deficiencies of existing research, the technical problem that the present invention intends to solve is to provide an online diagnosis method for potential faults in the power supply circuit of an intelligent release based on QPSO-SVM. Without applying a constant excitation signal or changing the circuit structure, online diagnosis of potential faults in the power supply circuit can be achieved, and the accuracy of fault diagnosis is improved by introducing quantum mechanical behavior to improve the particle swarm optimization algorithm.

[0005] The technical solution of the present invention to solve the technical problem is to provide a potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM, characterized in that the method comprises the following steps:

[0006] Step 1: According to the structure and working principle of the power circuit of the intelligent release, combined with the "Electronic Equipment Reliability Prediction Model and Data Manual", the power circuit is analyzed for failure mode and impact, the potential failure type of the power circuit is determined, and the fault diagnosis objects are selected as energy storage capacitors, resistors No. 1, No. 3, No. 4, No. 5, and No. 7. The power circuit failure type is set according to the different parameter offset ranges of the fault diagnosis objects;

[0007] The intelligent release power supply circuit includes a No. 1 current transformer, a No. 2 current transformer, a No. 3 current transformer, a No. 1 rectifier bridge, a No. 2 rectifier bridge, a No. 3 rectifier bridge, a No. 1 resistor, a No. 2 resistor, a No. 3 resistor, a No. 4 resistor, a No. 5 resistor, a No. 6 resistor, a No. 7 resistor, a diode, a No. 1 filter capacitor, a No. 2 filter capacitor, a No. 4 filter capacitor, an energy storage capacitor, a MOSFET, a voltage comparator, and a linear regulator;

[0008] The input ends of the No. 1 current transformer, the No. 2 current transformer, and the No. 3 current transformer are respectively connected to the three-phase AC lines A, B, and C, the AC input end of the No. 1 rectifier bridge is respectively connected to the two output ends of the No. 1 current transformer, the AC input end of the No. 2 rectifier bridge is respectively connected to the two output ends of the No. 2 current transformer, the AC input end of the No. 3 rectifier bridge is respectively connected to the two output ends of the No. 3 current transformer, the positive output ends of the three rectifier bridges are all connected to the anodes of the diodes, and the negative output ends of the three rectifier bridges are grounded;

[0009] The cathode of the diode is connected to the positive electrode of the energy storage capacitor, and the other end of the energy storage capacitor is grounded; the input pin 1 of the linear regulator is connected to the positive electrode of the energy storage capacitor, the output pin 3 of the linear regulator is connected to one end of the filter capacitor No. 4, and the other end of the filter capacitor No. 4 is grounded, and the pins 2 and 4 of the linear regulator are grounded; at the same time, the output pin 3 of the linear regulator is the wiring terminal of the external module;

[0010] The drain of the MOSFET is connected to the anode of the diode, the gate of the MOSFET, one end of the No. 1 filter capacitor, one end of the No. 6 resistor, and one end of the No. 7 resistor are all connected to the output end of the voltage comparator, and the source of the MOSFET, the other end of the No. 1 filter capacitor, and the other end of the No. 6 resistor are all grounded;

[0011] One end of the No. 1 resistor is connected to the cathode of the diode, the other end of the No. 1 resistor, the other end of the No. 7 resistor, and one end of the No. 5 resistor are all connected to one end of the No. 2 resistor, and the other end of the No. 2 resistor is connected to the output end of the voltage comparator; the other end of the No. 5 resistor is grounded; one end of the No. 3 resistor is connected to the No. 3 output pin of the linear regulator, and the other end of the No. 3 resistor is connected in parallel with one end of the No. 4 resistor and one end of the No. 2 filter capacitor, and then connected to the inverting end of the voltage comparator; the other end of the No. 4 resistor and the other end of the No. 2 filter capacitor are connected in parallel and then grounded;

[0012] Step 2, using Pspice software to build a simulation model of the power supply circuit of the intelligent release described in step 1, selecting the voltage at the energy storage capacitor terminal as the monitoring signal, and performing a Monte Carlo simulation experiment analysis on the power supply circuit under each fault type according to the power supply circuit fault type set in step 1, and under each fault type, the sampling range of the input current is set to between 0.5 times and 2 times the rated input current, thereby obtaining multiple groups of voltage waveforms at the energy storage capacitor terminal under different fault types;

[0013] Step 3, based on the charging and discharging characteristics of the energy storage capacitor, and the influence of the change in the value of the fault diagnosis object on the energy storage capacitor terminal voltage of the energy storage capacitor C3, select the energy storage capacitor discharge time, peak-to-peak value, maximum value, minimum value, voltage mean, standard deviation, and peak factor on the voltage waveform of the energy storage capacitor under each Monte Carlo simulation experiment as the first group of candidate fault feature parameters; the energy storage capacitor charging time, charge and discharge cycle, charge and discharge time ratio, effective value, kurtosis, margin factor, and skewness on the energy storage capacitor terminal voltage waveform are used as the second group of candidate fault feature parameters; extract the above-mentioned first group and second group of candidate fault feature parameter values ​​in the experimental sample data of each group of energy storage capacitor C3 terminal voltage waveform, and combine the two to obtain a composite fault feature data, and use the label of the power supply circuit fault type corresponding to the voltage waveform as the label of the data, and all the composite fault feature data constitute a training sample set S;

[0014] Step 4: Use the support vector machine recursive feature elimination method to sort the importance of the alternative fault feature parameters of all composite fault feature data in the training sample set S, and obtain the importance sorting result of the alternative fault feature parameters; at the same time, use the training sample set S and the particle swarm algorithm to optimize the hyperparameters of the power circuit fault diagnosis model based on the support vector machine, the extreme learning machine, the K-nearest neighbor algorithm, and the random forest, respectively, to obtain the optimal parameter group of each power circuit fault diagnosis model; then, according to the importance sorting order of the alternative fault feature parameters, starting from the first importance sorting, adopt a strategy of gradually increasing the number, and input the corresponding alternative fault feature parameters in each composite fault feature data into each power circuit fault diagnosis model of the above optimal parameter group, and the fault type corresponding to each composite fault feature data is used as the output of the model; calculate the fault diagnosis accuracy of the model corresponding to each number of alternative fault feature parameters; when the increase of the fault diagnosis accuracy of each model is not greater than 10 -6 When , the number of fault characteristic parameters is the optimal value, and the optimal fault characteristic parameter combination is determined accordingly, and an optimal fault characteristic data is obtained from the optimal fault characteristic parameter combination in each composite fault characteristic data, and the optimal characteristic parameter sample data set S1 is obtained from all the optimal fault characteristic data;

