Fault diagnosis method and device for electric power new energy equipment based on small samples

Abstract

The present invention provides a method and device for diagnosing faults of electric power renewable energy equipment based on small samples, which relates to the technical field of data processing, including: obtaining an original sample data set of electric power renewable energy equipment, wherein the original sample data set includes: a sensor data set and a state label set corresponding to each sensor data; based on a generative adversarial network under an improved Cartesian coordinate system, performing data expansion processing on the original sample data set to expand the original sample data set and determine a target sample data set; based on the target sample data set, performing model training processing on a preset electric power renewable energy equipment fault diagnosis model to determine a target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power renewable energy equipment fault data to be classified and determine the fault classification result. The present invention can significantly improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for diagnosing faults of new energy electric power equipment based on a small sample. Background Art

[0002] In the field of electric power and renewable energy, stable operation and high efficiency of equipment are crucial. Electric power and renewable energy equipment, such as wind turbines, solar panels, energy storage equipment, etc., are highly complex and expensive investments, and they will continue to generate a large amount of operating data during operation. These data come from various sensors of the equipment, such as temperature sensors, pressure sensors, current and voltage measuring devices, etc. Correctly analyzing and interpreting these data can effectively help detect the health of the equipment, predict potential failures, and take preventive measures to ensure the long life and maximum efficiency of the equipment.

[0003] Since high-quality labeled data is usually scarce, equipment failure is not a common event, which makes fault diagnosis based on traditional machine learning methods a small sample problem. At present, relevant technologies propose that equipment faults can be diagnosed after data expansion through traditional data expansion methods such as rotation, scaling and cropping. However, in the above schemes, traditional data expansion methods lack the ability to generate new samples that are high-quality, diverse and highly correlated with the original data, and existing classification methods are difficult to effectively process small sample data or capture complex nonlinear relationships in the data. In addition, traditional neural networks mainly rely on gradient descent or its variants for optimization, which are easily affected by gradient vanishing, gradient explosion or local optimality, resulting in difficulty in capturing and identifying key features related to faults when diagnosing faults of small sample electric power new energy equipment, which in turn leads to low accuracy of fault diagnosis. Summary of the invention

[0004] In view of this, an object of the present invention is to provide a method and device for diagnosing faults of new energy electric power equipment based on a small sample, which can significantly improve the accuracy of fault diagnosis.

[0005] In the first aspect, an embodiment of the present invention provides a new energy equipment fault diagnosis method based on small samples, the method comprising: obtaining an original sample data set of electric power new energy equipment, wherein the original sample data set comprises: a sensor data set and a state label set corresponding to each sensor data; based on a generative adversarial network under an improved Cartesian coordinate system, performing data expansion processing on the original sample data set to expand the original sample data set and determine a target sample data set; based on the target sample data set, performing model training processing on a preset electric power new energy equipment fault diagnosis model to determine a target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power new energy equipment fault data to be classified and determine the fault classification result.

[0006] In one embodiment, a generative adversarial network based on an improved Cartesian coordinate system includes: a generator, a discriminator and a coordinate mapper. Based on the generative adversarial network under the improved Cartesian coordinate system, data expansion processing is performed on the original sample data set to expand the original sample data set, and the step of determining the target sample data set includes: mapping the original sample data set to a high-dimensional space through the coordinate mapper, and using the generator to perform data expansion processing on the original sample data set in the high-dimensional space to determine the first expanded data; performing inverse mapping processing on the first expanded data through the coordinate mapper to map the first expanded data back to the original dimension to determine the second expanded data; performing data identification processing on the second expanded data through the discriminator to determine the target sample data set.

[0007] In one embodiment, a generator is used to perform data enumeration processing on an original sample data set in a high-dimensional space to determine the step of first expanded data, including: determining an anchor point set corresponding to the original sample data set based on an entropy weighted algorithm, wherein the anchor point set includes anchor points corresponding to main features of the data; and through the generator, based on the coordinate conversion rules of Cartesian coordinates in high-dimensional space, with each anchor point as the center, performing data enumeration processing on the original sample data to determine the first expanded data.

[0008] In one embodiment, the step of determining the anchor point set corresponding to the original sample data set based on an entropy weighted algorithm includes: determining a candidate anchor point set in a high-dimensional space through a clustering algorithm; performing entropy weighting processing on the candidate anchor point set based on the entropy of the category to which each data point in the original sample data set belongs through an entropy weighted algorithm to determine the entropy weight corresponding to each candidate anchor point in the candidate anchor point set; determining the anchor point set corresponding to the original sample data set based on the entropy weight corresponding to each candidate anchor point and a preset weight threshold.

[0009] In one embodiment, based on a target sample data set, a preset electric power new energy equipment fault diagnosis model is trained to determine the target fault diagnosis model, including: performing feature extraction on the target sample data set according to a neural network optimization algorithm of droplet impact and splashing mechanism to determine a target feature set; inputting the target sample data set and the target feature set into a preset classifier, performing classification training on the classifier based on an improved low-rank approximate extreme learning machine algorithm to determine a fault classification model.

[0010] In one embodiment, according to the neural network optimization algorithm of droplet impact and splash mechanism, feature extraction processing is performed on the target sample data set, and the step of determining the target feature set includes: performing parameter update processing on the parameter set in the preset initialized neural network model through the neural network optimization algorithm of droplet impact and splash mechanism to determine the target neural network model; using the target neural network model to perform feature extraction processing on the sample data to determine the target feature set.

[0011] In one embodiment, a parameter set in a preset initialized neural network model is updated through a neural network optimization algorithm for droplet impact and splashing mechanism, and the step of determining a target neural network model includes: mapping the error of the neural network to the energy of the droplet, and determining the droplet update position and the neural network parameters corresponding to the droplet update position by simulating the droplet splashing information generated by the droplet impact, wherein the droplet splashing information includes: the splashing droplet direction and the splashing droplet distance; performing adaptive adjustment processing and convergence detection processing on the neural network parameters corresponding to the droplet update position, determining the target neural network parameters, and determining the target neural network model based on the target neural network parameters.

[0012] In the second aspect, an embodiment of the present invention also provides a new energy equipment fault diagnosis device based on small samples, the device comprising: a sample data acquisition module, which obtains the original sample data set of the electric power new energy equipment, wherein the original sample data set comprises: a sensor data set and a state label set corresponding to each sensor data; a data expansion module, which performs data expansion processing on the original sample data set based on a generative adversarial network under an improved Cartesian coordinate system to expand the original sample data set and determine a target sample data set; a model training module, which performs model training processing on a preset electric power new energy equipment fault diagnosis model based on the target sample data set to determine a target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power new energy equipment fault data to be classified and determine the fault classification result.

[0013] In a third aspect, an embodiment of the present invention further provides a server, comprising a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement any one of the methods provided in the first aspect.

