JP2019022251A - Solar cell diagnosis method and solar cell diagnosis system - Google Patents

Abstract

To provide a method and system that includes a thermal image classification apparatus and a method with a high classification performance of f1-score 0.97 using a deep learning toward a diagnosis of the status of a bypass diode and a power generation circuit, in which a separation between the abnormal and normal states of the bypass diode that cannot be easily perceived can be performed.SOLUTION: A solar cell diagnosis method and system: photographs a solar cell surface as thermal imagery in a solar cell diagnosis method for diagnosing a failure of solar cell by photographing the solar cell surface as thermal imagery; separately detects both a heat generation state of a bypass circuit of a solar cell module or a whole or part of solar cell array and a heat generation state of a power generation circuit; detects the thermal distribution of these heat generation states; and determines an abnormal heat generation state expressed in the thermal images in combination with those states.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a solar cell diagnostic method and a solar cell diagnostic system for diagnosing a failure of a solar cell by thermal image analysis.

  In recent years, the increase in greenhouse gas emissions due to the use of petroleum resources has become a problem, and the use of natural renewable energy such as solar power generation and wind power generation has attracted attention, and the introduction of solar power generation systems is expanding .

  FIG. 10 shows a schematic structural diagram of a general photovoltaic power generation system 101 to be diagnosed by the solar cell diagnostic system of the present invention. The solar power generation system 101 generally includes a solar cell array 3, a power conditioner 6 that converts DC power generated by the solar cell array 3 into AC power, and a connection that connects the solar cell array 3 and the power conditioner 6. It consists of a box 7 and the like. The solar cell array 3 refers to the entire solar cell installed on a base B such as a house roof or garden. In the solar cell array 3, a string 30 that is a series connection body of a plurality of solar cell modules 300 is connected to a connection box, and a plurality of strings 30 are connected in parallel. The solar cell module 300 is generally called a solar panel and is a basic unit in construction. In the solar cell module 300, a plurality of solar cells 300a are connected in series and connected in parallel with one bypass diode 300b (BPD) to form a unit circuit called a cluster 300A, and a plurality of clusters 300A are connected in series. In a weather-resistant package. What is contained in this weather-resistant package is called a “solar cell module”. In the solar cell module 300, the entire circuit connecting the solar cells 300a and / or the solar cells 300a is referred to as a “power generation circuit”, and the entire circuit connecting the bypass diode 300b and / or the bypass diode 300b is referred to as a “bypass circuit”. The bypass circuit is stored in the junction box 300c. In the solar cell module 300, the P pole and the N pole of each solar cell 300a are connected by a strip-shaped conductor called an interconnector. The solar battery cell 300a is a functional minimum unit having a function of converting solar light energy into electrical energy.

  Here, in the present invention, the solar power generation system 101 can also include the wiring 8, the distribution board 9, the watt hour meter 10, and the like as a distribution system. The distribution board 9 and the watt hour meter 10 are used when the power conditioner 6 is connected to the AC power supply 11. Further, in the present invention, the solar battery includes the solar battery array 3, the string 30, the solar battery module 300, the solar battery cluster 300A, the solar battery cell 300a, etc., but is not particularly limited. Configurations can also be included. Solar cell 300a may also be referred to as a “cell” or “power generation element” (FIG. 10).

  The introduction of photovoltaic power generation is expanding, and accordingly, social needs for maintenance technology in solar cells are increasing. A deteriorated solar cell cannot sufficiently convert light energy into electric energy as compared with a normal solar cell, and thus exhibits a higher temperature than a solar cell in a normal state. The deterioration of the solar cell not only causes a decrease in the amount of power generation, but also causes a lack of safety of the power generation system, so periodic deterioration management is a very important issue.

  Development of a solar cell diagnostic system for diagnosing such a solar cell failure has been underway. There are various ways to divide the solar cell diagnostic system. For example, the solar cell diagnostic system is brought into contact with the solar cell to diagnose a failure, and the solar cell diagnostic system is not brought into contact with the solar cell. There is a non-contact type diagnosis system for diagnosing a failure. Some contact-type diagnostic systems, for example, connect a part of the solar cell diagnostic system directly to the solar cell and measure the voltage and / or current to diagnose the failure. Some batteries require a DC power supply connection. From a viewpoint different from the contact type and the non-contact type, there is a built-in type diagnostic system as a method for diagnosing a failure while a part of the solar cell diagnostic system is always incorporated in the solar cell.

  Further, when the solar cell diagnostic system is divided from the viewpoint of a more specific maintenance method, it can be classified into two methods, a method using current-voltage characteristics and a method using a thermal image. In the maintenance method using the current-voltage characteristic, the failure factor can be estimated from the analysis of the curve characteristic. However, it is necessary to connect the measuring device directly to the module for diagnosis, and since it corresponds to the contact type diagnosis system described above, it is necessary to separate the module in order to separate it from the power generation state at the time of diagnosis. There is a problem that it takes enormous time and cost to identify the system.

  The method using a thermal image is generally a non-contact type, but may be a contact type. The case where the method using a thermal image corresponds to a non-contact type is a method of photographing only with a camera. The case where the method using the thermal image corresponds to the contact type is a method that requires an external power source or the like to be brought into contact with the solar cell in addition to the camera for photographing. If it is a non-contact type, diagnosis can be performed without stopping power generation, and a heat generation point can be visually recognized, so that a faulty module can be easily identified. However, the thermal image of the solar cell depends on shooting conditions and the like. Even when there is a failure, there is a case where a heat generation pattern that can be easily distinguished is not displayed, and thus specialized knowledge on a correct measurement method is required. Moreover, even if it is the method using a thermal image, when it sees with the whole solar cell diagnostic system, it may correspond to a contact type, and in such a case, power generation is stopped and a diagnosis is required.