[0015] Step 5: Utilize the improved QPSO and the optimal characteristic parameter sample data set S1 in step 4 to optimize the penalty factor and kernel function coefficient of the SVM power circuit fault diagnosis model, and obtain the optimal SVM power circuit fault diagnosis model;

[0016] Step 6: For the power supply circuit of the intelligent release in step 1 connected to the power grid, the voltage signal at the energy storage capacitor end is collected online in real time, and the corresponding value is extracted according to the optimal fault characteristic parameter combination selected in step 4, and then input into the optimal SVM model in step 5. The SVM model outputs the power supply circuit fault type at the time of voltage signal collection to realize online diagnosis of the fault type.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) The present invention deeply analyzes the working principle of the power supply circuit of the intelligent release, and combines the electronic reliability prediction manual and the working profile of the power supply circuit of the intelligent release to analyze the fault mode and impact of the circuit, systematically evaluates the impact degree and occurrence probability of each potential fault mode, scientifically and accurately determines the core object of fault diagnosis, ensures the pertinence and efficiency of fault diagnosis work, and realizes the optimal allocation and utilization of resources.

[0019] (2) The present invention can realize online diagnosis of potential faults in the power supply circuit of the intelligent release without applying constant excitation by mining and optimizing fault characteristic parameters based on the charging and discharging characteristics of the energy storage capacitor and the characteristics of the control loop. The diagnostic process will not interfere with the normal operation of the circuit and has strong practical application value. Fault diagnosis is realized by extracting the overall response of the power supply circuit, selecting the terminal voltage of the energy storage capacitor as the monitoring signal, and extracting the fault characteristic parameters from the monitoring signal, rather than traditionally monitoring the specific parameter values ​​of each component, thereby effectively avoiding the complex component-level monitoring process and significantly improving the diagnostic efficiency.

[0020] (3) The present invention adopts a support vector machine recursive feature elimination method to sort the candidate fault feature parameters by importance, and selects the optimal set of fault feature parameters for fault diagnosis based on the sorting results. The selected fault feature parameters include basic time domain feature parameters and statistical feature parameters. These feature parameters are simple to calculate, easy to extract, and have good generalization and anti-interference capabilities. Compared with feature parameter extraction methods such as wavelet packets and variational mode decomposition, the present invention effectively reduces the computational complexity of feature parameter extraction, reduces resource consumption, and significantly improves the efficiency of fault diagnosis.

[0021] (4) The present invention introduces a QPSO algorithm to optimize the parameters of the SVM model and proposes a power circuit fault diagnosis model based on QPSO-SVM. By introducing quantum mechanical behavior, the problem that the traditional PSO algorithm is prone to fall into local optimality is solved, thereby improving the accuracy of power circuit fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a method flow chart of an embodiment of the method for online diagnosis of potential faults in the power supply circuit of an intelligent release based on QPSO-SVM of the present invention.

[0023] Figure 2 It is a structural schematic diagram of an intelligent release power supply circuit according to an embodiment of the method for online diagnosis of potential faults of an intelligent release power supply circuit based on QPSO-SVM of the present invention.

[0024] Figure 3 The present invention is a flowchart of power circuit failure mode and impact analysis of an embodiment of a method for online diagnosis of potential failures in a power circuit of an intelligent release based on QPSO-SVM.

[0025] Figure 4 yes Figure 2 Schematic diagram of the charging process of the energy storage capacitor in the power supply circuit of the intelligent release.

[0026] Figure 5 yes Figure 2Schematic diagram of the principle of the energy storage capacitor discharge process of the power supply circuit of the intelligent release.

[0027] Figure 6 yes Figure 2 The change curve of the charging and discharging cycle, charging time and discharging time of the energy storage capacitor when the input current of the power supply circuit of the intelligent release changes.

[0028] Figure 7 The present invention is a schematic diagram of an embodiment of a method for online diagnosis of potential faults in a power circuit of an intelligent release based on QPSO-SVM, which uses a power circuit fault diagnosis model based on support vector machine (SVM), extreme learning machine (ELM), K-nearest neighbor algorithm (KNN), and random forest (RF) to obtain the optimal fault feature parameter combination (i.e., the number of features). DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention are given below. The specific embodiments are only used to further illustrate the present invention in detail and do not limit the protection scope of the claims of the present invention.

[0030] The present invention provides a potential fault online diagnosis method for a power circuit of an intelligent release device based on QPSO-SVM (hereinafter referred to as the method), characterized in that the method comprises the following steps:

[0031] Step 1: According to the structure and working principle of the power circuit of the intelligent release, combined with the "Electronic Equipment Reliability Prediction Model and Data Manual", the power circuit is analyzed for failure mode and impact, the potential failure type of the power circuit is determined, and the fault diagnosis objects are selected as the energy storage capacitor C3, the first resistor R1, the third resistor R3, the fourth resistor R4, the fifth resistor R5, and the seventh resistor R7. The power circuit failure type is set according to the different parameter offset ranges of the fault diagnosis objects;

[0032] Step 1.1: Analysis of the working principle of the power supply circuit

[0033] The intelligent release in this embodiment is powered by a self-generated power supply. The structure of the power supply circuit is shown in FIG. Figure 2 As shown, it includes current transformer No. 1 CT1, current transformer No. 2 CT2, current transformer No. 3 CT3, rectifier bridge No. 1 D1, rectifier bridge No. 2 D2, rectifier bridge No. 3 D3, resistor No. 1 R1, resistor No. 2 R2, resistor No. 3 R3, resistor No. 4 R4, resistor No. 5 R5, resistor No. 6 R6, resistor No. 7 R7, diode D4, filter capacitor No. 1 C1, filter capacitor No. 2 C2, filter capacitor No. 4 C4, energy storage capacitor C3, MOSFET (field effect transistor) U5, voltage comparator U2, and linear regulator U7.