[0014] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0015] The embodiments of the present invention bring the following beneficial effects:

[0016] The embodiment of the present invention provides a method and device for diagnosing faults of electric power renewable energy equipment based on small samples. After obtaining the original sample data set of the electric power renewable energy equipment, the method performs data expansion processing on the original sample data set based on the improved generative adversarial network under the Cartesian coordinate system to expand the original sample data set, determine the target sample data set, and perform model training processing on the preset electric power renewable energy equipment fault diagnosis model based on the target sample data set to determine the target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the fault data of the electric power renewable energy equipment to be classified, and determine the fault classification result. When the above method performs data expansion on the original sample data set of the electric power renewable energy equipment with a small sample, the original sample data set is processed based on the generative adversarial network under the improved Cartesian coordinate system. Compared with the related art, the data expansion is performed by the traditional data expansion methods such as rotation, scaling and cropping, so that high-quality, diverse and highly correlated new samples with the original data can be generated, thereby optimizing the training model and significantly improving the accuracy of fault diagnosis.

[0017] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A schematic diagram of a flow chart of a new energy equipment fault diagnosis method based on a small sample provided by an embodiment of the present invention;

[0021] Figure 2 A flowchart of a model training method provided by an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of the structure of a new energy equipment fault diagnosis device based on a small sample provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the structure of a server provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] At present, in the field of electric power and renewable energy, the stable operation and high efficiency of equipment are crucial. Electric power and renewable energy equipment, such as wind turbines, solar panels, energy storage equipment, etc., are highly complex and expensive investments, and they will continue to generate a large amount of operating data during operation. These data come from various sensors of the equipment, such as temperature sensors, pressure sensors, current and voltage measuring devices, etc. Correctly analyzing and interpreting these data can effectively help detect the health of the equipment, predict potential failures, and take preventive measures to ensure the long life and maximum efficiency of the equipment. Since high-quality labeled data is usually scarce, equipment failure is not a common event, which makes fault diagnosis based on traditional machine learning methods a small sample problem. Related technologies propose that equipment failures can be diagnosed after data expansion through traditional data expansion methods such as rotation, scaling, and cropping. However, in the above scheme, the traditional data expansion method lacks the ability to generate high-quality, diverse and highly correlated new samples with the original data, and the existing classification methods are difficult to effectively process small sample data or capture complex nonlinear relationships in the data. In addition, the traditional neural network mainly relies on the gradient descent method or its variants for optimization, which is easily affected by gradient vanishing, gradient explosion or local optimality, resulting in difficulty in capturing and identifying key features related to the fault when diagnosing faults of small sample electric power new energy equipment, which leads to low accuracy of fault diagnosis. Based on this, the new energy equipment fault diagnosis method based on small samples provided by the present invention realizes efficient data expansion, optimized parameter adjustment, fast and efficient classification and better model generalization. By generating adversarial networks, the quantity and diversity of small sample data are effectively enhanced, thereby improving the generalization ability and accuracy of the model.

[0026] See also Figure 1 The flowchart of a method for diagnosing faults of new energy equipment based on a small sample is shown, and the method mainly includes the following steps S102 to S106:

[0027] Step S102 obtains the original sample data set of the new energy power equipment, wherein the original sample data set includes: a sensor data set and a state label set corresponding to each sensor data. In one embodiment, the new energy power equipment includes a wind turbine, a solar panel, an energy storage device, etc., which continuously generates a large amount of operation data during operation. These data come from various sensors of the equipment, such as temperature sensors, pressure sensors, current and voltage measuring devices, etc. The collected data is stored in a structured table form, each row represents multiple sensor readings at a time point, and each column corresponds to a specific sensor or data attribute. Specifically, the data attributes and characteristics can be expressed as: : ambient temperature, in °C; : Loaded power, in kW; : Phase current, unit is A; : Phase voltage, in V; : wind speed, in m / s; : The power generated, in kW; : rotor angle, in degrees; : device temperature, in °C; : Sensor status (normal / abnormal); : Fault label, used to identify whether a device fault occurs and its fault type.

[0028] In practical applications, consider a data sample of a device in a certain state. The attribute values ​​of the sample include the following: °C; kW; A; V; m / s; kW; °; °C; =Normal; = No fault. This data sample is a new energy power equipment in normal working state. In order to train and test the model, a large number of the above data samples need to be collected and annotated by experts to determine whether there is a fault and the type of fault. In particular, the number of data attributes in the embodiment of the present invention is 10, but the number of data attributes in the actual application scenario may be more, even up to hundreds. Therefore, it is unreliable to manually judge whether the equipment is faulty. Based on this, the new energy power equipment fault diagnosis method based on small samples proposed in the present invention uses artificial intelligence algorithm to realize the fault diagnosis of new energy equipment.

[0029] In another embodiment, data preprocessing is required before starting model training. For the classification of faults of new energy power equipment, the purpose of preprocessing is to ensure that each data point in the data set is complete, accurate, and in the correct format.

[0030] First, data normalization is performed. Considering the various data attributes of electric power and new energy equipment, normalization is the key to ensure that all attributes are on the same scale. It can be expressed as:

[0031]

[0032] in, is the original data, is the normalized data, and The data sets are The minimum and maximum values ​​of the corresponding attribute.

[0033] Furthermore, missing value processing is performed. Since there may be some missing values ​​in the equipment failure data, the present invention uses the mean interpolation method to process the missing values, which can be expressed as:

[0034]

[0035] in, is the filled data, is the number of non-missing data, It is non-missing data.

[0036] Furthermore, outlier processing is performed. The equipment failure data may contain outliers, such as erroneous readings caused by sensor failure. The median filter is used to process the outliers to maintain the stability of the data, which can be expressed as:

[0037]

[0038] in, is the processed data, is half the filter window size, is the data located in the window.

[0039] Step S104, based on the generative adversarial network under the improved Cartesian coordinate system, the original sample data set is subjected to data expansion processing to expand the original sample data set and determine the target sample data set, wherein the generative adversarial network based on the improved Cartesian coordinate system includes: a generator, a discriminator and a coordinate mapper. In one embodiment, data collection, labeling and preprocessing are time-consuming and labor-intensive, especially in the case of small samples. Insufficient data volume can easily lead to poor generalization and low precision of the model. Therefore, after feature extraction is completed, the present invention proposes a sample generation based on the algorithm principle of the generative adversarial network under the improved Cartesian coordinate system, thereby realizing sample expansion: the generative adversarial network in the generative adversarial network based on the improved Cartesian coordinate system consists of two parts: a generator (G) and a discriminator (D). The generator is used to create data from random noise, and the discriminator is used to try to distinguish data from a real data set from data created by the generator.

