An intelligent composite fault diagnosis method for photovoltaic modules

Abstract

The present invention discloses an intelligent photovoltaic module composite fault diagnosis method, which relates to the technical field of photovoltaic module operation and maintenance. The method includes the following steps: data preprocessing; local key point search; local key area interval selection; local key area interpolation reconstruction; and construction of a global-local dual-stream collaborative diagnosis framework. This method enables the global model and the local model to be trained using a CNN with the same network structure, converts the original single-label multi-classification problem into a multi-label binary classification problem for solution, and proposes a model parameter adaptive switching mechanism and a global-local dual-stream collaborative model fusion mechanism, effectively improving the accuracy of the model's composite fault diagnosis and facilitating rapid and accurate maintenance after multiple faults occur simultaneously.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic module operation and maintenance, and in particular to an intelligent photovoltaic module composite fault diagnosis method. Background Art

[0002] With the progress of technology, human society is facing unprecedented energy challenges, such as resource shortage, environmental pollution, climate change, uneven development, etc. These problems have prompted countries to vigorously develop clean energy, and new energy power generation has become a national strategy. As a most common renewable energy, photovoltaic power generation can not only be distributedly utilized, but also has the characteristics of cleanliness, high efficiency, easy access, etc., and is receiving increasing attention, and has rapidly become an important part of sustainable development and energy strategy. Photovoltaic modules are the core components of the operation of photovoltaic systems, and the health status of photovoltaic modules will greatly affect the operation safety and power generation efficiency of the entire system. However, since most photovoltaic power stations are built in harsh environments such as deserts or gobi, the components of the photovoltaic system are exposed to the outdoor environment for a long time, and a series of faults will inevitably occur during the operation of the photovoltaic modules, resulting in economic losses in power generation and even safety accidents. In addition, it is difficult to deploy precision sensors on a large scale in the actual field, and it is also difficult to manually screen a variety of large quantities of detection data collected on site, resulting in the difficulty of applying traditional mechanism modeling methods. It is an irresistible trend to realize intelligent fault diagnosis of photovoltaic systems based on artificial intelligence technology.

[0003] Most of the existing intelligent diagnosis methods for photovoltaic faults only consider the diagnosis of single fault types, that is, only one fault will occur in the photovoltaic module at the same time. However, with the continuous expansion of the scale of today's photovoltaic power stations, the series-parallel topological structure of photovoltaic systems has become very complex, and there are often composite faults during the operation of actual photovoltaic arrays, that is, multiple faults occur simultaneously. If the existing methods continue to be used for diagnosis, even if two or more interrelated and interacting faults occur in the array at the same time, the model will ultimately only predict one fault, which means that the existing single fault type diagnosis method cannot comprehensively and accurately reflect the health status of photovoltaic modules, resulting in the difficulty of discovering faults and untimely maintenance. Therefore, it is very necessary to study a new photovoltaic composite fault diagnosis method according to the characteristics of actual photovoltaic array data to quickly and accurately identify various fault types in the case of single fault and multiple faults occurring simultaneously, which is of great significance for reducing photovoltaic operation power generation losses and maintenance costs.

[0004] In the existing photovoltaic fault diagnosis schemes, some establish physical models based on parameter identification methods and compare the theoretical calculation values with the actual test conditions to analyze the fault types. Specifically, this scheme first needs to build a simulation model similar to the actual photovoltaic array structure, and collect the I-V curve data of the corresponding fault types in the actual array through the simulation model. After that, a set of parameters to be identified that can describe the output value of the simulation model is constructed, and the evolutionary algorithm is used to continuously adjust the parameters to be identified of the array simulation model, so that the I-V curve output by the simulation model continuously approaches the measured I-V curve, and finally the error between the two reaches the minimum, so as to identify the fault parameters and judge the fault type. Although this scheme can identify the fault parameters when multiple faults occur simultaneously, it is only applicable to photovoltaic arrays with small scale, simple topological structure and a small number of fault type diagnoses. The parameter identification method adopted by this scheme is essentially a mechanism modeling method, which requires building a corresponding simulation model first. However, for an actual photovoltaic power station, the photovoltaic array is not only huge in scale but also has various complex series-parallel topological structures. It is very difficult and costly to build a simulation model similar to the actual power station structure in advance. No matter how the parameters to be identified are optimized in the existing scheme, the I-V curve output by the simulation model will have a significant difference from the actual I-V curve, which means that the parameter identification method adopted by this scheme cannot be implemented in a large-scale photovoltaic power station. At the same time, when this scheme is applied to the fault diagnosis of different fault types or actual photovoltaic arrays with different topological structures and different scales, it is necessary to rebuild the fault simulation model according to the structure of the actual array and carry out parameter identification again, which means that this scheme does not have portability, and the generalization and accuracy of the model are poor, and the modeling efficiency is low. In addition, the simulation model built by this scheme inevitably has errors. Even if this scheme also expands the parameter search range accordingly, the fault judgment basis it adopts is self-defined and mostly based on empirical judgment. This empirical judgment is easily interfered by factors such as the scale, structure and geographical environment of the actual photovoltaic system, resulting in low accuracy of the identified parameters and inaccurate diagnostic results that are easily affected by the model error.

[0005] In the prior art, there is also a solution that realizes photovoltaic composite fault diagnosis based on the idea of multi-level fault diagnosis. Specifically, in the first-level fault diagnosis of this solution, the power increment ratio is calculated to determine whether a fault occurs in the branch where the photovoltaic grid-connected inverter is located and to judge some fault types; in the second-level fault diagnosis, the current-voltage similarity method is used to determine whether each photovoltaic string under the photovoltaic grid-connected inverter branch has a fault; in the third-level fault diagnosis process, the internal fault components of the photovoltaic string are located based on the trained three-layer BP fault diagnosis neural network. Although this solution can distinguish single faults and composite faults of photovoltaic modules, it simply classifies all single faults that occur simultaneously in two or more types into multiple faults. That is, this solution can only roughly identify various types of single faults and the entire category of multiple faults, and cannot diagnose the specific types of multiple faults. This means that this method can only give a vague and general composite fault prompt in practical applications (only knowing that multiple types of faults occur simultaneously in the system, but not knowing specifically which ones), and maintenance personnel still need to judge one by one for each single fault type, which is extremely inconvenient for actual on-site maintenance and repair. Obviously, in order to minimize the power generation loss caused by component faults to the system, photovoltaic operators not only hope to know that multiple faults have occurred in the array, but more importantly, they need to specifically and accurately know which single faults have occurred simultaneously in order to quickly repair and restore the system operation. At the same time, the fault types that this solution can identify are short circuit, open circuit, occlusion, and multiple faults, and it does not have the ability to identify aging faults, while the present invention can achieve this. The method of the present invention can diagnose more fault types, and can further give specific single fault types that occur simultaneously for multiple faults. In addition, this solution divides the fault diagnosis process into three levels of execution, and both the first level and the second level are based on the mechanism model to calculate the actual values and then compare them with the corresponding reference variables and thresholds to determine the fault types, resulting in low efficiency when analyzing actual data. In addition, the accuracy of the mechanism model will greatly affect the judgment accuracy of this method. For example, once there is a deviation in the result calculated by the mechanism model during the first-level or second-level diagnosis process, it will lead to incorrect diagnosis of the final model. At the same time, the mechanism model also needs to be used in conjunction with the topological structure of the actual photovoltaic array, and the generalization and migration of the model are poor.

[0006] In addition, existing invention solutions involving the idea of global and local information fusion are all targeted at the field of image recognition (face images) or the field of mechanical bearings (vibration time-series data). However, the data used for the photovoltaic module compound fault diagnosis problem are I-V curves and the corresponding temperature and irradiance information (neither images nor time-series data). The image processing methods such as image segmentation and feature map fusion used in existing image field solutions, as well as the time-domain and frequency-domain transformation methods for time-series data such as the fast Fourier transform and wavelet transform, cannot be applied to photovoltaic data. At the same time, most of the existing invention methods need to construct multiple models with different structures to train global and local information respectively, ignoring the implicit connections between various single faults, and the modeling cost is higher and the model stability is reduced. In addition, the information fusion scheme adopted by existing inventions is to weight the prediction probability vectors of multiple sub-models to obtain a fused prediction vector. However, this weighting method needs to be manually adjusted, which requires a high level of prior knowledge (previously accumulated experience) of the operator, and the model needs to be retrained after each adjustment, resulting in poor interpretability of the training process and often failing to achieve the expected accuracy.

[0007] In summary, the disadvantages of the existing technology at least include the following points:

[0008] 1. Most of the existing photovoltaic fault diagnosis technologies are modeled for the identification of single fault types and do not consider the compound fault diagnosis situation where multiple fault types occur simultaneously. However, with the continuous expansion of the scale of current photovoltaic power stations, the series-parallel topological structure of photovoltaic systems has become very complex. When the actual photovoltaic array operates, two or more interrelated and interacting faults often occur simultaneously. However, the single fault type diagnosis model established by the existing technology can only predict one fault and cannot accurately and comprehensively reflect the health status of photovoltaic modules, resulting in difficulties in detecting faults and untimely maintenance.