  For example, Patent Document 1 is a method for inspecting a solar cell array including a plurality of solar cell panels 1, a step of connecting a DC power source 6 to the solar cell array and energizing the solar cell array, and heat of the solar cell panel 1 of the solar cell array. The step of acquiring an image, the thermal image is multivalued, an area ratio of a plurality of levels is calculated, and the estimated value of the output power of the solar cell panel 1 is calculated using the output characteristic equation of the solar cell panel 1 with respect to the area ratio. And a step of determining whether or not the solar cell panel 1 should be replaced based on the estimated value after the time change of the estimated value becomes equal to or less than a predetermined value. The output characteristic equation includes a coefficient for each area ratio. There is disclosed a method for inspecting a solar cell array, wherein the coefficient and the constant are values obtained in advance by regression analysis.

  The solar cell array inspection method disclosed in Patent Document 1 is a solar cell diagnostic system based on thermal image analysis, and acquires a thermal image without contact between the solar cell array and the image capturing unit 2. However, since there is a process of connecting the DC power source 6 which is a part of the diagnostic system to the solar cell array and energizing it, it is a contact type diagnostic system, which is more laborious for diagnostic work compared to the non-contact type. It takes a long time and the diagnosis efficiency deteriorates.

  As described above, in general, in a contact-type diagnosis system, it is necessary to interrupt power generation for the operation of bringing the diagnosis system into contact with the solar cell, or the operation of sequentially changing the contact destination of the diagnosis system in units of strings, for example. Since it becomes necessary or the workers at that time are limited to qualified personnel, the efficiency of the diagnostic work is poor. In addition, in the built-in diagnostic system, it is necessary to keep the diagnostic system beside the solar power generation system or in a barn or the like in order to continuously monitor voltage, current, and the like. Therefore, not only installation space is taken, but installation cost also becomes high. Furthermore, the life of a diagnostic system is generally shorter than that of a solar cell, and the diagnostic system itself needs to be maintained. Therefore, there is an increasing expectation for a non-contact type diagnostic system that does not require interruption of power generation and does not need to be incorporated into the diagnostic system.

  By the way, when a failure occurs in a solar cell module due to abnormality or deterioration, the power generation capability of the solar cell module decreases. However, when the entire solar power generation system, solar cell array or solar cell string unit is viewed, the decrease in power generation capability is slight. In general, it has been difficult to notice a decrease in the power generation capacity until a predetermined value or less of a warning set in advance at an output location of a power conditioner where power generation output is measured. Therefore, even if the solar cell module is out of order, it cannot be noticed and may be used as it is for a long time. However, when a failure of the solar cell module occurs due to opening (disconnection), not only a decrease in power generation but also a high temperature or ignition due to arc discharge may occur. Due to this, there have also been cases of fires caused by long-term use of failed solar cell modules.

  Conventionally, when confirming a decrease in power generation capacity, the estimated power generation amount obtained by measuring the amount of solar radiation was compared with the actual power generation amount. It is difficult to specify whether there is a faulty solar cell module in the photovoltaic power generation system because it only understands the degree of power generation reduction in the solar cell array or string unit being measured. there were. Further, even if the user notices the presence of the failed solar cell module, it is difficult to specify the position where the failed solar cell module is installed. In order to specify the installation position of the failed solar cell module, for example, work such as checking the voltage of each solar cell module at the time of power generation is required, and time and cost are required. Also, it was dangerous to work at high places such as the roof of a house.

  Therefore, in order to solve the above-described problems, solar cell diagnostic systems for diagnosing a failure of the solar cell module have been studied in various places. For example, Patent Document 2 discloses a solar cell array that is generating electric power by measuring the temperature distribution of the surface of the solar cell module constituting the solar cell array for each aggregate of solar cell modules. A method for detecting a defect in a solar cell array, characterized by detecting a battery module, is described.

  However, Patent Document 2 does not have a function of correcting the apparent shape of the module imaged on the thermal image into an original rectangular shape after photographing. Therefore, the solar cell array arranged as a unit with respect to the outer wall of the building is photographed so that the solar cell module obtained as a thermal image has a rectangular shape. In other words, the solar cell module is originally rectangular, but depending on the shooting angle and distance, the shape of the solar cell module obtained as a thermal image becomes a trapezoid or the like and the diagnostic accuracy decreases, so that such a situation can be avoided, In the conventional method, the arrangement or angle of the solar cell array or the arrangement or angle of the thermal image capturing unit needs to be specified in advance. For example, as an arrangement configuration for photographing so that the shape of the solar cell module is rectangular, it is assumed that photographing is performed so that the panel surface of the solar cell array and the thermal image photographing unit face each other. Since the part is arranged on the ground, it is necessary to arrange the solar cell module substantially horizontally with respect to the ground plane. If the solar cell module is not arranged substantially horizontally with respect to the ground plane, it will be difficult to obtain a thermal image of a plurality of solar cell modules at the same time, or the thermal image capturing unit will be enlarged. . Moreover, even if aerial photography is performed using a drone, an error occurs in the shape of the solar cell module obtained as a thermal image, and the diagnostic accuracy decreases.

JP 2013-197173 A JP 2002-329879 A

  By the way, in the conventional non-contact solar cell diagnostic system and method based on thermal image analysis that can be diagnosed in units of solar cell modules, each component constituting the thermal image capturing unit and the thermal image analysis unit, for example, If the module extraction unit, determination unit, etc. are configured in an integrated manner, the state of the solar cell module, the presence or absence of the failure or deterioration of the solar cell module, and the position and cause thereof are accurately diagnosed, and the repair of the solar cell module Whether it is possible to make a diagnosis by distinguishing between a notification-free state that does not require replacement and a state that should be notified that repair or replacement of the solar cell module is required or recommended I couldn't say it was done.