[0040] In the generative adversarial network based on the improved Cartesian coordinate system of the present invention, the idea of ​​the Cartesian coordinate system is integrated, that is, the multi-dimensional data space is used to capture the complexity and intrinsic structure of the data. In addition, the present invention no longer generates data only from random noise, but uses points in the Cartesian coordinate system as seeds to enhance the quality of the data. The traditional generative adversarial network may miss some important data distributions, so the present invention introduces an improved anchor point strategy. The core idea of ​​this strategy is to determine the key points in the original data space (called "anchor points"), which capture the main features of the data, and then the generator creates new data points with these anchor points as the center. In addition, in addition to the generator in the traditional generative adversarial network and the discriminator In the generative adversarial network based on the improved Cartesian coordinate system of the present invention, a third component is introduced: a coordinate mapper (C).

[0041] Step S106, based on the target sample data set, the preset electric power new energy equipment fault diagnosis model is trained and processed to determine the target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power new energy equipment fault data to be classified, and determine the fault classification result. In one embodiment, the expanded data is input into the neural network for feature extraction: the traditional neural network adopts the gradient descent method to optimize the parameters. The present invention proposes a neural network optimization algorithm based on the droplet impact and splashing mechanism optimization strategy to avoid the gradient vanishing and gradient explosion phenomena caused by the traditional gradient descent method, and at the same time avoid the parameters from falling into the local optimal solution when searching for the best solution; inspired by the droplet impact and splashing mechanism in nature, when raindrops fall on the water surface, ripples and droplet splashing will be formed due to the interaction between liquids and liquids. This process involves a variety of physical phenomena, such as conservation of momentum, conservation of energy, etc. The present invention applies this droplet impact and splashing mechanism to the optimization of neural network parameters, and finds the best value for the network parameter by simulating the droplet impact process.

[0042] In another embodiment, the present invention proposes a classification based on an improved low-rank approximate extreme learning machine algorithm. The extreme learning machine is a single hidden layer feedforward neural network. Its key advantage is that it can directly solve the output weight without iteration. The present invention uses a low-rank structure to approximate the weights between the hidden layer and the output layer. In addition, traditional neural networks use integer-order derivatives, while the present invention uses fractional-order derivatives, which can better capture the nonlinear relationship between data in small sample scenarios.

[0043] The above-mentioned small sample-based new energy electric power equipment fault diagnosis method provided in an embodiment of the present invention performs data expansion processing on the original sample data set based on a generative adversarial network in an improved Cartesian coordinate system, which can effectively enhance the quantity and diversity of small sample data, thereby improving the generalization ability and accuracy of the model.

[0044] See also Figure 2 The flowchart of a model training method shown in FIG. 1 is a flowchart of a model training method. The embodiment of the present invention also provides an implementation method of the model training method. For details, refer to the following steps S202 to S206:

[0045] Step S202, based on the principle of the generative adversarial network algorithm in the improved Cartesian coordinate system, the sample data is expanded. The original sample data set is mapped to a high-dimensional space by a coordinate mapper, and the original sample data set in the high-dimensional space is expanded by a generator to determine the first expanded data, and the first expanded data is inversely mapped by the coordinate mapper to map the first expanded data back to the original dimension to determine the second expanded data. Finally, the second expanded data is identified by a discriminator. The steps of determining the target sample data set include the following (1) to (2):

[0046] (1) Determine the anchor point set corresponding to the original sample data set based on the entropy weighted algorithm. Through the generator, based on the coordinate transformation rules of the coordinates in the Cartesian coordinate system in the high-dimensional space, with each anchor point as the center, perform data expansion processing on the original sample data to determine the first expanded data, wherein the anchor point set includes anchor points corresponding to the main features of the data. Specifically, the following steps are required for data expansion processing: initialization, high-dimensional mapping, high-dimensional anchor point selection, seed generation and transformation in the Cartesian coordinate system, high-dimensional data generation, data inverse mapping, data identification, joint optimization of the generator, discriminator and mapper, and anchor point update strategy.

[0047] First, initialize the discriminator with the original data set , and train a basic generator , and then through high-dimensional mapping, using the coordinate mapper , maps the original data to a high-dimensional space. Define the coordinate mapper As a function, it can be expressed as:

[0048]

[0049] in, is the original data vector; is the mapping matrix.

[0050] Furthermore, the mapping matrix The role of is to keep some characteristics of the original data while transferring it to a high-dimensional space. The present invention defines is a diagonal matrix whose diagonal elements are the predicates of the principal component analysis (PCA) of the data. principal components, among which is the dimension of the data, and the formula is:

[0051]

[0052] in, It is the original data No. principal components; is the mapping matrix No. diagonal elements.

[0053] When performing the step of high-dimensional anchor point selection, it is necessary to select anchor points that capture the main features of the data in the high-dimensional space. . Assume that the present invention has anchor points, then:

[0054]

[0055] in, is the set of anchor points selected in the high-dimensional space; For the Anchor points.

[0056] Furthermore, when selecting anchor points, the present invention uses K-means clustering to select anchor points in a high-dimensional space, which can be expressed as:

[0057]

[0058] in, It is the first The center of a cluster.

[0059] When performing the steps of seed generation and transformation in the Cartesian coordinate system, it is necessary to randomly select seed points in the high-dimensional space based on the coordinates in the Cartesian coordinate system and use the coordinate mapper For any seed point in high-dimensional space , which translates to:

[0060]

[0061] in, is a randomly selected seed point in high-dimensional space; To use the coordinate mapper The converted seed point.

[0062] When performing the step of generating high-dimensional data, it is necessary to use the generator and anchors to generate data in high-dimensional space. Using the generator , then the generation method of high-dimensional data can be expressed as:

[0063]

[0064] in, is a scaling factor that controls the contribution of the anchor point. is the average position of all anchor points.

[0065] Furthermore, the generator in the present invention It consists of three layers of fully connected neural network with activation function , the formula is:

[0066]

[0067]

[0068]

[0069] in, and The generator Layer weights and biases; For the generator In a given seed and anchor point data generated in the case of is the ReLU activation function.

[0070] When performing the data inverse mapping step, you need to use a coordinate mapper The inverse mapping of , maps the high-dimensional data back to the original dimension. The inverse mapping formula is:

[0071]

[0072] in, is the mapping matrix The inverse of For the generator; is the discriminator; and are the coordinate mappers and their inverses.

[0073] Furthermore, the coordinate mapper The loss function is based on the reconstruction error, that is, the error between the original data and the data mapped back by the mapper. The formula is:

[0074]

[0075] In the step of data identification, a discriminator is needed Identification by Generated data points in the original dimension. Discriminator The goal is to minimize the The difference between the data and the real data can be expressed as:

[0076]

[0077] in, and are the loss functions representing the discriminator and coordinate mapper respectively.