[0009] 2. The existing photovoltaic compound fault diagnosis technology uses the mechanism modeling method. Whether it is the parameter identification scheme or the multi-level diagnosis scheme, it estimates some unknown variables or parameters in the mechanism model through some theoretical relationships between physical parameters to obtain an approximate I-V curve or a prediction expression of key reference quantities. However, it is difficult to deploy precision sensors on a large scale in the actual field, and it is also difficult to manually screen a variety of large quantities of detection data collected on-site. The parameters in these expressions may not all be collectable in the actual field, resulting in difficulties in applying the existing photovoltaic compound fault diagnosis method based on mechanism modeling. At the same time, there is a certain error between the results calculated by the approximate expressions in mechanism modeling and the true values, which is difficult to eliminate. The program operation volume is large, and it requires a high running speed of the processor, making it difficult to achieve engineeringization.

[0010] 3. Existing photovoltaic composite fault diagnosis technologies still need to be carried out in combination with the topological structure of the actual photovoltaic array during modeling, and often multiple models with different structures need to be established for comprehensive judgment. This not only has low modeling efficiency, but also the generalization and migration of the models are poor.

[0011] 4. Existing photovoltaic composite fault diagnosis technologies ignore the implicit connections between various single faults and do not design a dedicated method to fully utilize the local information of photovoltaic data according to the characteristics of photovoltaic fault data. The number of composite fault types that can be diagnosed is very limited, making it difficult to truly apply to the situation where there are many types of composite faults in the actual field.

[0012] 5. Most of the information fusion schemes in the existing technologies obtain the fusion prediction vector by weighting the prediction probability vectors of multiple sub-models. However, this weighting method requires manual adjustment, has high requirements for the prior knowledge (accumulated experience in advance) of the operator, and the model needs to be retrained after each adjustment. The interpretability of the training process is poor and often cannot achieve the expected accuracy.

[0013] Therefore, based on the above analysis, in the operation and maintenance of photovoltaic power stations, there is an urgent need for an effective method for diagnosing composite faults of photovoltaic components, so that the training and inference of the model are data-driven, and end-to-end rapid fault diagnosis can be achieved. Thus, in the complex situation where multiple fault types occur simultaneously in the actual photovoltaic system, the diagnosis speed of the model for composite faults, as well as the accuracy and generalization of the model, can be effectively improved. For this purpose, the technicians in this field are committed to developing an intelligent method for diagnosing composite faults of photovoltaic components, which can accurately identify the corresponding types when multiple single fault types occur simultaneously. Summary of the Invention

[0014] In view of the above-mentioned defects of the existing technology, the technical problem to be solved by the present invention is how to accurately identify the corresponding types in the complex situation where multiple fault types occur simultaneously in the actual photovoltaic system, and improve the diagnosis speed of the model for composite faults and the accuracy of the model.

[0015] To achieve the above object, the present invention provides an intelligent method for diagnosing composite faults of photovoltaic components, including the following steps:

[0016] Step 1. Data preprocessing;

[0017] Step 2. Local key point search;

[0018] Which includes:

[0019] Step 21. Global sample acquisition; The complete I-V curve after the data preprocessing is recorded as the global sample, and the global sample is composed of 40 discrete data points (x i , y i) is composed and evenly distributed on the curve;

[0020] Step 22, local first-order difference calculation; Define two adjacent sampling points (x m , y m ), (x m+1 , y m+1 ) on the I-V curve and (x m+1 > x m ). The local first-order difference value g′(x m , y m ) at (x m ) is:

[0021]

[0022] Step 23, flat point screening. Define the sampling points with the absolute value of the local first-order difference less than α in the complete I-V curve as flat sampling points; Based on the 39 first-order difference values of adjacent two sampling points in the global samples calculated successively in the above step S22, screen out the set T of flat points on the I-V curve:

[0023] T = find(abs|g′(x i ) < α|), i = 2,..., 39

[0024] Among them, the function of the find() function is to screen out the relative positions of the sampling points that meet the constraint conditions in the parentheses, abs|| is the absolute value function, and α is a pre-given constant value;

[0025] S24, local key point search;

[0026] Which includes: Define the sampling point where the absolute value of the local first-order difference of the flat sampling point is at least β times smaller than that of the adjacent left side as the first local key point, and the sampling point where the absolute value of the local first-order difference of the flat sampling point is at least γ times smaller than that of the adjacent right side as the second local key point, and screen out the set O of local key points on the I-V curve:

[0027] O1 = find(β * abs|g′(x i )| ≤ abs|.g′(x i-1 )|), i = 2,..., 38

[0028] 02 = find(γ * abs|g′(x i )| ≤ abs|g′(x i+1 )|), i = 2,..., 38

[0029] O = 01UO2

[0030] Among them, O1 represents the first subset of local key points, and O2 represents the second subset of local key points respectively. The symbol U represents the union, and β and γ are pre-given constant values;

[0031] Step 3: Selection of local key region intervals;

[0032] It includes:

[0033] Step 31: Group sampling of composite fault samples; The total number T of global samples of composite faults when four single fault types of occlusion, aging, short circuit, and open circuit occur simultaneously in the training set is evenly divided into z groups, and k random samples are obtained by sampling without replacement for each group, that is

[0034]

[0035] Step 32: Marking of local key points; The above-mentioned step S24 of local key point screening is respectively executed for all samples in each group. After execution, the corresponding number of local key points will be marked on the I-V curve of each global sample;

[0036] Step 33: Classification of local key points;

[0037] It includes:

[0038] The local key points are subdivided into local feature head marking points, local feature tail marking points, and local feature middle marking points; Among them, the local feature head marking point is the local key point with the most forward relative position; the local feature tail marking point is the local key point with the most backward relative position; the local feature middle marking point is the remaining marking points excluding the local feature head marking point and the local feature tail marking point;

[0039] Step 34: Classification of randomly grouped data distributions;

[0040] It includes: The data of the local feature head marking points is classified into the local feature head random group The total sample of the corresponding local feature head random group is denoted as X h , the data of the local feature tail marking points is classified into the local feature middle random group The total sample of the corresponding local feature middle random group is denoted as X m , the data of the local feature tail marking points is classified into the local feature tail random group The total sample of the corresponding local feature tail random group is denoted as X l ;

[0041] Step 35: Transformation of data distribution;

[0042] Including: Assume that the head random group sample or the middle random group sample or the tail random group sample is denoted as {X i}, the overall sample is denoted as X, the mean of the overall sample is denoted as μ, and the variance is denoted as σ 2 , the probability distribution function is denoted as Φ, i represents different random groups of the same type, i = 1, 2,..., z; Obviously, each group of samples {X i} satisfies independent and identical distribution, and where and represent the mean and variance of the overall sample respectively;

[0043] Let:

[0044]

[0045] At the same time, let be 's characteristic function, and we can get:

[0046]

[0047]

[0048] where, and are the first derivative and the second derivative of the characteristic function respectively, j is a complex number, so j 2 =-1; thus, we can further obtain the Taylor expansion of the characteristic function of at the origin as:

[0049]

[0050] Denote the characteristic function of X z as and we can get:

[0051]

[0052] When the number of groups z is large approximately follows a normal distribution and the corresponding expectation, variance, and probability distribution are respectively:

[0053]

[0054]

[0055]

[0056] Transform the local key-point distributions of the three types of random groups, namely the head random group samples, the middle random group samples, and the tail random group samples, into approximately following a normal distribution, that is

[0057] Step 36: Construct local key-region intervals; that is, respectively determine the final head local key region, middle local key region, and tail local key region on the premise of ensuring at least 95% confidence level;

[0058] Including:

[0059] Step 361: For the local feature head random group According to the following formula

[0060]

[0061]

[0062]

[0063]

[0064] Obtain the head local key-region interval as where i = 1, 2,..., z;

[0065] Step 362: For the local feature middle random group According to the following formula

[0066]

[0067]

[0068]

[0069] Obtain the middle local key-region interval as where i = 1, 2,...., z;

[0070] Step 363: For the local feature tail random group According to the following formula

[0071]

[0072]

[0073]

[0074]

[0075] The obtained tail partial key area interval is where i = 1, 2,..., z;

[0076] Step 4, Interpolation reconstruction of the partial key area; Reconstruct the head partial key area, the middle partial key area, and the tail partial key area according to the partial key area interpolation method;

[0077] where the partial key area interpolation method includes:

[0078] Step 41, Based on the global dimension, calculate the number of resampling points required between every two original sample points within the interval;

[0079] Step 42, Obtain the function values and local first-order difference values of two adjacent interpolation points x m , x m+1 on the I-V curve, which are respectively:

[0080] y i = g(x i ), t i = g′(x i ), (i = m, m + 1),

[0081] Step S43, Construct the local feature interpolation polynomial C(x) to satisfy:

[0082] C(x i ) = y i , C′(x i ) = t i , (i = m, m + 1)

[0083] Introduce four pointer functions r1(x), r2(x), s1(x), s2(x) to represent two components C i , (i = 1, 2) of the local feature interpolation function C(x), and they are all polynomials of order less than or equal to two:

[0084] C1(x) = y m r1(x)+t m s1(x)

[0085] C2(x) = y m+1 r2(x)+t m+1 s2(x)

[0086]

[0087] β i (x j ) = 0, β′ i (x j ) = δ ij, (i = 1, 2),

[0088] C(x i ) = y i , C′(x i ) = y′0, (i = m, m + 1),

[0089] The local feature interpolation function C(x) is expressed as:

[0090] C(x) = C1(x) + C2(x)