  For example, in a conventional non-contact type solar cell diagnosis system based on thermal image analysis that can be diagnosed in units of solar cell modules, generally, a function that automatically adjusts the temperature scale in the thermal image capturing unit is used. A temperature scale that encompasses the temperature distribution of the entire thermal image is set. Therefore, when a background other than the solar cell is reflected in the thermal image, the maximum value of the temperature scale becomes larger than necessary or the minimum value becomes smaller than necessary depending on the temperature of the background. Therefore, the temperature resolution with respect to the solar cell is lowered, and the subsequent determination accuracy is lowered. Even if the temperature scale is manually set to a fixed value in the thermal image capturing unit, it is not always possible to set the temperature scale to encompass the entire solar cell part. The temperature resolution of the part was not always fully utilized.

  Further, for example, in a conventional non-contact type solar cell diagnosis system based on thermal image analysis that can be diagnosed in units of solar cell modules, the solar cell module is originally rectangular, but depending on the angle and distance of shooting, The solar cell module obtained as an image has a trapezoidal shape or the like. For this reason, in the method for contrasting with other thermal images, it is necessary to shoot a solar cell of a similar size from a similar shooting angle and a similar shooting distance. For example, as in Patent Document 2, it has been necessary to photograph the solar cell module obtained as a thermal image after it has been configured in advance to be rectangular. Conventionally, it has been necessary to shoot with a camera or lens with similar or equivalent specifications.

  In addition, for example, in a conventional non-contact solar cell diagnosis system based on thermal image analysis that can be diagnosed in units of solar cell modules, the accumulation of thermal image data for diagnosis, the type of diagnostic pattern, and the diagnostic accuracy are reduced. The system and method to be improved are not fully studied focusing on whether or not the solar cell module needs to be replaced, and the normal state or minor failure that does not require replacement of the solar cell module The condition was also diagnosed and / or notified.

  In general, the state of high temperature and heat generated in the thermal image of the solar cell includes, for example, (1) current concentration due to contamination of the solar cell, cell crack, etc., power consumption of the shadow area, or BPD in the junction box Cluster unit, which is caused by spot-like high-temperature part (mainly high temperature / heat generation in a cell unit, so-called hot spot), or (2) power circuit disconnection, solder failure, BPD short circuit, etc. (3) When all the clusters in the module become high temperature / heat generation parts, etc., high temperature / heat generation in module units, or ( 4) There is high temperature and heat generation in string units caused by connector disconnection / damage or cable damage.

  All of these appear on the thermal image as high temperature and heat generation, but these include normal or minor failure conditions that do not require replacement of the solar cell module, and are free of charge under the warranty of the solar cell module manufacturer. Some are not eligible for replacement. As described above, although the state of high temperature / heat generation that appears in the thermal image and the state that requires replacement of the solar cell module do not coincide with each other immediately and sufficiently, the conventional, non-contact type, In the solar cell diagnosis system based on thermal image analysis, the focus is on notifying the state of high temperature and heat generation that appears in the thermal image, and the state of high temperature and heat generation that appears in the thermal image is indeed the replacement of the solar cell module There was no focus on whether or not it is a state that needs to be notified (a state to be notified). Therefore, for example, a normal state or a minor failure state that does not require replacement of the solar cell module has been diagnosed and / or notified.

  In addition, for example, in a conventional non-contact type solar cell diagnosis system based on thermal image analysis that can be diagnosed in units of solar cell modules, it is possible to separate the bypass diode from an abnormal state that cannot be easily detected. And the provision of methods remained difficult.

  In general, among solar cell modules, the cause of failure is a bypass circuit and a power generation circuit. There is a short circuit and an open circuit in the bypass circuit failure, and a disconnection in the power generation circuit failure. FIG. 11 shows a representative thermal image for each state. When the bypass circuit is normal and the power generation circuit is disconnected (FIG. 11, state 04), the generated current flows through the diode, and the junction box in which the bypass diode is stored generates heat. Moreover, since the solar energy irradiated to the cluster is not converted into electric energy but becomes thermal energy, the entire cluster generates heat and exhibits a belt-like heat generation characteristic. In the present invention, this belt-like heat generation is called a hot stripe.

  When the bypass circuit is short-circuited, the bypass diode does not generate heat even if the power generation circuit is disconnected, or it generates heat weakly enough that it cannot be noticed, and only hot stripes are generated (state 05 in FIG. 11). Further, when the bypass circuit is short-circuited and the power generation circuit is normal, the generated current flows in the closed circuit of the power generation circuit, and only specific cells generate heat due to variations in cell characteristics (FIG. 11, state 02). ). When the bypass circuit is in an open state, heat is generated uniformly regardless of the state of the solar battery cell (FIG. 11, states 03 and 06). This is because when the solar cell is disconnected, power generation is not performed, and when it is normal, no current is originally passed through the bypass circuit. Note that state 03 is the same as normal (state 01) and has a uniform temperature distribution, and state 06 has a uniform temperature distribution as the hot stripe portion expands over the entire module surface. It is in a state.

  However, when the bypass circuit is in an open state, a clear high-temperature portion such as a hot spot or hot stripe does not occur in the thermal image of the solar cell. In addition, even when local heat generation occurs, the pattern changes depending on the position of the heat generation and the temperature scale determined at the time of thermal image generation, and a high level of expertise is required to make a specific diagnosis. Although a solar battery maintenance system using a drone equipped with a thermal imaging camera is available on the market, it has only been able to detect clear high-temperature parts such as hot spots and hot stripes, and has not yet reached the point of providing fault diagnosis. . That is, it has still been difficult to provide an apparatus and a method that can separate the abnormality and normal state of the bypass diode that cannot be easily detected.

  Therefore, the object of the present invention is to solve the above-mentioned problems, diagnose the state of the solar cell module, the presence or absence of the failure or deterioration of the solar cell module, and the cause, and the notification-free state that does not require replacement of the solar cell module. Another object of the present invention is to provide a solar cell diagnostic method and a solar cell diagnostic system that perform diagnosis by distinguishing from a state to be notified that requires replacement of a solar cell module. Furthermore, it is an object of the present invention to provide a thermal image classification apparatus and method using deep learning for diagnosis of the state of a bypass diode and a power generation circuit that are main abnormal parts. In other words, an object of the present invention is to provide a solar cell diagnostic method and a solar cell diagnostic system capable of separating an abnormal state and a normal state of a bypass diode that cannot be easily detected.