[0078] Furthermore, the discriminator in the present invention It also consists of three layers of fully connected neural networks, but its last layer uses the Sigmoid activation function. The formula is:

[0079]

[0080]

[0081]

[0082] When performing the joint optimization step of the generator, discriminator, and mapper, it is necessary to Feedback, joint optimization generator , Discriminator and coordinate mapper In the process of iterative optimization, the total loss can be expressed as:

[0083]

[0084] in, It is a coordinate mapper The loss represents the accuracy of the mapping. is the weight, control importance in the total loss. By using the gradient descent method, we update the generator , discriminator , and coordinate mapper The weight of .

[0085] When performing the anchor point update strategy, it is necessary to regularly update the anchor point based on the newly generated data. , anchor point The update method can be expressed as:

[0086]

[0087] in, is a weight factor between 0 and 1, is the entropy weight of the i-th anchor point, is the learning rate of the current iteration.

[0088] (2) Determine a set of candidate anchor points in a high-dimensional space through a clustering algorithm, and perform entropy weighting on the candidate anchor point set based on the entropy of the category to which each data point in the original sample data set belongs through an entropy weighting algorithm to determine the entropy weight corresponding to each candidate anchor point in the candidate anchor point set. According to the entropy weight corresponding to each candidate anchor point and a preset weight threshold, determine the anchor point set corresponding to the original sample data set. Specifically, in order to ensure that the selected anchor point can more representatively capture the characteristics of the data, the present invention introduces an entropy weighting strategy to enhance the selection of anchor points. Entropy, as an information theory measurement tool, can provide information about the uncertainty of data distribution. By using the entropy weighting strategy, it can be ensured that the selected anchor point can capture the diversity of the data. Specifically, the entropy of the category to which each data point belongs is calculated. For the data point and its label distribution ,entropy Defined as:

[0089]

[0090] in, is the number of categories, It is a data point Belong to The probability of the class.

[0091] Furthermore, entropy-weighted anchor point selection is performed. Each candidate anchor point in the candidate anchor point set selected using the above K-means method will get an entropy weight , the calculation formula is:

[0092]

[0093] in, is the number of candidate anchor points.

[0094] Furthermore, the learning rate In order to enhance the training speed and accuracy of the model, different learning rates may be required at different stages of the learning process. When the loss decreases slowly, the learning rate may need to be increased to skip the local minimum, while when the loss decreases rapidly, the learning rate may need to be reduced to ensure model convergence. That is, when calculating the dynamic learning rate, first, define a base learning rate and a learning rate adjustment factor , whose initial value is 1. In each training cycle, the change in loss is calculated:

[0095]

[0096] in, and are the losses of the current and previous training cycles, respectively.

[0097] Furthermore, if Less than a predetermined small threshold , increasing the learning rate, can be expressed as:

[0098]

[0099] if Greater than a predetermined large value threshold , reducing the learning rate, can be expressed as:

[0100]

[0101] in, is a small adjustment value.

[0102] Furthermore, when updating the learning rate, using the above adjustment factor, the way to calculate the dynamic learning rate can be expressed as:

[0103]

[0104] Repeat the above steps of seed generation and transformation to anchor point update strategy in Cartesian coordinate system until a specific convergence criterion is met or a predetermined number of iterations is reached, then perform data fusion to fuse the generated data with the original data set to form an extended data set.

[0105] Step S204, input the expanded data into the neural network for feature extraction. According to the neural network optimization algorithm of the droplet impact and splash mechanism, feature extraction processing is performed on the target sample data set to determine the target feature set: the parameter set in the preset initialization neural network model is updated through the neural network optimization algorithm of the droplet impact and splash mechanism, the target neural network model is determined, and the sample data is subjected to feature extraction processing using the target neural network model to determine the target feature set. Specifically, the feature extraction steps include: initialization, multi-scale selection, evaluation and energy conversion, impact and splash, parameter update, adaptive adjustment, convergence detection and annealing strategy and other steps for feature extraction processing.

[0106] During the initialization and multi-scale selection steps, it is necessary to assign random values ​​to the parameters of the neural network, randomly set an initial droplet impact force and direction, and select a suitable scale to simulate the droplet based on the current optimization state.

[0107] When performing the evaluation and energy conversion steps, the error of the neural network needs to be calculated and converted into the energy of the droplet. , which is mapped to the energy of the droplet , further, in order to enhance the nonlinear characteristics of this relationship, the present invention introduces an exponential function to represent the relationship between error and energy:

[0108]

[0109] in, is a real number greater than 1, which enhances the nonlinear relationship between error and energy. is the output error of the neural network; is the energy representing the droplet; is a constant used to control the mapping relationship between error and energy; is a real number greater than 1, which enhances the nonlinear relationship between error and energy.

[0110] When performing the impact and splashing steps, it is necessary to simulate the impact of the droplet at the selected scale. Further, based on the energy and momentum of the droplet, the direction and distance of the splashing droplet are calculated. The impact of the droplet can be calculated by momentum. To describe, it is expressed as:

[0111]

[0112] in, is the initial velocity of the droplet; is the momentum of the droplet.

[0113] Furthermore, the direction of the splashing droplets generated by the impact of the droplets and distance It can be obtained by the following formula:

[0114]

[0115]

[0116] in, is a constant used to control the splash distance; is the direction of the splashing droplets; is the distance of the splashing droplets.

[0117] Furthermore, chain splash is simulated and the momentum of splashing droplets is taken into account. Number of droplets generated by chain splash With its momentum It can be expressed as:

[0118]

[0119]

[0120] in, is a constant involving chain splashes, is the weight of the ith droplet. The number of droplets generated for chain splash; is a constant involving chain splash; is the momentum of the chain splash droplets.

[0121] Furthermore, considering that the size of the droplet may affect its momentum, a weight is introduced for each droplet , which can be expressed as:

[0122]

[0123] in, It is The energy of a droplet.

[0124] When performing the parameter update step, the new position of the simulated droplet needs to be mapped to the parameters of the neural network and updated. The movement of the simulated droplet can be expressed as:

[0125]

[0126]

[0127] in, is the Gaussian term affected by the splash; , are the weight values ​​before and after the update; , is the bias value before and after the update; is the learning rate.

[0128] Furthermore, in order to consider the impact area of ​​droplet splashing, the splashing affects the Gaussian term Defined as:

[0129]

[0130] in, Controls the range of the splash.

[0131] When performing the adaptive adjustment step, it is necessary to analyze the current optimization state and adjust the impact force and direction of the droplets. Furthermore, the droplet size for the next iteration is determined based on the historical optimization records. and direction Adjustments can be made based on historical optimization records, which can be expressed as:

[0132]

[0133] in, It is the impact strength of the previous iteration; is the impact force of the droplet.

[0134] Furthermore, the direction The adjustment takes into account the delay effect and can be expressed as:

[0135]

[0136] in, is the change in energy, is the number of iterations, is the direction of the previous iteration; is the impact direction of the droplet; is the change in energy; is the number of iterations; is a delay factor between 0 and 1.