[0091] x m+1 is a double zero of r1(x), and the pointer function value and the local difference function value of r1(x) at x m+1 are the same and both are zero. Let:

[0092] r1(x) = px 2 + px m+1 2 + qx 2 - 2px * x m+1 + qx m+1 2 - 2qx * x m+1

[0093] From r1(x m+1 ) = r1′(x m+1 ) = 0, we can get:

[0094]

[0095]

[0096]

[0097] By interchanging x m , x m+1 , we can get:

[0098]

[0099] Similarly, we can get:

[0100]

[0101]

[0102] The local interpolation formula between two points is:

[0103] C(x) = C1(x) + C2(x) = y m r1(x) + t m s1(x) + y m+1 r2(x) + t m+1 s2(x);

[0104] Step 5. Construct a global-local dual-stream collaborative diagnosis framework;

[0105] It includes:

[0106] Step 51. Obtain global samples;

[0107] Step 52. Obtain local samples to obtain initial head local samples, initial middle local samples, and initial tail local samples;

[0108] Step 53. Label encoding and dataset construction;

[0109] It includes:

[0110] Encode the label of each fault type into a binary vector with a dimension of 1×4, corresponding to 4 single fault types respectively. The first column represents the occlusion fault, the second column represents the aging fault, the third column represents the short-circuit fault, and the fourth column represents the open-circuit fault;

[0111] Encode a flag bit for the global samples, the initial head local samples, the initial middle local samples, and the initial tail local samples respectively;

[0112] Step 54. Build a shared CNN network model;

[0113] Step 55. Perform adaptive switching of model parameters;

[0114] It includes:

[0115] Step 551. Initialize a shared CNN network, denote the network parameter vector of the global multi-label classification model as W g , denote the network parameter vector of the head local multi-label classification model as W h , the network parameter vector of the middle local multi-label classification model as W m , and the network parameter vector of the tail local multi-label classification model as W l , and combine them to construct a global-local model parameter vector W switch :

[0116] W switch =[W g , W h , W m , W l

[0117] Step 552. Input the sample to be diagnosed into the shared CNN network and preferentially read the flag bit of the sample to identify which one of the global samples, the initial head local samples, the initial middle local samples, or the initial tail local samples the sample belongs to;​

[0118] Step 553: Based on the recognition result of the flag bit, the shared CNN network model adaptively selects the corresponding model parameter vector from the global-local model parameter vector W switch to load and implement model parameter switching. Denote the model parameter switching as W g , W h , W m and W l . The shared CNN network models when W g , CNN h , CNN m and CNN l are CNN g denoted as the multi-label classification global model, CNN h , CNN m and CNN l denoted as the multi-label classification model head local model, multi-label classification middle local model, and multi-label classification tail local model respectively;

[0119] Step 554: Use the shared CNN network model after model parameter switching to predict the sample to be diagnosed, and obtain the multi-label prediction results corresponding to the multi-label classification global model, the multi-label classification head local model, the multi-label classification middle local model, and the multi-label classification tail local model respectively;

[0120] Step 56: Execute the global-local model fusion mechanism to obtain the diagnosis result.

[0121] Furthermore, the step 52 includes:

[0122] Step 521: Positioning; that is, perform the local key point search on the global sample, and locate and mark the corresponding local key points on the global sample;

[0123] Step 522: Screening; that is, perform the construction of the local key region interval on the global sample with annotations, and respectively screen out the head local key region interval, the middle local key region interval, and the tail local key region interval from the global sample to obtain the head initial local sample, the middle initial local sample, and the tail initial local sample;

[0124] Step 523: Reconstruction; perform the local key region interpolation method on the head initial local sample, the middle initial local sample, and the tail initial local sample to ensure that the finally interpolated and reconstructed sample has the same data dimension as the global sample and can be trained using a neural network model with the same structure.

[0125] Furthermore, the step 56 includes:

[0126] Step 561: Based on the model parameter adaptive switching mechanism, use the global model, the head local model, the middle local model, and the tail local model to test the samples in the validation set respectively, and obtain the global information flow, the head local information flow, the middle local information flow, and the tail local information flow respectively; define the global sample to input the global multi-label classification model CNN g The final predicted vector obtained is the global information flow, and define three local samples of the head, middle, and tail to input the corresponding local multi-label classification model CNN h , CNN m and CNN l The final predicted vector obtained and are the local information flows, as shown below.

[0127]

[0128]

[0129]

[0130]

[0131] Furthermore, step 56 further includes:

[0132] Step 562: By testing all the samples in the validation set and inputting the above and information flows into the threshold function calculate the accuracy rate of the shared CNN model for identifying various single fault types under different model parameters, and the average per-class accuracy rate vectors corresponding to the four information flows obtained are:

[0133]

[0134]

[0135]

[0136]

[0137] Among them, represents the recognition accuracy rate of the multi-label classification global model for occlusive faults on the validation set, and respectively represent the recognition accuracy rates of the multi-label classification model head local model, the multi-label classification middle local model, and the multi-label classification tail local model for occlusive faults on the validation set.

[0138] Furthermore, step 56 further includes:

[0139] In step 563, according to the test situation of the validation set, find the model with the most accurate recognition of various single fault types, then fuse the corresponding information flows to obtain the final global-local dual-stream collaborative probability prediction vector, and further input it into the threshold function to obtain the final fault diagnosis result of the fusion model:

[0140]

[0141]

[0142] where i represents four single fault types, A max The function is to return the information flow probability prediction value corresponding to the maximum value in the same column of the four accuracy vectors in step 562;

[0143] By judging i = 1, 2, 3, 4 respectively, A max A total of 4 probability prediction values are returned to form to obtain the fusion prediction vector of the global-local dual-stream collaborative model, that is, to determine which prediction result of the corresponding column of the information flow is taken for the 1st to 4th columns of the final fusion model respectively. Based on the threshold function calculate to obtain the diagnosis result Result.

[0144] Furthermore, in the shared CNN network model, in cooperation with the Sigmoid function and the threshold function the display expression of the output photovoltaic composite fault diagnosis classification result is:

[0145] x = [x i , i = 1, 2, 3, 4

[0146]

[0147]

[0148] where x is the output vector representing the last fully connected layer, which contains 4 items in total. Input it into the Sigmoid function to obtain the prediction probability vector for the occurrence of 4 single fault types represents the prediction probability of the occurrence of single fault type i, and τ represents the threshold; use the threshold function for make a further discrimination. If the prediction probability is greater than or equal to τ, it is considered that this type of single fault has occurred. Finally, the final photovoltaic composite fault diagnosis classification result is obtained by synthesizing the discrimination results of 4 columns;

[0149] The selected loss function is Binary Cross Entropy (BCE), and the corresponding mathematical expression is as follows:

[0150]

[0151]

[0152] Among them, and represent the global model loss function and the local model loss function respectively, C represents the number of single fault types, K represents the total number of training samples, represents the true label of single fault type j in the i-th sample, which is 1 if it occurs and 0 if it does not occur, and are output by the threshold function and represent the predicted labels of the global model and the local model for single fault type j in the i-th sample respectively.

[0153] Furthermore, the global-local model fusion is obtained by adopting other model fusion methods.

[0154] Furthermore, the α is set to 10 -4 .

[0155] Furthermore, the β is set to 5 and the γ is set to 3.

[0156] Furthermore, the τ is set to 0.5.

[0157] Compared with the existing technical solutions, the beneficial technical effects of the present invention at least include:

[0158] (1) The present invention considers the complex situation that multiple faults may occur simultaneously in photovoltaic modules. Different from most of the existing technologies that adopt mechanism modeling methods for compound fault diagnosis, the present invention proposes a new global-local dual-stream collaborative photovoltaic compound fault diagnosis framework. The training and inference of the model are data-driven, and end-to-end fast fault diagnosis can be realized. Therefore, when multiple fault types occur simultaneously in the actual photovoltaic system, the problems of low modeling efficiency, poor accuracy and generalization of the existing technology model are solved.

[0159] (2) Different from the existing technologies that only use global information for photovoltaic fault diagnosis, the present invention designs a new local key point search algorithm, a local key area interval selection method and a local key area interpolation algorithm according to the characteristics of photovoltaic data to further obtain local information.

[0160] (3) The method of the present invention starts from the perspective of the compound fault mechanism, regards the compound fault as the superposition of multiple known faults, transforms the compound fault diagnosis problem into the identification of multiple single faults, that is, transforms the original single-label multi-classification problem into a multi-label binary classification problem for solution. By synergistically utilizing the global information and key local information of the photovoltaic fault data, the problems of fewer diagnosed fault types and low accuracy of compound fault diagnosis in the prior art are solved.

[0161] (4) The present invention designs a new model parameter adaptive switching mechanism, enabling the global and local samples to share the same structure of the CNN model for training, overcoming the deficiency of the prior invention method that requires multiple different structure network models to be trained separately. This not only improves the modeling efficiency but also enhances the robustness of the diagnostic model.

[0162] (5) The present invention designs a new global-local model fusion mechanism, which can directly correct the recognition accuracy of the global model for some single fault categories through the local model, solving the problem that the prior invention method needs to manually weight and fuse the prediction probability vectors based on prior knowledge or manual experience.