The solar cell diagnostic method of the present invention is a solar cell diagnostic method for diagnosing a failure of a solar cell by photographing the solar cell surface as a thermal image. The solar cell module or solar cell array is obtained by photographing the solar cell surface as a thermal image. The heat generation state of all or part of the bypass circuit and the heat generation state of the power generation circuit are detected separately, and the temperature distribution of these heat generation states is detected, and these are combined to feature the abnormal heat generation state that appears in the thermal image It is characterized by determining as.
Moreover, the solar cell diagnostic method of the present invention is a solar cell diagnostic method for photographing a solar cell surface as a thermal image and diagnosing a failure of the solar cell. A feature amount extracting step for extracting a feature amount from the heat generation state of the entire or a part of the bypass circuit of the battery module or the solar cell array and the heat generation state of the power generation circuit; and the abnormal heat generation state appearing in the photographed thermal image and the thermal image. And a determination step of determining similarity based on the feature amount.
Further, the solar cell diagnostic system of the present invention is a solar cell diagnostic system based on thermal image analysis obtained by photographing the surface of the solar cell. The solar cell diagnostic system is a heat generation state of the entire solar cell module or the solar cell array or a partial bypass circuit. And a heat generation state of the power generation circuit separately, comprising a thermal image capturing unit that captures the solar cell surface as a thermal image, and a thermal image diagnostic unit for diagnosing the captured thermal image, The image diagnostic unit includes a determination unit that determines a thermal image by dividing a diagnostic pattern based on a state of a bypass circuit and a state of a power generation circuit.
In the solar cell diagnostic system of the present invention, the determination unit divides the bypass circuit state into three types of normal, short circuit, and open, and divides the state of the power generation circuit into two types of normal and open. The diagnosis pattern is divided into six types based on the combination of the above state and the heat generation / high temperature state of the power generation circuit, and feature amounts are extracted based on the distribution and shape of the heat generation source in each failure state.

  When the state of high temperature and heat generation appearing in the thermal image is caused by the above-mentioned (2) band-like high temperature and heat generation (hot stripe) in cluster units caused by disconnection of the power generation circuit, solder failure, BPD short circuit, etc. This is a case where, for example, in one solar cell module composed of three clusters, one or more clusters do not generate power, and the output of the solar cell module decreases to 2/3 or less. Modules are generally subject to free replacement under the warranty of the solar cell module manufacturer. To describe the phenomenon of hot stripes from the principle of heat generation, in one solar cell module, the cluster becomes defective in conduction due to the disconnection of the cluster, or the power is not generated due to the short circuit of the bypass circuit, and the light is received in order to change into the electric energy. It is caused by the fact that the light energy of the sun changes mainly to thermal energy instead of electric energy.

  Therefore, in the present invention, by utilizing the correlation between the high temperature / heat generation state appearing in the thermal image due to the hot stripes and the heat generation / high temperature state requiring replacement of the solar cell module to be notified, A focused solar cell diagnostic system was developed.

According to the present invention, diagnosis is performed by dividing a diagnosis pattern from the heat generation / high temperature state of the bypass circuit and the heat generation / high temperature state of the power generation circuit, so that the solar cell diagnosis by the thermal image analysis of the solar cell surface is taken. In the system and / or method, it is possible to determine whether the state is a high temperature / heat generation state appearing in the thermal image due to hot stripes, and a normal state or minor failure state that does not require replacement of the solar cell module And a state that requires replacement of the solar cell module, and notification of a normal state or a minor failure state that does not require replacement of the solar cell module can be reduced.
According to the invention, a feature amount is extracted from the heat generation state of the bypass circuit and the heat generation state of the power generation circuit, and the similarity between the captured thermal image and the abnormal heat generation state appearing in the thermal image is determined based on the feature amount. Providing a device and a method that can separate abnormalities of a bypass diode from normal states that cannot be easily detected in the past because feature values are extracted based on the distribution and shape of heat sources in each failure state. Is possible.

The solar cell diagnosis method of the present invention displays, as a characteristic of the abnormal heat generation state, a point-like abnormal heat generation state called a hot spot and / or a linear abnormal heat generation state called a hot stripe appearing in a strip shape together with a temperature distribution. Or by extracting the feature value from the heat generation state of the bypass circuit in the solar cell module and the heat generation state of the power generation circuit by an analysis algorithm and displaying the identification result by the analysis algorithm as an abnormal heat generation state for determination. Features.
The solar cell diagnosis method of the present invention is characterized in that the abnormal heat generation state is characterized by the fact that the temperature distribution of the heat generation state is equal to or higher than a predetermined temperature, and a linear abnormal heat generation state called a hot stripe appearing in a strip shape is displayed. The circuit state is determined as either a short circuit or an open circuit, or as a short circuit or an open circuit circuit.

  According to the solar cell diagnostic system of the present invention, in the solar cell diagnostic system based on the thermal image analysis in which the surface of the solar cell is photographed, the diagnostic accuracy of the presence or absence of hot stripes in the solar cell module is improved. Therefore, the accuracy of distinguishing between a normal state or a minor failure state that does not require replacement of the solar cell module and a heat generation / high temperature state that requires replacement of the solar cell module is improved.

  In the solar cell diagnostic method of the present invention, the similarity of the abnormal heat generation state of the determination unit (determination step) and / or the distortion correction of the module extraction unit is detected using a convolutional neural network that is deep learning of artificial intelligence. It is characterized by that.

According to the present invention, it is possible to separate an abnormality (heat generation / temperature) of a bypass diode from a normal state, which has not been easily detected in the past.
Further, according to the present invention, for example, conventional image classification problems such as an increase in similarity due to affine transformation, a problem of feature amount extraction, and a problem of position dependency can be solved.