[0137] When performing the convergence detection and annealing strategy steps, it is necessary to check whether the parameters are stable. If the expected accuracy is not achieved, the energy of the droplet is adjusted according to the annealing strategy to reduce the impact force of the droplet and return to step 2. The annealing strategy can be adjusted by adjusting the energy To achieve this, it can be expressed as:

[0138]

[0139] in, is a constant less than 1 in the annealing strategy, used to reduce the energy.

[0140] During the save and end steps, it is necessary to save the current optimal parameters and end the algorithm when the end conditions are met. This complex DOA process ensures the careful optimization of parameters at various scales, and through adaptive mechanisms, chain splashing and multi-scale simulation, the parameter update is more comprehensive and in-depth.

[0141] Step S206, training a classifier based on an improved low-rank approximate extreme learning machine algorithm. Input the target sample data set and the target feature set into a preset classifier, and perform classification training on the classifier based on the improved low-rank approximate extreme learning machine algorithm to determine a fault classification model: map the error of the neural network to the energy of the droplet, and determine the droplet update position and the neural network parameters corresponding to the droplet update position by simulating the droplet splash information generated by the droplet impact, perform adaptive adjustment processing and convergence detection processing on the neural network parameters corresponding to the droplet update position, determine the target neural network parameters, and determine the target neural network model based on the target neural network parameters, wherein the droplet splash information includes: the direction of the scattered droplets and the distance of the scattered droplets. Specifically, the steps of training the classifier include: initialization and weight setting, data mapping to the hidden layer, dynamic low-rank approximation weight solution, fractional-order backpropagation and gradient correction, adaptive batch selection, online learning and data enhancement, termination conditions and model verification and integration strategy, etc. to train the classifier.

[0142] When performing the initialization and weight setting steps, it is necessary to randomly select the weights and biases of the hidden layer, and use regularization techniques (such as L2 regularization) to adjust the initialization weights to prevent overfitting. Specifically, the hidden layer weights and biases are randomly initialized as:

[0143]

[0144]

[0145] in, is the input dimension, is the number of neurons in the hidden layer. Further, the weights are adjusted using L2 regularization, which can be expressed as:

[0146]

[0147] in, is the regularization coefficient.

[0148] When performing the step of mapping data to the hidden layer, the input data needs to be mapped to the hidden layer. Use a variety of activation functions (such as Sigmoid, tanh, ReLU, etc.) for nonlinear transformation, and assign a coefficient to the result of each activation function for subsequent weighted averaging. Specifically, the data is mapped to the output of the hidden layer The calculation method can be expressed as:

[0149]

[0150] in, is the Sigmoid function, is the weight coefficient of each activation function, and . is the weight matrix of the hidden layer; is the bias vector of the hidden layer; is the hidden layer weight matrix after L2 regularization; is the output of the hidden layer;

[0151] Furthermore, the present invention considers three activation functions: Sigmoid, tanh, and ReLU, which are respectively expressed as:

[0152] Sigmoid function:

[0153]

[0154] Tanh function:

[0155]

[0156] ReLU function:

[0157]

[0158] When performing the step of solving the weight of dynamic low-rank approximation, it is necessary to use matrix decomposition to find multiple low-rank approximation representations of the weights between the hidden layer and the output layer, and dynamically select the optimal low-rank approximation strategy based on the performance of the validation set. Specifically, assuming that the weight from the hidden layer to the output layer is , using SVD for low-rank approximation, we have:

[0159]

[0160] in, is the result of SVD decomposition. The present invention only selects the former The columns corresponding to the singular values ​​can be approximated as follows:

[0161]

[0162] in, is the weight matrix from the hidden layer to the output layer; are the left singular vector matrix, singular value diagonal matrix and right singular vector matrix after SVD decomposition respectively; is the number of singular values ​​selected for approximation; is the weight matrix after low-rank approximation.

[0163] When performing fractional-order backpropagation and gradient correction steps, fractional-order derivatives are needed for backpropagation, and the loss function About weight The fractional gradient of is:

[0164]

[0165] in, represents a 1.5-order differential operation; is the loss function; is the 1.5th order gradient of the loss function with respect to the weight; It is a 1.5-order differential operation.

[0166] Furthermore, fractional-order backpropagation and gradient correction are performed, and the way to update the weights can be expressed as:

[0167]

[0168] in, is the learning rate; is the updated hidden layer weight matrix.

[0169] When performing the adaptive batch selection step, it is necessary to dynamically adjust the batch size of each epoch based on the performance of the model in the first three epochs. Specifically, the adjustment method is:

[0170] Given an initial batch size , after each epoch, calculate the loss of the current epoch And the loss of the previous epoch .

[0171] if (in is a small threshold), then ,otherwise .

[0172] in, are the initial and new batch sizes, respectively; The losses of the current and previous epochs respectively; is the small threshold used for the adaptive batch selection strategy.

[0173] When performing online learning and data enhancement, it is necessary to randomly select a portion of the training data after each epoch, perform a slight transformation (such as rotation, scaling, or adding noise) and add it to the training set to improve the generalization ability of the model. The new data obtained after the transformation of part of the data can be expressed as:

[0174]

[0175] in, Is a transformation function, the specific data transformation function It can be expressed as a combinatorial function, such as:

[0176]

[0177] in, is the function that adds noise, is the scaling function, is the rotation function; For input data; The input data after data enhancement; is the data transformation function.

[0178] When performing the termination condition and model validation steps, it is necessary to introduce an early stopping mechanism. When the performance on the validation set does not improve in multiple consecutive epochs, stop training. Furthermore, use K-fold cross validation to evaluate the model and make necessary hyperparameter adjustments. Specifically, use the early stopping mechanism. If the performance on the validation set does not improve in multiple consecutive epochs, stop training. If the verification loss does not decrease in epochs, the training is stopped, which can be expressed as:

[0179]

[0180] in, is the validation set loss.

[0181] When performing the integration strategy, it is necessary to train multiple models based on the improved low-rank approximation extreme learning machine, each model using different initialization and parameters. Specifically, the output weight set of multiple models is obtained by training ,The final classification decision is the weighted average of these models, and finally, the prediction results of all models are combined through a voting mechanism or an averaging strategy.