[0163] The following will further illustrate the concept, specific structure and technical effects of the present invention in conjunction with the accompanying drawings to fully understand the purpose, features and effects of the present invention. Description of the Drawings

[0164] Figure 1 is the overall flowchart of a preferred embodiment of the present invention;

[0165] Figure 2 is the local first-order difference schematic diagram of a preferred embodiment of the present invention;

[0166] Figure 3 is the interpolation effect diagram of the head local key area of a preferred embodiment of the present invention;

[0167] Figure 4 is the interpolation effect diagram of the middle local key area of a preferred embodiment of the present invention;

[0168] Figure 5 is the interpolation effect diagram of the tail local key area of a preferred embodiment of the present invention;

[0169] Figure 6 is the photovoltaic compound fault diagnosis framework based on global-local dual-stream collaboration of a preferred embodiment of the present invention;

[0170] Figure 7 is the multi-label classification CNN network structure diagram of a preferred embodiment of the present invention. Detailed Embodiments

[0171] The following describes multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0172] During the daily operation of a photovoltaic system, the single fault types that most easily occur in photovoltaic modules include: occlusion faults (shadows), aging faults, short-circuit faults, and open-circuit faults. However, in actual operation of a photovoltaic array, compound faults often exist, that is, two or more of the above four types of single faults occur simultaneously. Therefore, based on these four common single faults of occlusion, aging, short circuit, and open circuit, the present invention conducts research on compound fault diagnosis. Through permutation and combination, a total of sixteen fault types are obtained, including the normal state, four single faults, and eleven compound fault types. The original volt-ampere characteristic curves (I-V curves), as well as temperature and irradiance, are collected respectively under the corresponding states. That is, the original data includes four columns of features: current, voltage, temperature, and irradiance. The proposed compound fault diagnosis method for photovoltaic modules includes the following steps:

[0173] Step 1: Data preprocessing. On the one hand, the number of sampling points on different original I-V curves is different (data dimensions are different), and it is impossible to use a unified network input dimension for model training. On the other hand, these sampling points are unevenly distributed on the curve, and directly training the model will lead to network weight mismatch. Therefore, the present invention uses the methods of bilinear interpolation, upsampling, and data reconstruction to uniformly preprocess the original I-V curves. Specifically, it includes:

[0174] S11: Read the original I-V curve and record the open-circuit voltage Vo and short-circuit current Isc;

[0175] S12: Downsample the data, and resample 20 voltages V equidistantly within the range of [0, Vo], Rx and resample 20 currents I equidistantly within the range of [0, Isc], Rx to compress the data volume of the original volt-ampere characteristic curve and thus save computing resources;

[0176] S13: Perform bilinear interpolation on the data to calculate the voltage values corresponding to the 20 resampled currents I Rx and the current values corresponding to the 20 resampled voltages V Rx ; and use the following formula to calculate the voltage V Rx corresponding to the 20 resampled currents I Rx_n and the voltage I Rx corresponding to the 20 resampled voltages V Rx_n :

[0177]

[0178]

[0179] Among them, (I1, V1) and (I2, V2) are the sampling points closest to both sides of I Rx and V Rx on the original volt-ampere curve.

[0180] S14. Obtain the resampled volt-ampere characteristic curve, arrange the 40 resampled points obtained by bilinear interpolation in descending order of voltage to obtain a 40×2 array;

[0181] S15. Data reconstruction: Construct the temperature and irradiance corresponding to the 40 resampled points into a 40×2 environmental vector, and reconstruct it with the 40×2 volt-ampere characteristic curve array after resampling into a 40×4 two-dimensional array.

[0182] Step 2: Local key point search. Considering that the more single fault types occur simultaneously, the more distorted the corresponding composite fault I-V characteristic curve is compared with the I-V curve under normal conditions, that is, the more fault feature regions are superimposed. Therefore, the number of key local regions on different types of original I-V curves is different. In order to ensure that local samples can contain key fault feature information as much as possible, the key local search algorithm proposed in the present invention selects the most complex composite fault situation (simultaneous occurrence of four single fault types of occlusion, aging, short circuit, and open circuit) to locate the relative positions of key local sampling points. Here, the relative position refers to which point among the 40 discrete data points the marked local key sampling point is. It mainly includes:

[0183] S21. Global sample acquisition: Uniformly perform data preprocessing on the original I-V curve. The complete I-V curve after data preprocessing is recorded as the global sample, which is composed of 40 discrete data points, and these resampled points are evenly distributed on the curve;

[0184] S22. Local first-order difference calculation: Considering that the global sample (complete photovoltaic I-V curve) obtained after data preprocessing is composed of 40 discrete data points (x i , y i ), the derivative value of each point cannot be calculated like a continuous function. Therefore, the present invention starts from the perspective of data difference and defines the local first-order difference g′(x m , y m ), (x m+1 , y m+1 ) and (x m+1 >x m ) at (x m , y m ) as: m ) is:

[0185]

[0186] S23. Gentle point screening: Define the sampling points where the absolute value of the local first-order difference in the complete I-V curve is less than α as gentle points. Based on the method in S1 above, calculate the first-order difference values between adjacent two sampling points in the global sample in turn (a total of 39 local first-order difference values are calculated for 40 discrete points). Taking a composite fault sample where four fault types of a certain occlusion, aging, short circuit, and open circuit occur simultaneously as an example, the local first-order differences corresponding to its complete I-V curve are as shown in the appendix Figure 2 and screen out the set T of gentle points on the I-V curve:

[0187] T = find(abs|g′(x i ) < α|), i = 2,..., 39

[0188] where the function of find() is to screen out the relative positions of the sampling points that meet the constraint conditions in the parentheses, abs|| is the absolute value function, and α is a constant value given in advance. In the present invention, the constant value α is set to 10 -4 .

[0189] S24. Local key point screening: Define the sampling points where the gentle sampling point is at least β times smaller than the absolute value of the local first-order difference on the adjacent left side or the gentle sampling point is at least γ times smaller than the absolute value of the local first-order difference on the adjacent right side as local key points, and screen out the set O of local key points on the I-V curve:

[0190] O1 = find(β * abs|g′(x i )| ≤ abs|g′(x i - 1)|), i = 2,..., 38

[0191] O2 = find(γ * abs|g′(x i )| ≤ abs|g′(x i+1 )|), i = 2,..., 38

[0192] O = O1 U O2

[0193] where O1 and O2 respectively represent two subsets of local key points, the symbol U represents the union, and β and γ are constant values given in advance. In the present invention, the constant value β is set to 5, and the constant value γ is set to 3.

[0194] Step 3: Selection of the interval of the local key area. Since there are significant differences in the original I-V curves of compound faults collected under different temperatures and irradiation states, and the I-V curves after data preprocessing are composed of 40 discrete data points, the relative positions of the local key sampling points obtained in Step 2 on the I-V curves are different among different samples. To unify the intervals of the local key areas of samples of different fault types, the present invention further designs a method for selecting the interval of the local key area, which specifically includes:

[0195] S31. Sampling by grouping compound fault samples. The present invention evenly divides the total number T of global samples of compound faults when four single fault types of occlusion, aging, short circuit, and open circuit occur simultaneously in the training set into z groups, and k random samples are obtained by sampling without replacement for each group, so that:

[0196]

[0197] S32. Marking local key points. The previously described local key point search algorithm is executed for all samples in each group. After execution, the corresponding number of local key sampling points will be marked on the I-V curve of each global sample;

[0198] S33. Classification of local key points. The present invention considers how to determine the local key area from several local key sampling points marked on the same I-V curve. In the present invention, according to the relative positions of the key points obtained on a compound fault I-V curve by the search algorithm, these local key points are further divided into three major categories: (1) The local key point with the most forward relative position (the first marked point) is denoted as the head marked point of the local feature; (2) The local key point with the most backward relative position (the last marked point) is denoted as the tail marked point of the local feature; (3) The median of the remaining marked points (excluding the head and tail marked points) is denoted as the middle marked point of the local feature;

[0199] S34. Classification of randomly grouped data distribution. To more accurately obtain the data distribution of these key points, the present invention further divides the relative positions of the local key sampling points in each group of randomly selected compound fault samples into three major categories according to the above head, middle, and tail rules, that is, the local feature head random group, the local feature middle random group, and the local feature tail random group. The corresponding random group samples are respectively denoted as and For the total samples of the three types of local key points described above, they are denoted as X h , X m and X l ;

[0200] S35. Data distribution transformation. Based on the above step S34, perform data distribution transformation on the random groups of the three types of local key points. Since the analysis methods for the three types of local key points are similar, only one type will be taken as an example below to specifically introduce the data distribution analysis method proposed by the present invention. Assume that the random group samples of a certain type (head, middle, or tail) are denoted as {X i}, the overall sample is denoted as x, the mean of the overall sample is denoted as μ, the variance is denoted as σ 2 , the probability distribution function is denoted as Φ, and i represents different random groups of the same type, i = 1, 2,..., z. Obviously, each group of samples {X i} satisfies independent and identical distribution, and where and respectively represent the mean and variance of the overall sample. Further, in the present invention, let:

[0201]

[0202] Meanwhile, let be 's characteristic function, and we can get:

[0203]

[0204]

[0205] where and are respectively the first derivative and the second derivative of the characteristic function , j is a complex number, so j 2 = -1. Thus, we can further obtain that the characteristic function of at the origin has the Taylor expansion:

[0206]

[0207] For the convenience of expression, in the present invention, the characteristic function of X z is denoted as and we can get:

[0208]

[0209] Therefore, when the number of groups z is relatively large approximately follows a normal distribution and the corresponding expectation, variance, and probability distribution are respectively:

[0210]

[0211]

[0212]

[0213] After the above steps, the local key point distributions of the three types of random groups can be transformed into approximately following a normal distribution, that is Suppose there are 40 sampling points on the I-V curve after data preprocessing, which are respectively denoted as IV(x), where x = 1, 2,..., 40. In the present invention, V|b| is defined as the voltage value (the abscissa value of the I-V curve) of the nearest sampling point on the I-V curve after data preprocessing whose voltage value is not less than b, and is denoted as the upper bound of optimization. U|b| is the voltage value (the abscissa value of the I-V curve) of the nearest sampling point on the I-V curve after data preprocessing whose voltage value is not greater than b, and is denoted as the lower bound of optimization. L(V|b|) and L(U|b|) respectively correspond to the relative coordinates of the upper bound of optimization and the lower bound of optimization on the corresponding I-V curve (that is, which point among the 40 sampling points on the I-V curve the corresponding abscissa corresponds to);

[0214] S36. On the basis of steps S31 - S35, the present invention further improves the confidence interval method for the local sample data distributions of three different categories, so as to respectively determine the final local key regions of the head, middle, and tail on the premise of ensuring at least 95% confidence level. The following will be specifically introduced in three major categories:

[0215] ① For the random group of local feature heads It is only necessary to determine the upper bound of the head region interval and the upper bound of optimization, then determine the relative coordinate of the upper bound of optimization, and take the maximum value of the relative coordinates of the upper bounds of optimization for all random groups of local feature heads. That is, the region between the first sampling point IV(0) and the maximum relative coordinate of the upper bound of optimization on the I-V curve after data preprocessing is divided into the local key region of the head of the present invention. From the probability function of the normal distribution, it can be obtained that:

[0216]

[0217]

[0218]

[0219]

[0220] It can be obtained that the interval of the local key region of the head is where i = 1, 2,..., z.

[0221] ② For the random group of local feature middles It is necessary to separately determine the upper bound, lower bound, optimized lower bound, and optimized upper bound of the middle region interval. Then, determine the relative coordinates of the optimized upper bound and optimized lower bound, take the maximum value of the relative coordinates of the optimized upper bound for all randomly selected groups of local features in the middle, and take the minimum value of the relative coordinates of the optimized lower bound for all randomly selected groups of local features in the middle. Then, expand three sampling points on each side, that is, divide the region between the third sampling point to the left of the minimum optimized lower bound relative coordinate and the third sampling point to the right of the maximum optimized upper bound relative coordinate into the middle local key region of the present invention. From the probability function of the normal distribution, it can be obtained that:

[0222]

[0223]

[0224]

[0225] It can be obtained that the interval of the middle local key region is where i = 1, 2,..., z.

[0226] ③ For the randomly selected groups of local features at the tail It is only necessary to determine the lower bound and optimized lower bound of the tail region interval. Then, determine the relative coordinate of the optimized lower bound, and take the minimum value of the relative coordinates of the optimized lower bound for all randomly selected groups of local features at the tail, that is, divide the region between the maximum optimized upper bound relative coordinate and the last sampling point IV(40) on the I-V curve after data preprocessing into the tail local key region of the present invention. From the probability function of the normal distribution, it can be obtained that:

[0227]

[0228]

[0229]

[0230]

[0231] It can be obtained that the interval of the tail local key region is where i = 1, 2,..., z.

[0232] Step 4: Interpolation reconstruction of the local key region. Considering that the number of sample points included in different local key regions obtained by the above local key point search algorithm and local key region interval selection method is different, the present invention further designs a local key region interpolation algorithm to ensure that the local samples after final interpolation reconstruction have the same data dimension as the global samples and can be trained using a CNN sharing the same network structure. Specifically, it includes:

[0233] S41. Each local region calculates the number of resampling points required between every two original sample points within the interval based on the global dimension (the number of complete I-V curve sampling points is 40).

[0234] S42. Based on the difference calculation method described in the previous local key point search algorithm, the function values and local first-order difference values of two adjacent interpolation points x m , x m+1 on the I-V curve are respectively:

[0235] y i = g(x i ), t i = g′(x i ), (i = m, m + 1),

[0236] S43. In order to further supplement and amplify the information of the local key region, the present invention constructs a local feature interpolation polynomial C(x) such that it satisfies:

[0237] C(x i ) = y i , C′(x i ) = t i , (i = m, m + 1)

[0238] To facilitate the solution and description of the interpolation function, the present invention further introduces four pointer functions r1(x), r2(x), s1(x), s2(x) to represent two components C i , (i = 1, 2) of the local feature interpolation function C(x), and they are all polynomials of order less than or equal to two:

[0239] C1(x) = y m r1(x) + t m s1(x)

[0240] C2(x) = y m+1 r2(x) + t m+1 s2(x)

[0241]

[0242] β i (x j ) = 0, β′ i (x j ) = δ ij , (i = 1, 2),

[0243] C(x i ) = y i , C′(x i ) = y′0, (i = m, m + 1),

[0244] Therefore, the local feature interpolation function C(x) can be expressed by four interpolation basis functions as follows:

[0245] C(x) = C1(x) + C2(x)

[0246] From the function values of the pointer function and the local difference function mentioned above, it can be known that: x m+1 is a double zero of r1(x), and the function values of the pointer function and the local difference function of r1(x) at x m+1 are the same and both are zero. Let:

[0247] r1(x) = px 2 + px m+1 2 + qx 2 - 2px * x m+1 + qx m+1 2 - 2qx * x m+1

[0248] From r1(x m+1 ) = r1′(x m+1 ) = 0, it can be obtained that:

[0249]

[0250]

[0251]

[0252] Using the above method, by interchanging x m , x m+1 , it can be obtained that:

[0253]

[0254] Similarly, it can be obtained that:

[0255]

[0256]

[0257] To sum up, the local interpolation formula between two points can be obtained as follows:

[0258] C(x) = C1(x) + C2(x) = y m r1(x) + t m s1(x) + y m+1 r2(x) + t m+1 s2(x)

[0259] Taking the composite fault samples when four single fault types of occlusion, aging, short circuit, and open circuit occur simultaneously as an example, the results before and after interpolation in the local key areas of the head, middle, and tail are shown in Appendix Figure 3 , 4 and Figure 5.

[0260] Step 5: Construct a global-local dual-stream collaborative diagnosis framework. For the diagnosis of photovoltaic composite faults, relying solely on the complete I-V curve for global diagnosis or local key areas for local diagnosis cannot accurately identify similar faults. This is because global information often fails to highlight local details, while local information is difficult to comprehensively and accurately capture the overall fault characteristics. Based on the above local key point search algorithm, local key area interval selection method, and local key area interpolation algorithm, the present invention further explores key local features and integrates them with global information, and proposes a new global-local dual-stream collaborative diagnosis framework for the characteristics of photovoltaic data, as shown in Appendix Figure 6 as follows:

[0261] S51. Global sample acquisition: Preprocess the original I-V curve data based on bilinear interpolation, upsampling, and data reconstruction. In the present invention, the complete I-V curve after data preprocessing is called the global sample, the information contained therein is called global information, and the model trained based on the global sample is called the global model;

[0262] S52. Local sample acquisition: The acquisition of local samples mainly includes three parts: positioning, screening, and reconstruction. Considering that the distortion degrees of the global sample I-V curves of different fault types are different compared with the normal I-V curve, in order to ensure that the local samples can contain as much key fault feature information as possible, the positioning and screening steps are carried out based on the most complex composite fault situation (the simultaneous occurrence of four single fault types of occlusion, aging, short circuit, and open circuit). It should be noted that after the optimal local key area interval is screened out, this interval will be uniformly applied to the samples of all fault types, that is, the local key area intervals of all fault types are finally fixed and the same, specifically including:

[0263] ① Positioning: Execute the above local key point search algorithm on the global sample, so that the corresponding local key points can be located and marked on the global sample;

[0264] ② Screening: Execute the above local key area interval selection method on the marked global sample, so as to screen out the local key area intervals of the head, middle, and tail from the global sample respectively (all fault types use these three intervals). It should be noted that these three intervals can ensure a confidence level of more than 95%, guaranteeing the effectiveness of local features. At this time, three initial local samples can be obtained from the three intervals screened out from each global sample;

[0265] ③ Reconstruction: Since the dimensions of different initial local samples are different (the number of original sampling points included is different), the above-mentioned local key area interpolation algorithm is further executed on these initial local samples, so as to ensure that the finally interpolated and reconstructed samples have the same data dimension as the global samples and can share the neural network model with the same structure for training;

[0266] In the present invention, the I-V curve obtained after positioning, screening and reconstruction is called a local sample, the information contained therein is called local information, and the model trained based on the local sample is called a local model.