In the present invention, the thermal image diagnosis unit includes a module extraction unit that automatically or manually extracts an image corresponding to the solar cell module from the captured thermal image and corrects the distortion. A predetermined outermost contour point serving as an outer contour is calculated, and distortion captured by the thermal image capturing unit is corrected.
In the present invention, the module extraction unit detects vertical lines and horizontal lines that are the outermost contour of the solar cell module using a straight line detection method of image processing, calculates an intersection of the vertical lines and the horizontal lines, and calculates the outermost contour. Four points are automatically calculated.
As a straight line detection method for image processing, for example, Hough conversion is used. Hough transform (Hough transform) is one of feature extraction methods used in digital image processing. The thermal image diagnosis unit includes a module extraction unit that extracts and distorts an image corresponding to the solar cell module or the solar cell array from the captured thermal image, and the module extraction unit includes the solar cell module. A predetermined outermost point that is the outermost contour is calculated, and distortion captured by the thermal image diagnostic unit is corrected.

According to the present invention, in the solar cell diagnostic system and / or method based on thermal image analysis in which the surface of the solar cell is photographed, the shape of the solar cell module obtained as a thermal image does not become a trapezoid or the like depending on the photographing angle and distance. The solar cell module can be extracted as a rectangle which is the shape of the original solar cell module. For this reason, even with a method for comparing with other thermal images, it is not necessary to shoot a solar cell of a similar size from a similar shooting angle and a similar shooting distance.
Further, according to the straight line detection method by the module extraction unit, since the distance and angle may be arbitrary, it is possible to make a diagnosis based on a thermal image taken by a drone, and the trapezoidal image is taken at an inclination angle. However, it can be determined by extracting feature values from rectangular data.
In addition, according to the present invention, since it is a device capable of diagnosis in units of solar cell modules, diagnosis in units of solar cell strings (or units of solar cell arrays) that are larger units than solar cell modules is also possible. By the diagnosis in units of solar cell strings, it is possible to diagnose disconnection of the string (fuse blown, connector missing, etc.).

In the solar cell diagnostic method of the present invention, the temperature distribution of the heat generation state is automatically adjusted within a range in which the temperature distribution of the whole or a part of the solar cell module or solar cell array extracted by photographing is comprehensively included. Or, it is set manually to detect an abnormal heat generation state.
In the solar cell diagnostic system of the present invention, the thermal image diagnostic unit automatically or manually sets the temperature scale within a range in which the temperature distribution of the whole or a part of the solar cell module or the solar cell array is comprehensively included. The thermal image capturing unit includes a setting unit, and the thermal image capturing unit captures a thermal image again based on the set temperature scale.

  According to the present invention, the background other than the solar cell is not reflected in the extracted thermal image, and the maximum value of the temperature scale is increased more than necessary or the minimum value is more than necessary depending on the background temperature. It doesn't get smaller. That is, the temperature scale including the entire solar cell portion can be set, and the temperature resolution of the thermal image capturing unit can be sufficiently utilized for the solar cell. Therefore, the temperature resolution with respect to the solar cell is improved, and the subsequent determination accuracy is improved.

  According to the present invention, in a non-contact type solar cell diagnostic system and method based on thermal image analysis that can be diagnosed in units of solar cell modules, the above-described problems are solved, and the solar cell module is in a heat generation / high temperature state. , Or the presence or absence of solar cell module failure or deterioration and the cause of the diagnosis, no need to replace the solar cell module, notification-free heat generation, high temperature, and heat generation to be notified that requires solar cell module replacement It is possible to provide a solar cell diagnostic system and method for making a diagnosis by distinguishing from a high temperature state. Furthermore, according to the present invention, it is possible to provide a thermal image classification apparatus and method using deep learning for diagnosis of the state of a bypass diode as a main abnormal part and a power generation circuit. That is, it is possible to provide an apparatus and a method capable of separating an abnormal state and a normal state of a bypass diode that cannot be easily detected. Furthermore, the present invention aims to provide an apparatus and method having high classification performance, and the verification based on 1400 thermal images collected from six solar cell modules over a half year makes the apparatus and method of the present invention f1-score. It has a classification performance of 0.97. Furthermore, according to the straight line detection method, the outermost contour point can be extracted by the module extraction unit, and even if the trapezoidal image is captured by the inclination angle, the feature amount can be extracted and determined by rectangular data.

It is a figure which shows the example of installation of the solar cell diagnostic system of 1st Embodiment. It is a block diagram which shows one structural example of the solar cell diagnostic system of the said embodiment. It is a figure explaining the diagnostic screen of the display part of the said embodiment. It is a figure explaining the diagnosis history screen of the display part of the said embodiment. It is a block diagram which shows the example of 1 structure of the solar cell diagnostic system of 2nd Embodiment. It is a thermal image photograph which shows the deformation | transformation of the thermal image by the imaging angle of 3rd Embodiment. It is an image which shows the example of Data Augmentation of the said embodiment. It is a structure figure of the convolution neural network of the said embodiment. It is a figure which shows the extracted 1024-dimensional feature-value of the said embodiment. It is a graph which shows Train Loss and Valid Loss for every epoxy of the said embodiment. It is a simulation result using the test set of the said embodiment. It is an image which shows the convolution layer load value matrix (Left: 1st layer, Right: 2nd layer) of the said embodiment. It is a figure of the solar energy power generation system for demonstrating this invention. It is a thermal image which shows the typical failure of the bypass circuit or power generation circuit of a photovoltaic cell.

  Specific embodiments to which the present invention is applied will be described in detail with reference to the drawings.

(First embodiment)
FIG. 1 is a diagram illustrating an installation example of the solar cell diagnostic system 100 according to the first embodiment to which the present invention is applied. The solar cell diagnostic system 100 includes a thermal image photographing unit 1 and a thermal image diagnostic unit 2. The solar cell array 3 to be imaged by the thermal image capturing unit 1 is installed on the base B.