[0034] The input terminals of the No. 1 current transformer CT1, the No. 2 current transformer CT2, and the No. 3 current transformer CT3 are respectively connected to the three-phase AC lines A, B, and C, the AC input terminal of the No. 1 rectifier bridge D1 is respectively connected to the two output terminals of the No. 1 current transformer CT1, the AC input terminal of the No. 2 rectifier bridge D2 is respectively connected to the two output terminals of the No. 2 current transformer CT2, the AC input terminal of the No. 3 rectifier bridge D3 is respectively connected to the two output terminals of the No. 3 current transformer CT3, the positive output terminals of the three rectifier bridges D1 to D3 are all connected to the anode of the diode D4, and the negative output terminals of the three rectifier bridges D1 to D3 are grounded;

[0035] The cathode of the diode D4 is connected to the positive electrode of the energy storage capacitor C3, and the other end of the energy storage capacitor C3 is grounded; the input pin No. 1 of the linear regulator U7 is connected to the positive electrode of the energy storage capacitor C3, the output pin No. 3 of the linear regulator U7 is connected to one end of the No. 4 filter capacitor C4, the other end of the No. 4 filter capacitor C4 is grounded, and the pins No. 2 and No. 4 of the linear regulator U7 are grounded; at the same time, the output pin No. 3 of the linear regulator is the connection terminal of the external module (i.e., other modules of the intelligent release).

[0036] A control loop is composed of a MOSFET (field effect transistor) U5, a first filter capacitor C1, a second filter capacitor C2, a voltage comparator U2, a first resistor R1, a second resistor R2, a third resistor R3, a fourth resistor R4, a fifth resistor R5, a sixth resistor R6, and a seventh resistor R7; the specific structure of the control loop is as follows: the drain of the MOSFET U5 is connected to the anode of the diode D4, the gate of the MOSFET U5, one end of the first filter capacitor C1, one end of the sixth resistor R6, and one end of the seventh resistor R7 are all connected to the output end of the voltage comparator U2, and the source of the MOSFET U5, the other end of the first filter capacitor C1, and the other end of the sixth resistor R6 are all grounded.

[0037] One end of the first resistor R1 is connected to the cathode of the diode D4, and the other end of the first resistor R1, the other end of the seventh resistor R7, and one end of the fifth resistor R5 are all connected to one end of the second resistor R2, and the other end of the second resistor R2 is connected to the output end of the voltage comparator U2; the other end of the fifth resistor R5 is grounded. One end of the third resistor R3 is connected to the output pin 3 of the linear regulator U7, and the other end of the third resistor R3 is connected in parallel with one end of the fourth resistor R4 and one end of the second filter capacitor C2, and then connected to the inverting end of the voltage comparator U2; the other end of the fourth resistor R4 and the other end of the second filter capacitor C2 are connected in parallel and then grounded.

[0038] like Figure 2As shown, the current transformer CT senses the current from the AC line, and the current rectified by the diode D4 charges the energy storage capacitor C3. To ensure the stable operation of the subsequent circuit, the voltage of the energy storage capacitor C3 must be maintained within the safe range specified by the module technical specifications. Its upper and lower thresholds are determined by the control circuit according to the energy required by the subsequent load, and the voltage divider circuit and hysteresis comparison circuit composed of resistors R1 and R5 are used to achieve control. When the voltage of the energy storage capacitor C3 is higher than the upper threshold, the voltage divided on the resistor R5 is higher than the voltage at the reverse input terminal of the comparator U2, where the reverse input terminal voltage is obtained by dividing the output voltage of the linear regulator U7 through resistors R3 and R4, triggering the control circuit to output a high-level signal, so that U5 is turned on and U5 enters the working state. At this time, the current stops charging the energy storage capacitor, but flows to the ground through U5; when the voltage of the capacitor C3 is low, the voltage on the resistor R5 is lower than the voltage at the reverse input terminal of the comparator U2, and the control circuit outputs a low-level signal. At this time, U5 is not turned on, which will not affect the power supply to the subsequent modules. In addition to supplying energy to some devices, the voltage output by the energy storage capacitor C3 also outputs voltages of other amplitudes through the linear voltage regulator U7 to meet the different input voltage requirements of multiple modules.

[0039] Step 1.2, Fault mode and impact analysis, clarify the fault diagnosis object. According to the power circuit fault mode and impact analysis method of reference "Luo Yong. Application of FMEA technology in power module design [J]. Quality and Reliability, 2009, (05): 37-40.", the fault mode and impact analysis of the power circuit of the intelligent release is carried out. The fault mode and impact analysis process is as follows: Figure 3 As shown. First, the components to be analyzed are selected in combination with the working principle of the power supply circuit. Then, the main failure modes of the components are determined according to the GB / T 37963-2019 electronic equipment reliability prediction model and data manual. The probability of occurrence of the corresponding failure mode is calculated according to the component failure rate model in the manual, and the impact of these component failure modes on the power supply circuit is determined. The severity level is defined according to the magnitude of the impact. The definition of the severity level is shown in Table 1. Finally, the results of the failure mode and impact analysis of the power supply circuit of the intelligent release are shown in Table 2.

[0040] Table 1 Severity Level

[0041]

[0042] Combined with the comprehensive consideration of the impact of the fault on the power supply circuit and its probability of occurrence, fault diagnosis is performed on fault types with large impact and high probability of occurrence, and online fault diagnosis is performed on the parameter offset faults of the energy storage capacitor C3, the voltage divider resistors R1, R3, R4, R5 and the feedback resistor R7. Compared with open circuit, short circuit and other fault types, the offset faults of capacitor and resistor parameters have a higher probability of occurrence; at the same time, compared with other components such as filter capacitors and current limiting resistors, the impact of the faults of the energy storage capacitor C3, the voltage divider resistors R1, R3, R4, R5 and the feedback resistor R7 on the circuit is more serious. And the faults of these components belong to the potential fault type, that is, they have not yet caused serious damage to the equipment. Timely discovery and handling of these potential faults can effectively avoid the occurrence of catastrophic failures.

[0043] Table 2 Power supply circuit failure mode and impact analysis

[0044]

[0045]

[0046] Step 1.3: According to the fault diagnosis object obtained in step 1.2, set the fault type of the power supply circuit.