[0267] S53. Label encoding and dataset construction: In order to train a supervised learning model using a neural network subsequently, it is necessary to further perform label encoding on the global samples and local samples. Most existing patents adopt a single-label multi-classification scheme. Whether it is a single fault type or a compound fault type, the data label corresponding to the sample is a single value (for example, the labels of four single faults are 1-4 respectively, and the label of a certain compound fault is 5). That is, the compound fault is simply regarded as a new type and trained together with the single fault samples. As the number of compound fault types increases, the data labels will increase exponentially, making it difficult to train the model. Therefore, starting from the perspective of the compound fault mechanism, the present invention regards the compound fault as the superposition of multiple known faults, transforms the compound fault diagnosis problem into the identification of multiple single faults, that is, transforms the original single-label multi-classification problem into a multi-label binary classification problem for solution. Specifically, in the present invention, all compound faults can be subdivided into the superposition of 4 single fault types: occlusion, aging, short circuit, and open circuit. Therefore, the present invention adopts a multi-label encoding scheme, encoding the label of each fault type into a binary (0 and 1) vector with a dimension of 1×4, corresponding to 4 single fault types respectively (0 means that the fault has not occurred, 1 means that the fault has occurred). The first column represents the occlusion fault, the second column represents the aging fault, the third column represents the short circuit fault, and the fourth column represents the open circuit fault. For example, the label [1, 0, 0, 0] means that only the occlusion fault has occurred at this time, and the label [1, 0, 1, 0] means that a compound fault has occurred and is composed of the occlusion and short circuit faults. In addition, a flag bit is separately encoded for the global samples and the three types of local samples, which is convenient for subsequent model training and diagnosis. The flag bit of the global model is g, and the flag bits of the head, middle and tail local samples are h, m and l respectively;

[0268] S54. Build a shared CNN network model: Since the method in the present invention can ensure that the global samples and local samples have the same data dimension, a CNN with the same network structure can be shared to train the global and local models. The shared CNN network is specifically introduced below:

[0269] ① Convolution layer: The introduction of the standard two-dimensional convolution layer (Conv2d) can facilitate the implementation of network weight sharing, thus significantly reducing the free parameters in the model training process. At the same time, it also helps to improve the performance of the network. The calculation processes of the two-dimensional convolution layer and the one-dimensional convolution layer are defined as follows:

[0270]

[0271]

[0272] Among them, x represents the input data (two-dimensional matrix), i and j respectively correspond to the coordinates in the input two-dimensional matrix. The kernel matrix is represented by K, and its dimension size is m×m, which is a parameter that can be learned during model training. The coordinate indices of the two-dimensional kernel matrix are represented by w and h respectively, and the kernel size of the k-th input neuron is represented by K k denotes, and the symbol * represents the cross-correlation operator without zero-padding.

[0273] ② Max pooling layer: Its main function is to increase the network receptive field, reduce information redundancy, reduce the model calculation amount, reduce the network optimization difficulty, prevent network overfitting, and make the model more robust to feature changes in the input data. Max Pooling is to divide the input data into several rectangular regions and output the maximum value for each sub-region. Its definition is as follows:

[0274]

[0275] Among them, x kpq represents the element at the (p, q) position in the rectangular region related to the k-th feature map, and y kij represents the maximum pooling output value related to the k-th feature map in the rectangular region ;

[0276] ③ ReLU activation function: The full name of ReLU is Rectified Linear Unit, which means rectified linear unit. The ReLU activation function is a commonly used neural activation function. It is essentially a piecewise linear function that changes all negative values to 0 while keeping positive values unchanged. This operation is called unilateral inhibition. Its mathematical expression is as follows:

[0277]

[0278] Among them, x is the input feature. ReLU has sparsity, which can enable the sparse model to better mine relevant features and fit the training data; in the region where x≥0, there will be no problems of gradient saturation and gradient disappearance, the calculation complexity is low, and no exponential operation is required. As long as a threshold is used, the activation value can be obtained.

[0279] ④ Fully Connected Layer (FC) and Sigmoid Function: The main function of the fully connected layer is to map the feature space samples obtained from the previous layers (such as convolutional and pooling layers) to the sample label space. Simply put, it integrates the feature representations into a single value. Its advantage is that it reduces the influence of feature positions on the classification results and improves the robustness of the entire network. Its mathematical expression is as follows:

[0280] y = W N×D x D×M + b N×M

[0281] Another function of the fully connection is to act as a classifier. By combining with an appropriate output function, it can output the final classification result of the network. For the present invention, since the most complex compound fault situation is the simultaneous occurrence of four single faults: occlusion, aging, short circuit, and open circuit, the final output of the network should be a vector with a dimension of 1×4, where the value of each column corresponds to the predicted probability of the occurrence of a single fault type. Since the FC is placed in the last layer, combined with the Sigmoid function and the threshold function the output is the display expression of the photovoltaic compound fault diagnosis classification result:

[0282] x = [x i , i = 1, 2, 3, 4

[0283]

[0284]

[0285] where x is the output vector representing the last fully connected layer, which contains a total of 4 items. Inputting it into the Sigmoid function gives the predicted probability vector corresponding to the occurrence of 4 single fault types represents the predicted probability of the occurrence of single fault type i, and τ represents the threshold. Using the threshold function for to make a further discrimination. If the predicted probability is greater than or equal to τ, it is considered that the single fault of this type has occurred. Finally, combining the discrimination results of 4 columns gives the final photovoltaic compound fault diagnosis classification result. In the present invention, the constant value τ is set to 0.5.

[0286] Since the present invention transforms the compound fault diagnosis problem into a multi-label binary classification problem for solution, the selected loss function is Binary Cross Entropy (BCE), and the corresponding mathematical expression is as follows:

[0287]

[0288]

[0289] Among them, and respectively represent the global and local model loss functions, C represents the number of single fault types, K represents the total number of training samples, represents the true label (occurrence is 1, non-occurrence is 0) of single fault type j in the i-th sample, and are output by the above threshold function and respectively represent the predicted labels of the global and local models for single fault type j in the i-th sample.

[0290] It should be noted that different from the existing patent methods that require multiple network models with different structures, the present invention reconstructs the local information to ensure that the global and local samples can share the same structure of CNN model training, which not only improves the modeling efficiency but also enhances the robustness of the diagnostic model. By transforming the original compound fault diagnosis problem of photovoltaic modules into a multi-label binary classification problem, the present invention can train an accurate multi-label classification model only using one shared CNN. The CNN network structure includes a two-dimensional convolutional layer, a one-dimensional convolutional layer, a max pooling layer, a fully connected layer, and a Sigmoid output function. The specific multi-label classification CNN structure is as shown in the appendix Figure 7 as follows.

[0291] S55. Model parameter adaptive switching mechanism: Use the global samples to train the global multi-label classification model, and use the head, middle, and tail local samples to train three local multi-label classification models according to different flag bits. During actual diagnosis, each unknown fault sample to be diagnosed (the original I-V curve with unknown true label) first obtains the global sample and the three local samples of the head, middle, and tail through the above steps 1 to 3. As described in step 4, since the present invention ensures that the global samples and local samples have the same data dimension, they can share the same structure of the CNN network. Therefore, when performing model inference diagnosis, there is no need to input the samples into different machine learning models one by one to calculate the results separately as in the existing patents. Instead, it can be achieved through the model parameter adaptive switching mechanism designed by the present invention, which specifically includes:

[0292] ① Based on step 4, initialize a shared CNN network, denote the network parameter vector of the global multi-label classification model as W g , denote the network parameter vectors of the three local multi-label classification models of the head, middle, and tail as W h , W m and W l , and combine them to construct the global-local model parameter vector W switch :

[0293] W switch = [Wg ,W h ,W m ,W l

[0294] ② Input the sample to be diagnosed into the shared CNN network and preferentially read the flag bit of the sample, so as to identify whether the sample belongs to the global sample or the three types of local samples;

[0295] ③ Based on the recognition result of the flag bit, the shared CNN network model adaptively selects the corresponding model parameter vector from the global-local model parameter vectors W switch to load and implement model parameter switching. Denote the model parameter switching as W g ,W h ,W m and W l The shared CNN network models at this time are CNN g ,CNN h ,CNN m and CNN l respectively, where CNN g is denoted as the global model, and CNN h ,CNN m and CNN l are denoted as local models;

[0296] ④ Use the shared CNN network model after model parameter switching to predict the sample to be diagnosed, and the multi-label prediction results of the corresponding global model and local model can be obtained respectively;

[0297] S56. Global-local model fusion mechanism: Considering that global information often fails to highlight local details, while local information is difficult to comprehensively and accurately grasp the overall fault characteristics, the present invention further realizes compound fault diagnosis through global-local two-stream fusion. It should be noted that most existing patents obtain the fusion prediction vector by weighting the prediction probability vector for fault diagnosis. However, this weighting method requires manual adjustment, has high requirements for the prior knowledge (previously accumulated experience) of the operator, and the model needs to be retrained after each adjustment, resulting in poor interpretability of the training process and often unable to achieve the expected accuracy. Therefore, the present invention improves the fusion mechanism of global-local information flow, specifically including:

[0298] ① Based on the model parameter adaptive switching mechanism in S5 above, use the global model and the local model to test the samples in the validation set respectively, and obtain the global information flow and the local information flow respectively. The present invention defines the final prediction vector g obtained by inputting the global sample into the global multi-label classification model CNN as the global information flow, and defines the head, middle and tail three local samples to input the corresponding local multi-label classification model CNN​h , CNN m and CNN l The final predicted vector obtained and is the local information flow:

[0299]

[0300]

[0301]

[0302]

[0303] ② By testing all samples in the validation set and inputting the above and information flow into the threshold function to calculate the accuracy of the shared CNN model in identifying various single fault types under different model parameters (the recognition accuracy of a single column corresponding to the probability prediction vector of different information flows), the average per-class accuracy vectors corresponding to the four information flows obtained are respectively:

[0304]

[0305]