  FIG. 10 shows a schematic structural diagram of a general photovoltaic power generation system 101 that is a target of diagnosis by the solar cell diagnostic system 100 of the present invention. The solar power generation system 101 generally includes a solar cell array 3, a power conditioner 6 that converts DC power generated by the solar cell array 3 into AC power, and a connection that connects the solar cell array 3 and the power conditioner 6. It consists of a box 7 and the like. The solar cell array 3 refers to the entire solar cell installed on a base B such as a house roof or garden. In the solar cell array 3, a string 30 that is a series connection body of a plurality of solar cell modules 300 is connected to a connection box, and a plurality of strings 30 are connected in parallel. The solar cell module 300 is generally called a solar panel and is a basic unit in construction. In the solar cell module 300, a plurality of solar cells 300a are connected in series and connected in parallel with one bypass diode 300b (BPD) to form a unit circuit called a cluster 300A, and a plurality of clusters 300A are connected in series. In a weather-resistant package. What is contained in this weather-resistant package is called a “solar cell module”. In the solar cell module 300, the entire circuit connecting the solar cells 300a and / or the solar cells 300a is referred to as a “power generation circuit”, and the entire circuit connecting the bypass diode 300b and / or the bypass diode 300b is referred to as a “bypass circuit”. The bypass circuit is stored in the junction box 300c. In the solar cell module 300, the P pole and the N pole of each solar cell 300a are connected by a strip-shaped conductor called an interconnector. The solar battery cell 300a is a functional minimum unit having a function of converting solar light energy into electrical energy.

  Here, in the present invention, the photovoltaic power generation system 100 can also include a wiring 8, a distribution board 9, a watt hour meter 10 and the like as a distribution system. The distribution board 9 and the watt hour meter 10 are used when the power conditioner 6 is connected to the AC power supply 11. In the present invention, the solar cell includes the solar cell array 3, the string 30, the solar cell module 300, the solar cell cluster 300 </ b> A, the solar cell 300 a, and the like. Configurations can also be included. Solar cell 300a may also be referred to as a “cell” or “power generation element” (FIG. 10).

  The thermal image capturing unit 1 is used for capturing a thermal image of the surface temperature of the solar cell array 3 and sending the captured thermal image to the thermal image diagnostic unit 2. The thermal image data may be transmitted in a wired manner by the cable 4, or may be transmitted by a recording device or a wireless method. The thermal image capturing unit 1 is a device that can perform thermal image capturing (infrared image), such as an infrared camera (infrared thermography, etc.). Shooting may be performed by hand, or fixed-point shooting may be performed by adjusting the position and angle by the camera mount 5. Further, the thermal image capturing unit 1 may be mounted on the drone to perform aerial photography. In this case, for example, a wide area can be photographed at once so that the entire solar cell array 3 can be accommodated, and photographing is performed in close proximity. You can also

  The angle at which the solar cell array 3 is imaged by the thermal image capturing unit 1 is defined as the imaging angle α, and the imaging angle α = 0 ° in the perpendicular (normal) n direction to the surface of the solar cell array 3. α is preferably within 50 °.

  In general, since the emissivity decreases when the imaging angle α is 50 ° or more, in FIG. 1, the imaging angle α becomes shallower in the portion close to the thermal image capturing unit 1 in the solar cell array 3 and the temperature is relatively high. It is easy to measure, and in a distant part, since a photographing angle becomes deep and emissivity falls, it is easy to measure temperature low. Moreover, the temperature of the object which exists around the solar cell array 3 may be reflected in a thermal image by the influence of the specular reflection by the glass which covers the solar cell array 3 surface. By making the imaging angle α within 50 °, the diagnostic accuracy of the present invention is enhanced.

Optimal conditions for thermal image capturing are solar radiation amount: 200 W / m ² or more, weather: clear or cloudy, shooting angle α: within 50 °, temperature scale: 10 ° C. to 20 ° C., time zone: 9:00 to 15:00 is there. The temperature scale refers to the temperature range of the thermal image. By setting the temperature scale to 10 ° C. to 20 ° C., it is easy to check the heat generation location. However, since the appropriate condition of the temperature scale is affected by the ambient temperature, it is more preferably adjusted according to the ambient temperature. For example, the temperature scale is obtained by providing the temperature scale setting unit 28 in the solar cell diagnostic system 100 of the present invention, or the temperature distribution range of the entire solar cell extracted, or a desired part of the solar cell, for example, the temperature of the solar cell module It is preferable that the distribution range is automatically set within a comprehensive range (FIG. 3).

  FIG. 2A is a block diagram illustrating a configuration example of the solar cell diagnostic system 100. The thermal image diagnosis unit 2 includes a display unit 21, an input unit 22, a module extraction unit 23, a determination unit 24, a feedback unit 25, a thermal image storage unit 26, and a notification unit 27, which are software (image control means). Controlled by.

  2B and 2C are diagrams illustrating the display unit 21. FIG. The display unit 21 selectively displays the diagnosis screen 21a (FIG. 2B) or the diagnosis history screen 21b (FIG. 2C) when either the diagnosis button a1 or the diagnosis history button a2 is selected by the operator. That is, both the diagnostic screen 21a and the diagnostic history screen 21b have a diagnostic button a1 and a diagnostic history button a2, and when the diagnostic button a1 is selected, the diagnostic screen 21a is displayed and the diagnostic history button a2 is selected. As a result, the diagnosis history screen 21b is displayed.

  FIG. 2B is a diagram showing a diagnostic screen 21 a of the display unit 21. The diagnosis screen 21a includes a thermal image display field a3, a module number input field a4, an operation comment field a5, a diagnosis start button a6, and an input clear button a7. The captured thermal image is displayed in the thermal image display field a3. In the module number input field a4, the matrix number of the solar cell modules 300 is input. The operation comment field a5 is used by an operator to leave a comment. The diagnosis start button a6 is used for starting diagnosis. The input clear button a7 is used to clear the contents input on the diagnostic screen 21a. The thermal image displayed in the thermal image display field a3 displays the surrounding environment in addition to the solar cell array 3 or a part of the thermal image.