[0047] Combined with the reliability of the components themselves and the actual working performance of the power supply circuit, the corresponding fault type judgment standard is set. The energy storage capacitor of the self-generated power supply circuit uses an electrolytic capacitor. Combined with relevant literature, when the capacitance value parameter of the electrolytic capacitor offsets by more than 20%, the reliability of the capacitor can no longer meet the circuit working requirements, and it is considered that a circuit fault has occurred; according to the working performance of the power supply circuit, when the resistance value parameter offsets by more than 10%, the capacitor terminal voltage value of the power supply circuit will not be able to meet the normal working requirements of the circuit, and it can be determined that the circuit has a fault; and it is generally believed that when a circuit fault occurs, the resistance value parameter offsets upward (i.e. increases), and the capacitance value parameter offsets downward (i.e. decreases). The fault types set for the power supply circuit are shown in Table 3.

[0048] Table 3 Fault types of power supply circuit

[0049]

[0050] Step 2, use Pspice software to build a simulation model of the intelligent release power supply circuit described in step 1, select the voltage at the end of the energy storage capacitor C3 as the monitoring signal, and according to the power circuit fault type set in step 1, perform Monte Carlo simulation experimental analysis on the power supply circuit under each fault type, and under each fault type, the sampling range of the input current is set to between 0.5 times and 2 times the rated input current, so as to obtain multiple groups of voltage waveforms at the end of the energy storage capacitor C3 under different fault types.

[0051] Preferably, step 2 specifically includes:

[0052] Step 2.1: According to the working principle diagram of the power supply circuit described in step 1, a simulation model of the power supply circuit is constructed, and the terminal voltage of the energy storage capacitor C3 is selected as the monitoring signal.

[0053] Step 2.2, according to the 7 types of power circuit faults set in step 1, set the corresponding component parameters in the simulation model, and use a uniformly distributed sampling strategy to sample the component parameters for the parameter range of each fault type to ensure that the parameter variation range can be fully covered. Considering that the power circuit will change during actual operation, in order to more realistically simulate the working conditions of the power circuit under different transmission line currents, the input current is sampled at the same time, and the sampling range of the input current is set to 0.5 to 2 times the rated input current. 300 Monte Carlo simulation experiments were performed for each fault type, and a total of 2100 sets of experimental sample data of the voltage waveform at the end of the energy storage capacitor C3 were obtained.

[0054] Step 3, based on the charging and discharging characteristics of the energy storage capacitor, and the influence of the change in the value of the fault diagnosis object on the energy storage capacitor terminal voltage of the energy storage capacitor C3, select the energy storage capacitor discharge time, peak-to-peak value, maximum value, minimum value, voltage mean, standard deviation, and peak factor on the voltage waveform of the energy storage capacitor under each Monte Carlo simulation experiment as the first group of candidate fault feature parameters; the energy storage capacitor charging time, charge and discharge cycle, charge and discharge time ratio, effective value, kurtosis, margin factor, and skewness on the energy storage capacitor terminal voltage waveform are used as the second group of candidate fault feature parameters; extract the above-mentioned first and second groups of candidate fault feature parameter values ​​in the experimental sample data of each group of energy storage capacitor C3 terminal voltage waveform, and combine the two to obtain a composite fault feature data, and use the label of the power supply circuit fault type corresponding to the voltage waveform as the label of the data, and all the composite fault feature data constitute a training sample set S.

[0055] Preferably, step 3 specifically includes:

[0056] Step 3.1: Mining and optimizing fault characteristic parameters based on capacitor charging and discharging characteristics: The charging and discharging process of the energy storage capacitor is as follows: Figure 4-5 As shown, the current transformer CT induces current from the AC line, and the current after diode rectification charges the energy storage capacitor C3. According to the KCL theorem of the circuit, when the energy storage capacitor C3 is charged:

[0057] I in =I c1 +I R (1)

[0058] When the energy storage capacitor C3 is discharged:

[0059] I c2 =I R (2)

[0060] Among them, I in is the input current of the power circuit, that is, the secondary current of the current transformer after rectification, I c1 is the charging current of the capacitor, I R is the power supply load current, I c2 is the current of the capacitor discharge. Therefore, the current when the energy storage capacitor is charged is determined by the input current and the load current. A part of the input current is used to provide energy for the load, and the remaining part is used to charge the energy storage capacitor. The current when the energy storage capacitor is discharged is only related to the load current, but not to the input current.

[0061] According to the characteristics of capacitors, we have:

[0062]

[0063] Therefore, when the energy storage capacitor is charged, there is:

[0064]

[0065] When the energy storage capacitor is discharged, there are:

[0066]

[0067] Among them, Δu c is the voltage change across the energy storage capacitor, which depends on the control loop, Δt1 is the charging time of the energy storage capacitor, Δt2 is the discharging time of the energy storage capacitor, and the charging and discharging cycle of the energy storage capacitor is: T = Δt1 + Δt2. Therefore, when Δu c When determined, the discharge time of the energy storage capacitor depends on the capacitance value and the discharge current I c2 , while I c2 =I R , load current I R The discharge time of the energy storage capacitor depends only on the size of the capacitor C3 and has nothing to do with the input current. Therefore, the discharge time Δt2 of the energy storage capacitor is selected as the alternative fault characteristic parameter to eliminate the influence of the change of the transmission line current on the fault diagnosis of the energy storage capacitor C3, thereby realizing the online fault diagnosis of the power supply circuit. When the input current of the power supply circuit of the intelligent release changes, the charge and discharge cycle, charging time and discharge time of the energy storage capacitor C3 change as shown in the figure below. Figure 6 shown.

[0068] According to the characteristics of the control loop, when the resistance values ​​of resistors R1, R3, R4, R5 and R7 in the control loop change, the peak-to-peak value, maximum value, minimum value, voltage mean, standard deviation and peak factor of the energy storage capacitor terminal voltage will change accordingly and will not be affected by the input current. That is, when the current in the transmission line changes, these characteristic parameters do not change, and are only related to the parameter values ​​of components and have nothing to do with the circuit operating conditions. Therefore, selecting these characteristic parameters as alternative fault characteristic parameters can provide a basis for online fault diagnosis of resistors R1, R3, R4, R5 and R7.

[0069] In addition, when the component to be diagnosed is in a faulty state, the fault characteristic parameters such as the energy storage capacitor charging time, charge and discharge cycle, charge and discharge time ratio, effective value, kurtosis, margin factor and skewness will also change significantly compared to the values ​​when the component is operating normally. Even if affected by the input current, it can provide a classification basis for fault diagnosis. Therefore, this part of the characteristic parameters is also selected as the second group of alternative fault characteristic parameters to provide support for subsequent fault diagnosis.