[0182] In summary, the present invention can combine the Cartesian coordinate system with the traditional generative adversarial network, which not only enhances the quality and diversity of data, but also more effectively captures the complexity and intrinsic structure of data. Through the generative adversarial network, the quantity and diversity of small sample data are effectively enhanced, thereby improving the generalization ability and accuracy of the model; the neural network optimization strategy of the droplet impact and splash mechanism simulates the droplet impact process, provides an innovative method for neural network parameter optimization, and effectively avoids the problems in the traditional gradient descent method. The droplet impact and splash mechanism provides a more comprehensive and in-depth parameter update method for the neural network, thereby better Optimize model performance; based on the improved low-rank approximate extreme learning machine algorithm: combining low-rank structure and fractional order derivative, it not only improves computational efficiency, but also better captures the nonlinear relationship of data, especially in small sample scenarios; dynamic anchor point update strategy and entropy weighting strategy can ensure that the selected anchor point is more representative and better captures the characteristics of the data; multi-scale simulation and adaptive mechanism can be in the droplet impact and splash mechanism optimization strategy, these mechanisms ensure the meticulous optimization of parameters, through the droplet impact and splash mechanism optimization strategy and based on the improved low-rank approximate extreme learning machine algorithm, the model has stronger generalization ability and higher classification accuracy. Therefore, the above-mentioned small sample-based new energy power equipment fault diagnosis method provided by the embodiment of the present invention can realize efficient data expansion based on the improved generative adversarial network in the Cartesian coordinate system, realize optimization parameter adjustment based on the neural network optimization strategy of the droplet impact and splash mechanism, and realize fast and efficient classification and better model generalization based on the improved low-rank approximate extreme learning machine algorithm.

[0183] For the new energy equipment fault diagnosis method based on a small sample provided in the above embodiment, the embodiment of the present invention provides a new energy equipment fault diagnosis device based on a small sample, see Figure 3 The schematic diagram of the structure of a new energy equipment fault diagnosis device based on a small sample is shown, and the device includes the following parts:

[0184] The sample data collection module 302 obtains the original sample data set of the new energy power equipment, wherein the original sample data set includes: a sensor data set and a state label set corresponding to each sensor data;

[0185] The data expansion module 304 performs data expansion processing on the original sample data set based on the generative adversarial network in the improved Cartesian coordinate system to expand the original sample data set and determine the target sample data set;

[0186] The model training module 306 performs model training processing on the preset new energy electric power equipment fault diagnosis model based on the target sample data set to determine the target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the new energy electric power equipment fault data to be classified to determine the fault classification result.

[0187] The above-mentioned small-sample-based electric power new energy equipment fault diagnosis device provided in the embodiment of the present application performs data expansion processing on the original sample data set based on the improved generative adversarial network in the Cartesian coordinate system, which can effectively enhance the quantity and diversity of small sample data, thereby improving the generalization ability and accuracy of the model.

[0188] In one embodiment, a generative adversarial network based on an improved Cartesian coordinate system includes: a generator, a discriminator and a coordinate mapper. When performing a generative adversarial network based on an improved Cartesian coordinate system, performing data expansion processing on the original sample data set to expand the original sample data set and determine the target sample data set, the above-mentioned data expansion module 304 is also used to: map the original sample data set to a high-dimensional space through the coordinate mapper, and use the generator to perform data expansion processing on the original sample data set in the high-dimensional space to determine the first expanded data; perform inverse mapping processing on the first expanded data through the coordinate mapper to map the first expanded data back to the original dimension to determine the second expanded data; perform data identification processing on the second expanded data through the discriminator to determine the target sample data set.

[0189] In one embodiment, when using a generator to perform data expansion processing on the original sample data set in the high-dimensional space to determine the first expanded data, the above-mentioned data expansion module 304 is also used to: determine the anchor point set corresponding to the original sample data set based on an entropy weighted algorithm, wherein the anchor point set includes anchor points corresponding to the main features of the data; through the generator, based on the coordinate conversion rules of Cartesian coordinates in the high-dimensional space, with each anchor point as the center, perform data expansion processing on the original sample data to determine the first expanded data.

[0190] In one embodiment, when performing the step of determining the anchor point set corresponding to the original sample data set based on the entropy weighted algorithm, the above-mentioned data expansion module 304 is also used to: determine the candidate anchor point set in the high-dimensional space through the clustering algorithm; perform entropy weighting processing on the candidate anchor point set based on the entropy of the category to which each data point in the original sample data set belongs through the entropy weighted algorithm, and determine the entropy weight corresponding to each candidate anchor point in the candidate anchor point set; determine the anchor point set corresponding to the original sample data set based on the entropy weight corresponding to each candidate anchor point and a preset weight threshold.

[0191] In one embodiment, when performing model training processing on a preset electric power new energy equipment fault diagnosis model based on a target sample data set to determine the step of the target fault diagnosis model, the above-mentioned model training module 306 is also used to: perform feature extraction processing on the target sample data set according to a neural network optimization algorithm of the droplet impact and splashing mechanism to determine the target feature set; input the target sample data set and the target feature set into a preset classifier, perform classification training on the classifier based on an improved low-rank approximate extreme learning machine algorithm to determine the fault classification model.

[0192] In one embodiment, when performing feature extraction processing on a target sample data set and determining a target feature set according to a neural network optimization algorithm for droplet impact and splashing mechanism, the above-mentioned model training module 306 is also used to: perform parameter update processing on a parameter set in a preset initialized neural network model and determine a target neural network model through a neural network optimization algorithm for droplet impact and splashing mechanism; perform feature extraction processing on sample data using the target neural network model and determine a target feature set.

[0193] In one embodiment, when performing a neural network optimization algorithm through droplet impact and splashing mechanism, performing parameter update processing on the parameter set in the preset initialized neural network model, and determining the step of the target neural network model, the above-mentioned model training module 306 is also used to: map the error of the neural network to the energy of the droplet, and determine the droplet update position and the neural network parameters corresponding to the droplet update position by simulating the droplet splashing information generated by the droplet impact, wherein the droplet splashing information includes: the splashing droplet direction and the splashing droplet distance; adaptively adjust the neural network parameters corresponding to the droplet update position and convergence detection processing to determine the target neural network parameters, and determine the target neural network model based on the target neural network parameters.

[0194] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0195] An embodiment of the present invention provides a server. Specifically, the server includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above-mentioned embodiments.

[0196] Figure 4A structural diagram of a server provided in an embodiment of the present invention, the server 100 includes: a processor 40, a memory 41, a bus 42 and a communication interface 43, wherein the processor 40, the communication interface 44 and the memory 41 are connected via the bus 42; the processor 40 is used to execute an executable module stored in the memory 41, such as a computer program.

[0197] The memory 41 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 43 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0198] The bus 42 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0199] Among them, the memory 41 is used to store programs, and the processor 40 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor 40 or implemented by the processor 40.

[0200] The processor 40 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 40. The above processor 40 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present invention can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 41, and the processor 40 reads the information in the memory 41 and completes the steps of the above method in combination with its hardware.