FIG. 2C is a diagram showing a diagnosis history screen 21 b of the display unit 21. On the diagnosis history screen 21b, a diagnosis image list b1, a cut image b2, thermal image information b3, a diagnosis result b4, and a similarity b5 are displayed. The diagnostic image list b1 displays a list of thermal images taken so far in a list format. For one thermal image selected from the diagnostic image list b1, a cut image b2, thermal image information b3, and a diagnostic result b4 are displayed. The cut image b2 displays a thermal image cut by the module extraction unit 23 in units of 300 solar cell modules. In FIG. 2C, two solar cell modules 300 are displayed as the cutout image b2. The thermal image information b3 is displayed when there is a comment input on the diagnosis screen 21a and ELIF information of the image file of the diagnosis source. The diagnosis result b4 includes the panel number (solar cell module number) b6 of the captured thermal image, the similarity b5 between the captured thermal image and the diagnostic pattern, and the determination result b7 in a list format for each solar cell module 300. indicate. The panel number (solar cell module number) b6 is a number assigned to each cut out solar cell module 300, and the diagnosis result b4 has a similarity b5 for each solar cell module 300 (panel number b6). And the determination result b7 is displayed, and the diagnosis pattern having the highest similarity in the similarity b5 is displayed in the determination result b7. There are six types of diagnosis patterns, state 01 to state 06 (details of each diagnosis pattern will be described later), and the diagnosis result b4 displays the similarity b5 to each diagnosis pattern. The value displayed in the similarity 1 column is a numerical value indicating the possibility of the state 01 (similarity of the state 01), and the value displayed in the similarity 2 column indicates the possibility of the state 02. The value displayed in the similarity 3 column is a numerical value indicating the possibility of the state 03 (similarity in the state 03), and is displayed in the similarity 4 column. The value to be displayed is a numerical value indicating the possibility of the state 04 (similarity of the state 04), and the value displayed in the similarity 5 column is a numerical value indicating the possibility of the state 05 (similarity of the state 05) The value displayed in the similarity 6 column is a numerical value (similarity in state 06) indicating the possibility of being in state 06. In the memo field b8, which of the states 01 to 06 corresponds to each of the similarities 1 to 6 is displayed so that it can be easily understood.
For example, in FIG. 2C, two solar cell modules 300 are cut out, and numbers are assigned as panel number 1 and panel number 2 for each solar cell module 300. Regarding the solar cell module 300 with the panel number 1, the similarity b5 column includes a value of similarity 1 (similarity in state 01), a value of similarity 2 (similarity in state 02), and a value of similarity 3 (Similarity of state 03), value of similarity 4 (similarity of state 04), value of similarity 5 (similarity of state 05), value of similarity 6 (similarity of state 06) are displayed. . Among the similarities 1 to 6 (similarities in states 01 to 06), the state having the highest numerical value (diagnostic pattern) is displayed as the determination result b7. The same applies to panel number 2. For the numbers indicating the possibility of the states 01 to 06, the similarity is calculated by using an artificial intelligence described later (third embodiment).

The input unit 22 is a mouse or a keyboard, for example, and is operated by a user and used for selecting a thermal image, inputting the number of solar cell modules 300, and the like (FIGS. 2A and 2B). The thermal image to be diagnosed is displayed on the diagnostic screen 21 a by being designated by the input unit 22. Preferably, it is specified and displayed by drag and drop or click by mouse operation (FIG. 2B).
Based on the displayed thermal image, the range of coordinates of the four outermost points (outermost points m1, m2, m3, m4) of the solar cell module 300 is designated by a mouse operation. The outermost point is a solar cell module that can be cut out linearly when there are one or more solar cell modules 300 that can be cut out as a single solar cell module 300 in the displayed thermal image. It refers to the four vertices of 300. For example, in FIG. 2B, there are two solar cell modules 300 that can be cut out as a single solar cell module 300 in the thermal image displayed in the thermal image display field a3. The solar cell module 300 has a straight line connecting the outermost point m1 and the outermost point m2 (a horizontal line in the cut image) and a straight line connecting the outermost point m1 and the outermost point m3 (a vertical line in the cut image). ), The straight line connecting the outermost contour point m2 and the outermost contour point m4 (vertical line in the cut image), and the straight line connecting the outermost contour point m3 and the outermost contour point m4 (horizontal line in the cut image) Can be cut out linearly as a trapezoidal shape. The two solar cell modules 300 cut out linearly as this trapezoidal shape are each extracted as one solar cell module 300 by the module extraction unit 23 and corrected to a rectangular shape, so that on the diagnosis history screen 21b. Two cutout images b2 are displayed in a rectangular shape (FIG. 2C).
In addition, it is preferable to select the outermost points (m1, m2, m3, and m4 in the case of FIG. 2B), but it is necessary to select the outermost points at the time of extraction of the solar cell module 300 described above. Instead, for example, four vertices of one solar cell module 300 may be selected and only one solar cell module 300 may be cut out for diagnosis. For example, in FIG. 2B, by selecting four vertices of the outermost contour point m1, the outermost contour point m2, the vertex m5, and the vertex m6, the range of coordinates of the four points is specified, and one solar cell module 300 Module extraction and diagnosis can be performed. Similarly, module extraction and diagnosis of one solar cell module 300 can be performed by specifying a range of coordinates of four points by the four vertices of the outermost contour point m3, the outermost contour point m4, the vertex m5, and the vertex m6. it can.
In the module number input field a4, the number of cut-out solar cell modules 300 is input as a matrix number (FIG. 2B). In the operation comment field a5, comments such as special notes are input as necessary. The diagnosis start button a6 is used to input diagnosis start.