[0070] Step 3.2: Extraction of candidate fault feature parameters

[0071] Each voltage waveform curve generated in step 2 contains 2100 sets of data, which are saved in the simulation directory in CSV file format. According to the calculation formulas of the above-mentioned alternative fault characteristic parameters, MTALAB is used to extract the voltage waveform at the energy storage capacitor end under each Monte Carlo simulation experiment to obtain the above-mentioned first and second groups of alternative fault characteristic parameters.

[0072] Step 3.3: Obtain composite fault feature data

[0073] The first set of candidate fault characteristic parameters of a voltage waveform curve is combined with the second set of candidate fault characteristic parameters to obtain a composite fault characteristic data X i , X i ={x1,x2,…,x k ,…,x m}, where i = 1, 2, ..., 2100, m is the total number of candidate fault characteristic parameters, x k represents the characteristic value of the kth candidate fault characteristic parameter, and a total of 2100 sets of candidate fault characteristic parameter sets are obtained. Each composite fault characteristic data is assigned a classification label v i ,v i ∈{1, 2, ..., l}, l is the total number of power circuit fault categories, and the value of l in this embodiment is 7. Therefore, the training sample set is obtained

[0074] Step 4: Use the support vector machine recursive feature elimination method (SVM-REF) to sort the importance of the alternative fault feature parameters of all composite fault feature data in the training sample set S, and obtain the importance sorting result of the alternative fault feature parameters; at the same time, use the training sample set s and the particle swarm algorithm to optimize the hyperparameters of the power circuit fault diagnosis model based on the support vector machine (SVM), extreme learning machine (ELM), K-nearest neighbor algorithm (KNN), and random forest (RF), and obtain the optimal parameter group of each power circuit fault diagnosis model; then, according to the importance sorting order of the alternative fault feature parameters, starting from the first importance, adopt a strategy of gradually increasing the number, and input the corresponding alternative fault feature parameters in each composite fault feature data into each power circuit fault diagnosis model of the above optimal parameter group, and the fault type corresponding to each composite fault feature data is used as the output of the model; calculate the fault diagnosis accuracy of the model corresponding to each number of alternative fault feature parameters. When the increase in the fault diagnosis accuracy of each model is no more than 10 -6 When , the number of fault feature parameters is the optimal value, and the optimal fault feature parameter combination is determined accordingly, and an optimal fault feature data is obtained from the optimal fault feature parameter combination in each composite fault feature data, and the optimal feature parameter sample data set S1 is obtained from all the optimal fault feature data.

[0075] Step 4.1 In order to optimize the selection process of alternative fault feature parameters, the alternative fault feature parameters obtained in step 3 are sorted according to the contribution of the normal vector to the hyperplane. The process first removes the alternative fault feature parameters that contribute the least to the definition of the hyperplane, and then retrains the model using the remaining set of alternative fault feature parameters, and updates the sorting of the fault feature parameters based on the performance of the new model. Repeat this process, removing the currently ranked alternative fault feature parameter at the bottom of each cycle until the original set of alternative fault feature parameters is completely traversed. Since power circuit fault diagnosis is a multi-classification problem, a multi-class SVM-RFE feature parameter selection strategy is adopted. This strategy decomposes the multi-class problem into multiple independent two-class sub-problems in a one-to-one manner. Each sub-problem uses SVM-RFE to select feature parameters independently, which not only simplifies the complexity of the problem, but also ensures the pertinence and accuracy of feature parameter selection. Then, multiple SVM-RFE algorithms are used to obtain multiple sorting criterion scores, and these scores are added to obtain the total score of the sorting criterion. The total score is used as the basis for sorting the candidate fault characteristic parameters, and the candidate fault characteristic parameter with the smallest total score is deleted in each iteration until the set of candidate fault characteristic parameters is empty.

[0076] The specific process is as follows:

[0077] Step (1): Initialize the feature parameter index sequence T = [1, 2, ..., m] and the candidate fault feature parameter importance ranking set R = [ ];

[0078] Step (2): Input training sample set X i ={x1,x2,…,x k ,…,x m}, where i = 1, 2, ..., 2100; m is the total number of candidate fault characteristic parameters; x k represents the characteristic value of the kth candidate fault characteristic parameter, v i is the classification label, v i ∈{1, 2, …, l}, l is the total number of power circuit fault categories;

[0079] Step (3): Generate l×(l-1) / 2 training sub-sample sets based on the training sample set S:

[0080]

[0081] Step (4): Repeat the following process until

[0082] A. Obtain (l×(l-1) / 2) new training sample sets S j ,

[0083] B. Use S j Train SVM to get the weight vector ω i , i=1,2,…,l;

[0084] C. Calculate the total score of the sorting criteria:

[0085]

[0086] D. Find the candidate fault feature parameter with the smallest ranking score:

[0087] p=argmin C k (8)

[0088] E. Update the candidate fault feature parameter set and add p to the last sorted position in the set R;

[0089] F. Remove the candidate fault feature parameter with the smallest score from the current remaining data set to obtain a new training sample set S' j ,

[0090] Step (5): Based on the new training sample set S' obtained in step (4) jRepeat the process of step (3) to step (4) until the candidate fault feature parameter with the second highest ranking score is screened out. At this point, the ranking position in the candidate fault feature parameter importance ranking set R is confirmed, and the candidate fault feature parameters are ranked by importance.

[0091] Finally, the importance ranking results of the candidate fault feature parameters are obtained, as shown in Table 4:

[0092] Table 4 Importance ranking

[0093]

[0094]

[0095] Step 4.2: Establish circuit fault diagnosis models based on support vector machine (SVM), extreme learning machine (ELM), K-nearest neighbor algorithm (KNN), and random forest (RF), respectively, and optimize the hyperparameters of these models through particle swarm algorithm (PSO). The process of particle swarm algorithm is as follows:

[0096] Assume that the entire population contains N particles, and the goal is to maximize a certain fitness function. For the i-th particle, the update of its individual optimal position and the selection method of the global optimal position of the entire particle swarm can be described as follows:

[0097]

[0098] P g (t) = argmax{f(P)}

[0099] f(P) is the accuracy of fault diagnosis, and t+1 refers to the next evolutionary generation after the tth evolutionary generation.