[0306]

[0307]

[0308] Among them, represents the recognition accuracy of the multi-label classification global model for the first single fault (occlusion fault) on the validation set, and respectively represent the recognition accuracies of the three multi-label classification local models for the first single fault (occlusion fault) on the validation set;

[0309] ③ According to the test situation of the validation set, find the model that is most accurate in identifying various single fault types, and then fuse the corresponding information flows, that is, use some columns of the local model probability prediction vector to replace some columns of the global model Sigmoid probability prediction vector to obtain the final global-local dual-stream collaborative probability prediction vector, and further input it into the threshold function M to obtain the final fault diagnosis result of the fusion model:

[0310]

[0311]

[0312] Among them, i represents four single fault types (i.e., the i-th column in the information flow), A max The function of the function is to return the information flow probability prediction value corresponding to the maximum value in the same column of the four accuracy rate vectors in ②. For example, for the accuracy rate vector of the first column If the maximum value is Then

[0313]

[0314] By judging i = 1, 2, 3, 4 respectively, A max A total of 4 probability prediction values are returned to form The fusion prediction vector of the global-local two-stream collaborative model is obtained, that is, it is determined which prediction result of the corresponding column of the information flow is taken for the 1st to 4th columns of the final fusion model respectively, and the diagnosis result Result can be further calculated based on the threshold function M. Based on the global-local two-stream collaborative composite fault diagnosis method proposed by the present invention, the local model can be used to correct the recognition accuracy of the global model for some single fault categories, thereby effectively improving the diagnosis accuracy of the final fusion model for the photovoltaic module composite fault.

[0315] The local key point search algorithm in the present invention is designed based on the local first-order difference. Here, the first-order difference can be replaced by other methods that can reflect the curve change trend, such as the second-order difference, Lagrange mean value theorem, etc. For the selection of the flat points and key points on the curve, the rules can also be appropriately adjusted to obtain an approximate solution.

[0316] The method for selecting the local key region interval in the present invention is obtained through data distribution transformation and improved normal distribution confidence interval method. This part can also be replaced by the confidence intervals of other types of data distributions, such as Poisson distribution, τ distribution, etc.

[0317] The threshold function in the present invention is not limited to the constant value set in the text, and this value can be changed, which can be either a fixed value or a dynamically changing value.

[0318] The fusion method of the global information and local information in the present invention can adopt other model fusion methods to obtain an approximate solution.

[0319] In summary, the technical contributions of the present invention at least include:

[0320] (1) In view of the complex situation that multiple faults may occur simultaneously in photovoltaic modules, the present invention designs a new global-local two-stream collaborative photovoltaic composite fault diagnosis framework. The training and inference of the model are data-driven, and it can achieve end-to-end rapid diagnosis of photovoltaic module composite faults, solving the problems of low efficiency of the existing photovoltaic module composite fault diagnosis mechanism modeling technology and poor model generalization.

[0321] (2) From the perspective of the compound fault mechanism, the present invention regards the compound fault as the superposition of multiple known faults, transforms the compound fault diagnosis problem into the identification of multiple single faults, that is, transforms the original single-label multi-classification problem into a multi-label binary classification problem for solution, which can well retain the potential connections between various single faults and solve the problems of fewer fault types diagnosed by the prior art and low accuracy of compound fault diagnosis;

[0322] (3) Different from the prior art that only uses global information for photovoltaic fault diagnosis, the present invention designs a new local key point search algorithm, a local key area interval selection method and a local key area interpolation algorithm according to the characteristics of photovoltaic data to further obtain local information, and effectively improves the recognition accuracy of the model for photovoltaic module compound faults by synergistically using the global information and key local information of photovoltaic fault data;

[0323] (4) The present invention designs a new model parameter adaptive switching mechanism, enabling the global and local samples to share the CNN model training with the same structure, overcoming the deficiency of the prior art that multiple network models with different structures need to be trained separately, improving both the modeling efficiency and the robustness of the diagnosis model;

[0324] (5) The present invention designs a new global-local model fusion mechanism, which can directly correct the recognition accuracy of the global model for some single fault categories through the local model, solving the problem that the prior art needs to manually weight and fuse the prediction probability vectors based on prior knowledge or manual experience.

[0325] In the present invention, by regarding the original complete I-V characteristic curve as global information, and based on the characteristics of photovoltaic I-V data, a new local key point search algorithm, a local key region interval selection method, and a local key region interpolation algorithm are designed to mine the local information of the original complete I-V curve. It can not only well retain the potential connections between various single faults, but also ensure that the local information and the global information have the same data dimension, enabling the global model and the local model to be trained using a CNN with the same network structure, thereby saving the model construction cost and improving the modeling efficiency. Starting from the perspective of the composite fault mechanism, the present invention regards the composite fault as the superposition of multiple known faults, transforms the composite fault diagnosis problem into the identification of multiple single faults, that is, transforms the original single-label multi-classification problem into a multi-label binary classification problem for solution, and establishes a multi-label classification global model and a multi-label classification local model based on the shared CNN. In addition, the present invention proposes a new model parameter adaptive switching mechanism and a global-local dual-stream collaborative model fusion mechanism, which can directly replace part of the columns of the global model Sigmoid probability prediction vector with part of the columns of the local model Sigmoid probability prediction vector based on the verification set situation, without the problem of weighting. Thus, the local model is used to correct the recognition accuracy of the global model for some single fault categories, greatly improving the situation of misjudgment and missed judgment of the model, effectively improving the composite fault diagnosis accuracy of the model, and having important significance for on-site maintenance personnel to quickly and accurately carry out rapid maintenance after multiple faults occur simultaneously, and avoid huge power generation economic losses and major safety accidents caused by the faulty operation of photovoltaic modules.