[0201] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0202] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0203] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for fault diagnosis of new energy power equipment based on small samples, characterized in that: The method comprises: Acquire an original sample data set of the new energy power equipment, wherein the original sample data set includes: a sensor data set and a state label set corresponding to each sensor data, wherein the sensor data set is operation data collected by sensors in the new energy power equipment and generated during operation of the new energy power equipment; Based on the improved generative adversarial network under the Cartesian coordinate system, the original sample data set is subjected to data expansion processing to expand the original sample data set and determine the target sample data set, wherein the generative adversarial network based on the improved Cartesian coordinate system is a generative adversarial network that integrates the Cartesian coordinate system concept and uses the multidimensional data space under the Cartesian coordinate system to capture the complexity and inherent structure of data. The generative adversarial network based on the improved Cartesian coordinate system is used to use points in the Cartesian coordinate system as seeds and perform seed generation and transformation processing under the Cartesian coordinate system to enhance the quality of data; Based on the target sample data set, a preset electric power new energy equipment fault diagnosis model is trained to determine a target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power new energy equipment fault data to be classified to determine a fault classification result; Wherein, the generative adversarial network based on the improved Cartesian coordinate system includes: a generator, a discriminator and a coordinate mapper. The generative adversarial network based on the improved Cartesian coordinate system performs data expansion processing on the original sample data set to expand the original sample data set and determine the step of the target sample data set, including: mapping the original sample data set to a high-dimensional space by the coordinate mapper, and using the generator to perform data expansion processing on the original sample data set in the high-dimensional space to determine the first expanded data; performing inverse mapping processing on the first expanded data by the coordinate mapper to map the first expanded data back to the original dimension to determine the second expanded data; performing data identification processing on the second expanded data by the discriminator to determine the target sample data set; The step of using the generator to perform data enlargement processing on the original sample data set in the high-dimensional space to determine the first extended data includes: determining an anchor point set corresponding to the original sample data set based on an entropy weighted algorithm, wherein the anchor point set includes anchor points corresponding to the main features of the data; performing data enlargement processing on the original sample data based on the coordinate conversion rule of Cartesian coordinates in the high-dimensional space and taking each anchor point as the center to determine the first extended data by using the generator; Wherein, the step of determining the anchor point set corresponding to the original sample data set based on the entropy weighting algorithm comprises: determining a candidate anchor point set in a high-dimensional space by a clustering algorithm; performing entropy weighting processing on the candidate anchor point set based on the entropy of the category to which each data point in the original sample data set belongs by an entropy weighting algorithm to determine the entropy weight corresponding to each candidate anchor point in the candidate anchor point set; determining the anchor point set corresponding to the original sample data set according to the entropy weight corresponding to each candidate anchor point and a preset weight threshold; The step of performing model training processing on a preset electric power new energy equipment fault diagnosis model based on the target sample data set to determine the target fault diagnosis model includes: According to the neural network optimization algorithm of droplet impact and splatter mechanism, feature extraction processing is performed on the target sample data set to determine the target feature set, wherein the neural network optimization algorithm according to the droplet impact and splatter mechanism is used to find the optimal value for the neural network parameter by simulating the droplet impact process according to the droplet impact and splatter mechanism in nature; Inputting the target sample data set and the target feature set into a preset classifier, performing classification training on the classifier based on an improved low-rank approximate extreme learning machine algorithm, and determining a fault classification model; Among them, when the target sample data set is subjected to feature extraction processing according to the neural network optimization algorithm of the droplet impact and splash mechanism and the target feature set is determined, initialization, multi-scale selection, evaluation and energy conversion, impact and splash, parameter update, adaptive adjustment and convergence detection processing need to be performed in sequence; Among them, when performing the initialization and multi-scale selection steps, it is necessary to assign random values ​​to the parameters of the neural network, and randomly set an initial droplet impact force and direction, and select the target scale to simulate the droplet according to the current optimization state; Among them, when performing the evaluation and energy conversion steps, it is necessary to calculate the error of the neural network and convert it into the energy of the droplet. , which is mapped to the energy of the droplet , and introduce an exponential function to represent the relationship between error and energy to enhance the nonlinear characteristics of this relationship: in, is a real number greater than 1, enhancing the nonlinear relationship between error and energy. is the output error of the neural network; is the energy representing the droplet; is a constant used to control the mapping relationship between error and energy; In the steps of impact and splash, it is necessary to simulate the impact of the droplet at the selected scale, and calculate the direction and distance of the splashing droplet according to the energy and momentum of the droplet. To describe, it is expressed as: in, is the initial velocity of the droplet; is the momentum of the droplet; The direction of the splashing droplets generated by the impact of the droplets is and distance It can be obtained by the following formula: in, is a constant used to control the splash distance; is the direction of the splashing droplets; is the distance of the splashing droplets; Among them, when simulating chain splash and considering the momentum of splashing droplets, the number of droplets generated by chain splash is With its momentum It can be expressed as: in, is a constant involving chain splashes, is the weight of the ith droplet, The number of droplets generated for chain splash; ; Among them, a weight is introduced for each droplet , which describes the image of the droplet size on its momentum, can be expressed as: in, It is The energy of a droplet; In the step of updating parameters, the new position of the simulated droplet is corresponded to the parameters of the neural network and updated. The movement of the simulated droplet can be expressed as: in, is the Gaussian term affected by the splash; , are the weight values ​​before and after the update; , is the bias value before and after the update; is the learning rate; Among them, the splashing influence Gaussian term is combined with the impact area of ​​droplet splashing Defined as: in, Control the impact range of splash; In the adaptive adjustment step, it is necessary to analyze the current optimization state, adjust the impact force and direction of the droplets, and determine the droplet size for the next iteration based on the historical optimization records. and direction Adjustments can be made based on historical optimization records, which can be expressed as: in, It is the impact strength of the previous iteration; is the impact force of the droplet; Furthermore, the direction The adjustment takes into account the delay effect and can be expressed as: in, is the change in energy, is the number of iterations, is the direction of the previous iteration; is the impact direction of the droplet; is the change in energy; is the number of iterations; is a delay factor between 0 and 1; When performing the convergence detection and annealing strategy steps, it is necessary to check whether the parameters tend to be stable; if the expected accuracy is not achieved, the energy of the droplet is adjusted according to the annealing strategy to reduce the impact force of the droplet. If the expected accuracy is achieved, it is determined that the end condition is met, the current optimal parameters are saved, and the algorithm is terminated to use the above algorithm for feature extraction processing to determine the target feature set.

2. The method for fault diagnosis of new energy power equipment based on small samples according to claim 1 is characterized in that: The step of performing feature extraction processing on the target sample data set based on the neural network optimization algorithm of the droplet impact and splash mechanism to determine the target feature set includes: Through the neural network optimization algorithm of droplet impact and splashing mechanism, the parameter set in the preset initialization neural network model is updated to determine the target neural network model; The target neural network model is used to perform feature extraction processing on the sample data to determine the target feature set.

3. The method for fault diagnosis of new energy power equipment based on small samples according to claim 1 is characterized in that: The steps of performing parameter updating processing on the parameter set in the preset initialization neural network model by using the neural network optimization algorithm of the droplet impact and splashing mechanism, and determining the target neural network model include: The error of the neural network is mapped to the energy of the droplet, and the droplet update position and the neural network parameters corresponding to the droplet update position are determined by simulating the droplet splash information generated by the droplet impact, wherein the droplet splash information includes: the splash drop direction and the splash drop distance; Adaptive adjustment processing and convergence detection processing are performed on the neural network parameters corresponding to the droplet update position to determine the target neural network parameters, and the target neural network model is determined based on the target neural network parameters.