[0100] During the entire iteration process, the position and velocity update equations of each particle are:

[0101] V i (t+1)=wV i (t)+c1r1(P i -X i (t))+c2r2(P g -X i (t))

[0102] X i (t+1)=X i (t)+V i (t+1))

[0103] c1 and c2 are acceleration factors, r1 and r2 are random numbers between [0,1], and w is the inertia weight, which decreases linearly with the number of evolutions in the standard particle swarm algorithm.

[0104] The algorithm evaluates the classification results of each model for the training sample set S under different hyperparameter combinations, and searches for potential optimal hyperparameter combinations based on these results. The particle swarm algorithm runs 100 iterations, with the model's fault diagnosis accuracy as the objective function value. The SVM model uses a Gaussian kernel function, and after optimization, the penalty coefficient C = 35.6 and the kernel function γ = 7.13. The activation function of ELM is selected as Softplus, and a regularization strategy is introduced. After optimization by the particle swarm algorithm, the two key parameters of the ELM model are determined to be CE = 4.34, which is used to adjust the strength of regularization, and LE = 183, which represents the number of nodes in the hidden layer. In addition, the elements of the weight matrix and bias matrix of ELM are designed to be uniformly randomly generated in the interval (-1, 1) to ensure that the initialization of the model has sufficient diversity; the KNN model is used to predict the number of nearest neighbors of the target point n_neighbors = 5; the maximum depth of the RF model decision tree max_depth = 35, and the number of decision trees n_estimators = 168.

[0105] Step 4.3 follows the order indicated by the ranking result of the importance of the candidate fault feature parameters, and adopts a step-by-step strategy to expand the set of fault feature parameters, that is, gradually increase the number of fault feature parameters and input the updated fault feature parameter sample set into the fault diagnosis model of the above optimal hyperparameter combination in sequence, and take the corresponding fault type as the output of the model; calculate the fault diagnosis accuracy of the model. At first, as the number of fault feature parameters increases, the fault diagnosis accuracy will gradually increase. When the increase in the fault diagnosis accuracy is no more than 10 -6 When , the number of fault characteristic parameters at this time is the optimal value, and the optimal fault characteristic parameter set is determined accordingly;

[0106] The accuracy of fault diagnosis is as follows: Figure 7 As shown in the figure, initially, as the number of fault feature parameters increases, the diagnostic accuracy of each fault diagnosis model gradually increases. When the number of fault feature parameters is 8, the diagnostic accuracy of the four models is close to the highest level, and as the number of fault feature parameters increases, the fault diagnosis accuracy no longer has an upward trend. Based on this, the optimal number of fault feature parameter sets is finally selected as 8, and the selected fault feature parameters are: peak-to-peak value, mean value, effective value, charge and discharge cycle, capacitor discharge time, charge and discharge time ratio, capacitor charging time, and maximum value. Based on the selected optimal fault feature parameters, the optimal feature parameter sample data set is established Among them, Z i It is a set of the above 8 fault characteristic parameters.

[0107] Step 5: Using the improved QPSO (quantum particle swarm algorithm) and the optimal characteristic parameter sample data set S1 in step 4, optimize the penalty factor (c) and kernel function coefficient (g) of the SVM power circuit fault diagnosis model to obtain the optimal SVM power circuit fault diagnosis model. The specific process of this step is:

[0108] The optimal feature parameter sample data set S1 is divided into a training set and a test set in a ratio of 7:3 according to the number of optimal fault feature data. The specific process of optimizing the penalty factor (c) and kernel function coefficient (g) of the SVM power circuit fault diagnosis model is as follows:

[0109] 1) Improved QPSO initialization; set the total number of particles, evolutionary generations, compression and expansion factor, initial position of particles, and optimization range, and map the penalty factor (c) and kernel function coefficient (g) of the SVM model to be optimized to the two-dimensional position of each particle;

[0110] 2) The classification accuracy of the SVM model for the training set with fault type labels is used as the fitness function to calculate the fitness of each particle in the particle swarm;

[0111] 3) Calculate the individual optimal position of each particle and the global optimal position of the group, calculate the average of the individual optimal positions of all particles, and update the position of each particle based on the improved QPSO;

[0112] Assume that in an N-dimensional space, there are M particles forming a particle swarm. If the current number of iterations is t, then the position of the i-th particle at the t-th iteration can be expressed as:

[0113] x i (t) = (x i1 (t),x i2 (t),…,x iN (t)),i=1,2,…,M (9)

[0114] The optimal position of the i-th particle is:

[0115]

[0116] Where g is the subscript of the best particle in the group, g∈{1,2,…,M};

[0117] The optimal position of all particles in the group is:

[0118] G(t)=(G i1 (t),G i2 (t),…,G iN (t)),i=1,2,…,M (11)

[0119] In each iteration cycle, before the particle performs the position update operation, it first performs a key step: re-evaluating and determining the optimal position within the group. This process includes comparing the individual optimal position of the current particle with the currently recorded global optimal position G(t). If the individual optimal position of the current particle is better than the global optimal position in terms of performance indicators, the global optimal position G(t) will be updated to the individual optimal position of the particle. Otherwise, if the individual optimal position of the current particle fails to surpass the global optimal position, the global optimal position G(t) remains unchanged.

[0120] The specific particle evolution rules are:

[0121]

[0122] Where α is the contraction expansion coefficient, which is set to decrease linearly with the number of evolutions; u ij (t) follows a random distribution between 0 and 1.

[0123] c j (t) is the mean of the optimal positions of all individual particles, and its expression is:

[0124]

[0125] p dj (t) Local attraction factor, which is determined by the individual optimal position and the global optimal position. In order to ensure the convergence of the algorithm, its expression is:

[0126]

[0127] In the formula, Obeys a random distribution between 0 and 1.

[0128] 4) Repeat step 2 Step 4, until the evolutionary generation set by initialization is reached;

[0129] 5) Output the global optimal position as the final optimization result, and obtain the optimal penalty factor and kernel function coefficient value of the SVM model;

[0130] 6) Input the test set with fault type labels into the SVM model with the best penalty factor and kernel function coefficient value. The SVM model outputs the fault type of the corresponding data sample. When the accuracy is above 95%, the SVM model is optimal. Otherwise, adjust the initialization method of the improved QPSO and repeat the process from step 1 to step 6 until the optimal SVM model is obtained.