[0326] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations according to the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. An intelligent composite fault diagnosis method for photovoltaic modules, characterized in that, It includes the following steps: Step 1, data preprocessing; Step 2, local key point search; It includes: Step 21, obtaining global samples; the complete I-V curve after the data preprocessing is denoted as the global sample, and the global sample is composed of 40 discrete data points (x i , y i ), and they are uniformly distributed on the curve; Step 22, local first-order difference calculation; define two adjacent sampling points (x m , y m ) and (x m+1 , y m+1 ) on the I-V line, and (x m+1 > x m ). The local first-order difference value g′(x m ) at (x m , y m ) is: Step 23, flat point screening. Define the sampling points with the absolute value of the local first-order difference less than α in the complete I-V curve as flat sampling points. Based on the 39 first-order difference values between adjacent two sampling points in the global samples calculated successively in step S22, screen out the set T of flat points on the I-V curve: T = find(abs|g′(x i ) < α|), i = 2,..., 39 Among them, the function of the find() function is to screen out the relative positions of the sampling points that meet the constraint conditions in the brackets, abs|| is the absolute value function, and α is a constant value given in advance; S24, local key point search; It includes: Define the flat sampling point whose absolute value of the local first-order difference from the adjacent left side is at least β times smaller as the first local key point, and the sampling point whose absolute value of the local first-order difference from the adjacent right side is at least γ times smaller as the second local key point, and screen out the set O of local key points on the I-V curve: O1 = find(β * abs|g′(x i )| ≤ abs|g′(x i -1)|), i = 2,..., 38 O2 = find(γ * abs|g′(x i )| ≤ abs|g′(x i+1 )|), i = 2,..., 38 O = O1 ∪ O2 Among them, O1 represents the first local key point subset, O2 represents the second local key point subset respectively, the symbol ∪ represents the union, and β and γ are constant values given in advance; Step 3, local key area interval selection; It includes: Step 31, composite fault sample grouped sampling. Divide the total number T of composite fault global samples when four single fault types of occlusion, aging, short circuit, and open circuit occur simultaneously in the training set into z groups on average, and each group uses the sampling without replacement method to obtain k random samples, that is Step 32, local key point marking. Respectively perform the local key point screening in step S24 on all samples in each group. After completion, the corresponding number of local key points will be marked on the I-V curve of each global sample; Step 33, local key point classification; It includes: Subdivide the local key points into local feature head marking points, local feature tail marking points, and local feature middle marking points. Among them, the local feature head marking point is the local key point with the most forward relative position; the local feature tail marking point is the local key point with the most backward relative position; the local feature middle marking points are the remaining marking points excluding the local feature head marking points and the local feature tail marking points; Step 34, data distribution random group classification; Including: classifying the data of the local feature head marker points into a local feature head random group The overall sample of the corresponding local feature head random group is denoted as X h , classifying the data of the local feature tail marker points into a local feature middle random group The overall sample of the corresponding local feature middle random group is denoted as X m , classifying the data of the local feature tail marker points into a local feature tail random group The overall sample of the corresponding local feature tail random group is denoted as X l ; Step 35, data distribution transformation; Including: Assume that the head random group sample or the middle random group sample or the tail random group sample is denoted as {X i}, the overall sample is denoted as X, the mean of the overall sample is denoted as μ, and the variance is denoted as σ 2 , the probability distribution function is denoted as Φ, i represents different random groups of the same type, i = 1, 2,..., z; Obviously, each group of samples {X i} satisfies independent and identically distributed, and where and represent the mean and variance of the overall sample respectively; Let: Meanwhile, let be 's characteristic function, and we can obtain: Among them, and are the first derivative and the second derivative of the characteristic function respectively. j is a complex number, so j 2 = -1; thus, the characteristic function of has the Taylor expansion at the origin as follows: Denote X z whose characteristic function is It can be obtained that: When the number of groups \(z\) is relatively large Approximately follows a normal distribution The corresponding expectations, variances, and probability distributions can be obtained as follows: Convert the local key point distributions of the three types of random groups, namely the head random group samples, the middle random group samples, and the tail random group samples, into an approximate normal distribution, that is Step 36, construct local key area intervals. That is, respectively determine the final head local key area, middle local key area, and tail local key area on the premise of ensuring at least 95% confidence; It includes: Step 361, for the random group of local feature headers According to the following formula The obtained local key region interval of the head is where i = 1, 2,..., z; Step 362, for the random group in the middle of the local features According to the following formula Get the interval of the central local key area as where i = 1, 2, ..., z; Step 363, for the random group at the tail of the local feature According to the following formula The obtained local key region interval of the tail is where i = 1, 2,..., z; Step 4, local key area interpolation reconstruction. Reconstruct the head local key area, the middle local key area, and the tail local key area according to the local key area interpolation method; The local key area interpolation method includes: Step 41, based on the global dimension, calculate the number of resampling points required between every two original sample points within the interval; Step 42: Obtain the function values and local first-order difference values of two adjacent interpolation points x m , x m+1 on the I-V curve, which are respectively: y i = g(x i ), t i = g'(x i ), (i = m, m + 1), Step S43, construct the local feature interpolation polynomial C(x) to make it satisfy: C(x i ) = y i , C′(x i ) = t i , (i = m, m + 1) Four pointer functions r1(x), r2(x), s1(x), and s2(x) are introduced to represent the two components C of the local feature interpolation function C(x) i , (i = 1, 2), and they are all polynomials of order less than or equal to two: C1(x) = y m r1(x) + t m s1(x) C2(x) = y m+1 r2(x) + t m+1 s2(x) β i (x j ) = 0, β′ i (x j ) = δ ij , (i = 1, 2), C(x i ) = y i , C′(x i ) = y′0, (i = m, m + 1), The local feature interpolation function C(x) is expressed as: C(x) = C1(x) + C2(x) x m+1 is a double zero of r1(x). The pointer function value and the local difference function value of r1(x) at x m+1 are the same and both are zero. Let: r1(x) = px 2 + px m+1 2 + qx 2 - 2px * x m+1 + qx m+1 2 - 2qx * x m+1 From r1(x m+1 ) = r1'(x m+1 ) = 0, we can obtain: Swap x m with x m+1 to get: Similarly, it can be obtained: The local interpolation formula between two points is as follows: C(x) = C1(x) + C2(x) = y m r1(x) + t m S1(x) + y m+1 r2(x) + t m+1 S2(x); Step 5: Construct a global-local dual-stream collaborative diagnosis framework; which includes: Step 51: Obtain global samples; Step 52: Obtain local samples to obtain initial head local samples, initial middle local samples, and initial tail local samples; Step 53: Label encoding and dataset construction; which includes: Encode the label of each fault type into a binary vector with a dimension of 1×4, corresponding to 4 single fault types respectively. The first column represents the occlusion fault, the second column represents the aging fault, the third column represents the short-circuit fault, and the fourth column represents the open-circuit fault; Encode a flag bit for the global samples, the initial head local samples, the initial middle local samples, and the initial tail local samples respectively; Step 54: Build a shared CNN network model; Step 55: Perform adaptive switching of model parameters; which includes: Step 551: Initialize a shared CNN network, and denote the network parameter vector of the global multi-label classification model as W g , denote the network parameter vector of the head local multi-label classification model as W h , the network parameter vector of the middle local multi-label classification model as W m , and the network parameter vector of the tail local multi-label classification model as W l , and combine them to construct the global-local model parameter vector W switch : W switch = [W g , W h , W m , W l ​ Step 552: Input the sample to be diagnosed into the shared CNN network and preferentially read the flag bit of the sample to identify which one of the global samples, or the initial head local samples, or the initial middle local samples, or the initial tail local samples the sample belongs to; Step 553: Based on the recognition result of the flag bit, the shared CNN network model adaptively selects the corresponding model parameter vector from the global-local model parameter vectors W switch to load and implement model parameter switching, and denote the model parameter switching as W g , W h , W m and W l . The shared CNN network models at W g , W h , W m and W l are respectively CNN g , which is denoted as the multi-label classification global model, and CNN h , CNN m and CNN l are respectively denoted as the multi-label classification model head local model, the multi-label classification middle local model, and the multi-label classification tail local model; Step 554: Use the shared CNN network model after model parameter switching to predict the sample to be diagnosed to obtain the multi-label prediction results of the corresponding multi-label classification global model, the multi-label classification head local model, the multi-label classification middle local model, and the multi-label classification tail local model respectively; Step 56: Perform a global-local model fusion mechanism to obtain a diagnosis result.

2. The intelligent photovoltaic module composite fault diagnosis method according to claim 1, wherein, The said Step 52 includes: Step 521: Localization; that is, perform the local key point search on the global sample, locate and mark the corresponding local key points on the global sample; Step 522: Screening; that is, perform the construction of the local key region interval on the globally sampled and labeled sample, and respectively screen out the head local key region interval, the middle local key region interval, and the tail local key region interval from the global sample to obtain the initial head local sample, the initial middle local sample, and the initial tail local sample; Step 523: Reconstruction; perform the local key region interpolation method on the initial head local sample, the initial middle local sample, and the initial tail local sample to ensure that the finally interpolated and reconstructed sample has the same data dimension as the global sample and can share the neural network model with the same structure for training.

3. The intelligent photovoltaic module composite fault diagnosis method according to claim 1, wherein,The said Step 56 includes: Step 561: Based on the model parameter adaptive switching mechanism, use the global model, the head local model, the middle local model, and the tail local model to test the samples in the validation set respectively, and obtain the global information flow, the head local information flow, the middle local information flow, and the tail local information flow respectively; define the global sample input to the global multi-label classification model CNN g The final predicted vector obtained is the global information flow. Define three local samples of the head, middle, and tail to input the corresponding local multi-label classification models CNN h , CNN m and CNN l The final predicted vectors obtained and are the local information flows, as follows.

4. The intelligent photovoltaic module composite fault diagnosis method according to claim 3, wherein, The said Step 56 further includes: Step 562: By testing all samples in the validation set and inputting the above and information flows into the threshold function calculate the accuracy of the shared CNN model in identifying various single fault types under different model parameters. The average per-class accuracy vectors corresponding to the four information flows obtained are as follows: Among them, represents the recognition accuracy of the multi-label classification global model for occlusion faults on the validation set, and respectively represent the recognition accuracy of the multi-label classification model head local model, the multi-label classification middle local model, and the multi-label classification tail local model for occlusion faults on the validation set.

5. The intelligent photovoltaic module composite fault diagnosis method according to claim 4, wherein, The said Step 56 further includes: Step 563: According to the test situation of the validation set, find the model that is most accurate in identifying various single fault types, then fuse the corresponding information flows to obtain the final global-local dual-stream collaborative probability prediction vector, and further input it into the threshold function Obtain the final fault diagnosis result of the fusion model: where i represents four single fault types, A max The function is to return the information flow probability prediction value corresponding to the maximum value in the same column of the four accuracy rate vectors in step 562; By making judgments for \(i = 1, 2, 3, 4\) respectively, A max A total of 4 probability prediction values are returned to form Obtain the fusion prediction vector of the global-local two-stream collaborative model, that is, determine which prediction result of the corresponding column of which information flow is taken for the 1st to 4th columns of the final fusion model respectively. Based on the threshold function Calculate to obtain the diagnostic result Result.

6. The intelligent photovoltaic module composite fault diagnosis method according to claim 5, wherein, In the shared CNN network model, in cooperation with the Sigmoid function and the threshold function The display expression of the photovoltaic composite fault diagnosis and classification result output is as follows: x = [x i , i = 1, 2, 3, 4 Among them, x represents the output vector of the last fully connected layer, which contains a total of 4 items. Inputting it into the Sigmoid function yields a prediction probability vector corresponding to the occurrence of 4 single fault types. represents the predicted probability of the occurrence of single fault type i, and τ represents the threshold; using the threshold function for to make a further discrimination. If the predicted probability is greater than or equal to τ, it is considered that this type of single fault has occurred. Finally, the final photovoltaic composite fault diagnosis classification result is obtained by synthesizing the discrimination results of 4 columns. The selected loss function is Binary Cross Entropy (BCE), and the corresponding mathematical expression is as follows: Among them, and represent the global model loss function and the local model loss function respectively, C represents the number of single fault types, K represents the total number of training samples, represents the true label of the single fault type j in the i-th sample, which is 1 if it occurs and 0 if it does not occur, and are output by the threshold function and represent the predicted labels of the global model and the local model for the single fault type j in the i-th sample respectively.

7. The intelligent photovoltaic module composite fault diagnosis method according to claim 1, wherein, The global-local model fusion is obtained by adopting other model fusion methods.

8. The intelligent photovoltaic module composite fault diagnosis method according to claim 1, wherein, The α is set to 10 -4 .

9. The intelligent photovoltaic module composite fault diagnosis method according to claim 1, wherein, The said β is set to 5, and the said γ is set to 3.

10. The intelligent photovoltaic module composite fault diagnosis method according to claim 6, wherein, The said τ is set to 0.5.

Images