4. A fault diagnosis device for electric power new energy equipment based on small samples, characterized in that: The device comprises: A sample data collection module is used to obtain an original sample data set of the new energy power equipment, wherein the original sample data set includes: a sensor data set and a state label set corresponding to each sensor data, wherein the sensor data set is the operation data collected by the sensor in the new energy power equipment and generated during the operation of the new energy power equipment; A data expansion module, based on a generative adversarial network under an improved Cartesian coordinate system, performs data expansion processing on the original sample data set to expand the original sample data set and determine a target sample data set, wherein the generative adversarial network based on the improved Cartesian coordinate system is an adversarial network that integrates the Cartesian coordinate system concept and uses the multidimensional data space under the Cartesian coordinate system to capture the complexity and inherent structure of data. The generative adversarial network based on the improved Cartesian coordinate system is used to use points in the Cartesian coordinate system as seeds and perform seed generation and transformation processing under the Cartesian coordinate system to enhance data quality; A model training module, based on the target sample data set, performs model training processing on a preset electric power new energy equipment fault diagnosis model to determine a target fault diagnosis model, wherein the target fault diagnosis model is used to perform fault classification processing on the electric power new energy equipment fault data to be classified to determine a fault classification result; Wherein, the generative adversarial network based on the improved Cartesian coordinate system includes: a generator, a discriminator and a coordinate mapper. The generative adversarial network based on the improved Cartesian coordinate system performs data expansion processing on the original sample data set to expand the original sample data set and determine the step of the target sample data set, including: mapping the original sample data set to a high-dimensional space by the coordinate mapper, and using the generator to perform data expansion processing on the original sample data set in the high-dimensional space to determine the first expanded data; performing inverse mapping processing on the first expanded data by the coordinate mapper to map the first expanded data back to the original dimension to determine the second expanded data; performing data identification processing on the second expanded data by the discriminator to determine the target sample data set; The step of using the generator to perform data enlargement processing on the original sample data set in the high-dimensional space to determine the first extended data includes: determining an anchor point set corresponding to the original sample data set based on an entropy weighted algorithm, wherein the anchor point set includes anchor points corresponding to the main features of the data; performing data enlargement processing on the original sample data based on the coordinate conversion rule of Cartesian coordinates in the high-dimensional space and taking each anchor point as the center to determine the first extended data by using the generator; Wherein, the step of determining the anchor point set corresponding to the original sample data set based on the entropy weighting algorithm comprises: determining a candidate anchor point set in a high-dimensional space by a clustering algorithm; performing entropy weighting processing on the candidate anchor point set based on the entropy of the category to which each data point in the original sample data set belongs by an entropy weighting algorithm to determine the entropy weight corresponding to each candidate anchor point in the candidate anchor point set; determining the anchor point set corresponding to the original sample data set according to the entropy weight corresponding to each candidate anchor point and a preset weight threshold; The step of performing model training processing on a preset electric power new energy equipment fault diagnosis model based on the target sample data set to determine the target fault diagnosis model includes: According to the neural network optimization algorithm of droplet impact and splatter mechanism, feature extraction processing is performed on the target sample data set to determine the target feature set, wherein the neural network optimization algorithm according to the droplet impact and splatter mechanism is used to find the optimal value for the neural network parameter by simulating the droplet impact process according to the droplet impact and splatter mechanism in nature; Inputting the target sample data set and the target feature set into a preset classifier, performing classification training on the classifier based on an improved low-rank approximate extreme learning machine algorithm, and determining a fault classification model; Among them, when the target sample data set is subjected to feature extraction processing according to the neural network optimization algorithm of the droplet impact and splash mechanism and the target feature set is determined, initialization, multi-scale selection, evaluation and energy conversion, impact and splash, parameter update, adaptive adjustment and convergence detection processing need to be performed in sequence; Among them, when performing the initialization and multi-scale selection steps, it is necessary to assign random values ​​to the parameters of the neural network, and randomly set an initial droplet impact force and direction, and select the target scale to simulate the droplet according to the current optimization state; Among them, when performing the evaluation and energy conversion steps, it is necessary to calculate the error of the neural network and convert it into the energy of the droplet. , which is mapped to the energy of the droplet , and introduce an exponential function to represent the relationship between error and energy to enhance the nonlinear characteristics of this relationship: in, is a real number greater than 1, enhancing the nonlinear relationship between error and energy. is the output error of the neural network; is the energy representing the droplet; is a constant used to control the mapping relationship between error and energy; In the steps of impact and splash, it is necessary to simulate the impact of the droplet at the selected scale, and calculate the direction and distance of the splashing droplet according to the energy and momentum of the droplet. To describe, it is expressed as: in, is the initial velocity of the droplet; is the momentum of the droplet; The direction of the splashing droplets generated by the impact of the droplets is and distance It can be obtained by the following formula: in, is a constant used to control the splash distance; is the direction of the splashing droplets; is the distance of the splashing droplets; Among them, when simulating chain splash and considering the momentum of splashing droplets, the number of droplets generated by chain splash is With its momentum It can be expressed as: in, is a constant involving chain splashes, is the weight of the ith droplet, The number of droplets generated for chain splash; ; Among them, a weight is introduced for each droplet , which describes the image of the droplet size on its momentum, can be expressed as: in, It is The energy of a droplet; In the step of updating parameters, the new position of the simulated droplet is corresponded to the parameters of the neural network and updated. The movement of the simulated droplet can be expressed as: in, is the Gaussian term affected by the splash; , are the weight values ​​before and after the update; , is the bias value before and after the update; is the learning rate; Among them, the splashing influence Gaussian term is combined with the impact area of ​​droplet splashing Defined as: in, Control the impact range of splash; In the adaptive adjustment step, it is necessary to analyze the current optimization state, adjust the impact force and direction of the droplets, and determine the droplet size for the next iteration based on the historical optimization records. and direction Adjustments can be made based on historical optimization records, which can be expressed as: in, It is the impact strength of the previous iteration; is the impact force of the droplet; Furthermore, the direction The adjustment takes into account the delay effect and can be expressed as: in, is the change in energy, is the number of iterations, is the direction of the previous iteration; is the impact direction of the droplet; is the change in energy; is the number of iterations; is a delay factor between 0 and 1; When performing the convergence detection and annealing strategy steps, it is necessary to check whether the parameters tend to be stable; if the expected accuracy is not achieved, the energy of the droplet is adjusted according to the annealing strategy to reduce the impact force of the droplet. If the expected accuracy is achieved, it is determined that the end condition is met, the current optimal parameters are saved, and the algorithm is terminated to use the above algorithm for feature extraction processing to determine the target feature set.

5. A server, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 3.

Images