[0131] Step 6: For the power supply circuit of the intelligent release in step 1 connected to the power grid, the voltage signal at the end of the energy storage capacitor C3 is collected online in real time, and the corresponding value is extracted according to the optimal fault characteristic parameter combination selected in step 4, and then input into the optimal SVM model in step 5. The SVM model outputs the power supply circuit fault type at the time of voltage signal collection to realize online diagnosis of the fault type.

[0132] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A potential fault online diagnosis method for the power circuit of an intelligent release based on QPSO-SVM, characterized in that: The method comprises the following steps: Step 1: According to the structure and working principle of the power circuit of the intelligent release, combined with the "Electronic Equipment Reliability Prediction Model and Data Manual", the power circuit is analyzed for failure mode and impact, the potential failure type of the power circuit is determined, and the fault diagnosis objects are selected as energy storage capacitors, resistors No. 1, No. 3, No. 4, No. 5, and No.

7. The power circuit failure type is set according to the different parameter offset ranges of the fault diagnosis objects; The intelligent release power supply circuit includes a No. 1 current transformer, a No. 2 current transformer, a No. 3 current transformer, a No. 1 rectifier bridge, a No. 2 rectifier bridge, a No. 3 rectifier bridge, a No. 1 resistor, a No. 2 resistor, a No. 3 resistor, a No. 4 resistor, a No. 5 resistor, a No. 6 resistor, a No. 7 resistor, a diode, a No. 1 filter capacitor, a No. 2 filter capacitor, a No. 4 filter capacitor, an energy storage capacitor, a MOSFET, a voltage comparator, and a linear regulator; The input ends of the No. 1 current transformer, the No. 2 current transformer, and the No. 3 current transformer are respectively connected to the three-phase AC lines A, B, and C, the AC input end of the No. 1 rectifier bridge is respectively connected to the two output ends of the No. 1 current transformer, the AC input end of the No. 2 rectifier bridge is respectively connected to the two output ends of the No. 2 current transformer, the AC input end of the No. 3 rectifier bridge is respectively connected to the two output ends of the No. 3 current transformer, the positive output ends of the three rectifier bridges are all connected to the anodes of the diodes, and the negative output ends of the three rectifier bridges are grounded; The cathode of the diode is connected to the positive electrode of the energy storage capacitor, and the other end of the energy storage capacitor is grounded; the input pin 1 of the linear regulator is connected to the positive electrode of the energy storage capacitor, the output pin 3 of the linear regulator is connected to one end of the filter capacitor No. 4, and the other end of the filter capacitor No. 4 is grounded, and the pins 2 and 4 of the linear regulator are grounded; at the same time, the output pin 3 of the linear regulator is the wiring terminal of the external module; The drain of MOSFET is connected to the anode of the diode, the gate of MOSFET, one end of the No. 1 filter capacitor, one end of the No. 6 resistor, and one end of the No. 7 resistor are all connected to the output end of the voltage comparator, and the source of MOSFET, the other end of the No. 1 filter capacitor, and the other end of the No. 6 resistor are all grounded; One end of the No. 1 resistor is connected to the cathode of the diode, the other end of the No. 1 resistor, the other end of the No. 7 resistor, and one end of the No. 5 resistor are all connected to one end of the No. 2 resistor, and the other end of the No. 2 resistor is connected to the output end of the voltage comparator; the other end of the No. 5 resistor is grounded; one end of the No. 3 resistor is connected to the No. 3 output pin of the linear regulator, and the other end of the No. 3 resistor is connected in parallel with one end of the No. 4 resistor and one end of the No. 2 filter capacitor, and then connected to the inverting end of the voltage comparator; the other end of the No. 4 resistor and the other end of the No. 2 filter capacitor are connected in parallel and then grounded; Step 2, using Pspice software to build a simulation model of the power supply circuit of the intelligent release described in step 1, selecting the voltage at the energy storage capacitor terminal as the monitoring signal, and performing a Monte Carlo simulation experiment analysis on the power supply circuit under each fault type according to the power supply circuit fault type set in step 1, and under each fault type, the sampling range of the input current is set to between 0.5 times and 2 times the rated input current, thereby obtaining multiple groups of voltage waveforms at the energy storage capacitor terminal under different fault types; Step 3: Based on the charge and discharge characteristics of the energy storage capacitor and the influence of the change of the value of the fault diagnosis object on the energy storage capacitor terminal voltage of the energy storage capacitor C3, the energy storage capacitor discharge time, peak-to-peak value, maximum value, minimum value, voltage mean, standard deviation, and peak factor on the voltage waveform of the energy storage capacitor under each Monte Carlo simulation experiment are selected as the first group of candidate fault characteristic parameters; The energy storage capacitor charging time, charge and discharge cycle, charge and discharge time ratio, effective value, kurtosis, margin factor and skewness on the voltage waveform at the energy storage capacitor terminal are used as the second group of candidate fault characteristic parameters; the first group and the second group of candidate fault characteristic parameter values ​​in the experimental sample data of each group of energy storage capacitor C3 terminal voltage waveform are extracted, and the two are combined to obtain a composite fault characteristic data, and the label of the power supply circuit fault type corresponding to the voltage waveform is used as the label of the data, and all the composite fault characteristic data constitute a training sample set S; Step 4: Use the support vector machine recursive feature elimination method to sort the importance of the alternative fault feature parameters of all composite fault feature data in the training sample set S, and obtain the importance sorting result of the alternative fault feature parameters; at the same time, use the training sample set S and the particle swarm algorithm to optimize the hyperparameters of the power circuit fault diagnosis model based on the support vector machine, the extreme learning machine, the K-nearest neighbor algorithm, and the random forest, respectively, to obtain the optimal parameter group of each power circuit fault diagnosis model; then, according to the importance sorting order of the alternative fault feature parameters, starting from the first importance sorting, adopt a strategy of gradually increasing the number, and input the corresponding alternative fault feature parameters in each composite fault feature data into each power circuit fault diagnosis model of the above optimal parameter group, and the fault type corresponding to each composite fault feature data is used as the output of the model; calculate the fault diagnosis accuracy of the model corresponding to each number of alternative fault feature parameters; when the increase of the fault diagnosis accuracy of each model is not greater than 10 -6 When , the number of fault characteristic parameters is the optimal value, and the optimal fault characteristic parameter combination is determined accordingly, and an optimal fault characteristic data is obtained from the optimal fault characteristic parameter combination in each composite fault characteristic data, and the optimal characteristic parameter sample data set S1 is obtained from all the optimal fault characteristic data; Step 5: Utilize the improved QPSO and the optimal characteristic parameter sample data set S1 in step 4 to optimize the penalty factor and kernel function coefficient of the SVM power circuit fault diagnosis model, and obtain the optimal SVM power circuit fault diagnosis model; Step 6: For the power supply circuit of the intelligent release in step 1 connected to the power grid, the voltage signal at the energy storage capacitor end is collected online in real time, and the corresponding value is extracted according to the optimal fault characteristic parameter combination selected in step 4, and then input into the optimal SVM model in step 5. The SVM model outputs the power supply circuit fault type at the time of voltage signal collection to realize online diagnosis of the fault type.

2. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: In step 1, the power circuit fault type is set as shown in the following table. Table 3 Fault types of power supply circuit 3. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: In step 2, 300 Monte Carlo simulation experiments are performed for each fault type.

4. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: In step 3, the first set of candidate fault characteristic parameters and the second set of candidate fault characteristic parameters are extracted from the voltage waveform at the energy storage capacitor terminal in each Monte Carlo simulation experiment using MTALAB.

5. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: In step 4, the specific process of using the support vector machine recursive feature elimination method to sort the importance of candidate fault feature parameters of all composite fault feature data in the training sample set S is as follows: Step (1): Initialize the feature parameter index sequence T = [1, 2, ..., m] and the candidate fault feature parameter importance ranking set R = []. Step (2): Input training sample set X i ={x1,x2,…,x k ,…,x m }, where i = 1, 2, ..., 2100; m is the total number of candidate fault characteristic parameters; x k represents the characteristic value of the kth candidate fault characteristic parameter, v i is the classification label, v i ∈{1, 2, …, l}, l is the total number of power circuit fault categories; Step (3): Generate l×(l-1) / 2 training sub-sample sets based on the training sample set S: Step (4): Repeat the following process until A. Obtain (l×(l-1) / 2) new training sample sets S j , B. Use S j Train SVM to get the weight vector ω i , i=1,2,…,l; C. Calculate the total score of the sorting criteria: D. Find the candidate fault feature parameter with the smallest ranking score: p=argmin C k (8) E. Update the candidate fault feature parameter set and add p to the last sorted position in the set R; F. Remove the candidate fault feature parameter with the smallest score from the current remaining data set to obtain a new training sample set S′ j , Step (5): Based on the new training sample set s' obtained in step (4) j Repeat the process of step (3) to step (4) until the candidate fault feature parameter with the second highest ranking score is screened out. At this point, the ranking position in the candidate fault feature parameter importance ranking set R is confirmed, and the candidate fault feature parameters are ranked by importance.

6. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 5 is characterized in that: The importance ranking results of the candidate fault feature parameters are shown in Table 4. Table 4 Importance ranking 7. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: In step 4, when the training sample set S and the particle swarm algorithm are used to optimize the hyperparameters of the power circuit fault diagnosis model based on support vector machine, extreme learning machine, K-nearest neighbor algorithm, and random forest, the particle swarm algorithm runs 100 iterations, and the fault diagnosis accuracy of the model is used as the objective function value.

8. The potential fault online diagnosis method of the power supply circuit of the intelligent release based on QPSO-SVM according to claim 1 is characterized in that: The specific process of step 5 is: divide the optimal feature parameter sample data set S1 into a training set and a test set in a ratio of 7:3 according to the number of optimal fault feature data; the specific process of optimizing the penalty factor and kernel function coefficient of the SVM power circuit fault diagnosis model is: 1) Improved QPSO initialization; set the total number of particles, evolutionary generations, compression and expansion factor, initial position of particles, and optimization range, and map the penalty factor and kernel function coefficient of the SVM model to be optimized to the two-dimensional position of each particle; 2) The classification accuracy of the SVM model for the training set with fault type labels is used as the fitness function to calculate the fitness of each particle in the particle swarm; 3) Calculate the individual optimal position of each particle and the global optimal position of the group, calculate the average of the individual optimal positions of all particles, and update the position of each particle based on the improved QPSO; Assume that in an N-dimensional space, there are M particles forming a particle swarm. If the current number of iterations is t, then the position of the i-th particle at the t-th iteration can be expressed as: x i (t)=(x i1 (t),x i2 (t),…,x iN (t)),i=1,2,…,M (9) The optimal position of the i-th particle is: Where g is the subscript of the best particle in the group, g∈{1,2,…,M}; The optimal position of all particles in the group is: G(t)=(G i1 (t),G i2 (t),…,G iN (t)),i=1,2,…,M (11) In each iteration cycle, before the particle performs the position update operation, it will first perform a key step: re-evaluate and determine the optimal position within the group. This process includes comparing the individual optimal position of the current particle with the currently recorded global optimal position G(t); if the individual optimal position of the current particle is better than the global optimal position in terms of performance indicators, the global optimal position G(t) will be updated to the individual optimal position of the particle; otherwise, if the individual optimal position of the current particle fails to surpass the global optimal position, the global optimal position G(t) remains unchanged; The specific particle evolution rules are: Where α is the contraction expansion coefficient, which is set to decrease linearly with the number of evolutions; u ij (t) follows a random distribution between 0 and 1; c j (t) is the mean of the optimal positions of all individual particles, and its expression is: p dj (t) Local attraction factor, which is determined by the individual optimal position and the global optimal position. In order to ensure the convergence of the algorithm, its expression is: In the formula, Obey the random distribution of 0-1; 4) Repeat step 2 Step 4, until the evolutionary generation set by initialization is reached; 5) Output the global optimal position as the final optimization result, and obtain the optimal penalty factor and kernel function coefficient value of the SVM model; 6) Input the test set with fault type labels into the SVM model with the best penalty factor and kernel function coefficient value. The SVM model outputs the fault type of the corresponding data sample. When the accuracy is above 95%, the SVM model is optimal. Otherwise, adjust the initialization method of the improved QPSO and repeat the process from step 1 to step 6 until the optimal SVM model is obtained.

Images