The module extraction unit 23 automatically extracts the solar cell modules 300 based on the coordinates of the outermost four points of the solar cell module 300 specified by the input unit 22 and the number of the solar cell modules 300, for example, a diagnosis such as a rectangle (FIG. 2A).
The thermal image diagnosis unit 2 includes a module extraction unit 23 that automatically or manually extracts and distorts an image corresponding to the solar cell module 300 or the solar cell array from the captured thermal image. Are extracted and distortion corrected automatically or manually. For example, the solar cell module is rectangular, but depending on the shooting angle and distance, the shape of the solar cell module obtained as a thermal image becomes a trapezoid or the like and the diagnostic accuracy is lowered. In this method, the arrangement or angle of the solar cell array, or the arrangement or angle of the thermal image capturing unit is defined in advance, and the solar cell module is photographed as an arrangement configuration in which the shape of the solar cell module is rectangular.

  The module extraction unit 23 may extract a plurality of solar cell modules 300. In this case, a method of extracting and sequentially diagnosing a plurality of solar cell modules 300 (for example, string 30), or extracting a plurality of solar cell modules 300 and uniting a plurality of solar cell modules 300 (for example, string 30). A method of diagnosis by using the

  The image portion corresponding to the solar cell (for example, the solar cell array 3, the solar cell string 30, the solar cell module 300, and / or the solar cell 300a) projected as a thermal image is an imaging angle of the thermal image capturing unit 1, The size and shape change depending on the installation angle of the solar cell and the distance between the thermal image capturing unit 1 and the solar cell. Therefore, the image portion corresponding to the solar cell module 300 projected as a thermal image has a different size and shape each time, but the module extraction unit 23 extracts the image portion corresponding to the solar cell module 300 from the thermal image, for example, a rectangular shape. By automatically correcting to an image suitable for analysis such as the above, it is possible to compare with other solar cell modules 300 having different sizes and shapes at the time of thermal image capturing. Thereby, the accurate classification | category of the diagnostic pattern of the solar cell module 300 is attained.

  Table 1 is a table | surface explaining the diagnostic pattern (patterning determination) of the solar cell diagnostic system 100 of this embodiment. Table 2 shows an example of symptoms and causes of high temperature / fever appearing in the thermal image used for the patterning determination.

  The determination unit 24 classifies the state of the solar cell module 300 into two types according to the state of the bypass circuit in the solar cell module 300 and the state of the power generation circuit in the solar cell module 300 (FIG. 2A). ). The determination unit 24 classifies the state of the bypass circuit in the solar cell module 300 into normal, short-circuited, and open, and classifies the diagnosis pattern into three types. The state of the power generation circuit in the solar cell module 300 is normal, The diagnosis pattern is classified into two types by dividing into open and open. The determination unit 24 combines the three types of diagnostic patterns of the bypass circuit and the two types of diagnostic patterns of the power generation circuit to classify the diagnostic patterns of the solar cell module 300 into a total of six types. There may be six or more types of diagnostic patterns. By adding diagnostic patterns, it is possible to diagnose solar cell states other than the six patterns.

  In the present invention, the “bypass circuit” refers to a bypass circuit in the solar cell module 300. For example, when a failure or abnormality occurs in a specific cluster 300A, no current flows through each of the solar cells 300a constituting the cluster 300A, and the current bypasses one bypass diode 300b connected to the cluster 300A. May refer to a circuit through such a bypass diode 300b. Of the states of the bypass circuit, “short circuit” refers to a short circuit failure of the bypass circuit. Among the states of the bypass circuit, “open” indicates an open (disconnection) failure. Of the states of the bypass circuit, “normal” refers to a normal state in which a short circuit failure and / or an open failure has not occurred. The bypass circuit works so as to flow current by bypassing the power generation circuit when a problem occurs in the power generation circuit. For example, when a shadow is applied to a certain cluster 300A and power generation becomes impossible, a current flows through the cluster 300A which is shaded by passing a current through the bypass circuit, and energy loss can be avoided.

  In the present invention, the “power generation circuit” refers to a power generation circuit in the solar cell module 300. For example, in a normal case where no failure or abnormality has occurred in a specific cluster 300A, a current flows through each cell constituting the cluster 300A, but this indicates a circuit through each such solar cell 300a. Among the heat generation / high temperature states of the power generation circuit, “open” refers to an open (disconnection) failure, for example, due to a state where the power generation circuit is physically disconnected or a defect in the material constituting the power generation circuit. In other words, it is a state in which the power is cut off, and the state where the power cannot be supplied or the power supply cannot be confirmed. Among the heat generation / high temperature states of the power generation circuit, “normal” refers to normal in which an open circuit failure has not occurred.

  The determination unit 24 performs feature extraction so that a total of six types of diagnosis patterns can be correctly classified and determined based on the thermal images of the solar cells extracted and corrected and stored in the thermal image storage unit 26. Optimization is performed manually and / or automatically. The determination unit 24 uses the optimized feature amount to calculate a similarity indicating how similar the thermal image to be diagnosed and each diagnosis pattern are. The calculated similarity with each diagnosis pattern is output to the notification unit 27 and the feedback unit 25. By calculating the degree of similarity with each diagnostic pattern, the state inside the solar cell module 300 can be classified into six patterns.

  Here, the “feature amount” refers to a numerical value or mathematical expression of the “feature” possessed by the data. “Feature” means, for example, if the data is an image, an element that may be necessary to determine what is reflected in the image, such as the size, shape, arrangement, vertex position, color, and brightness of the subject. All points. This feature extraction is called “feature extraction”. In other words, “feature extraction” refers to extracting features that are considered to be effective in correctly judging and classifying from atypical data such as images and sentences. Point to. In addition, the “similarity” can be included in the “feature amount”.

  In the present invention, from the heat generation state of the bypass circuit in the solar cell module 300 and the heat generation state of the power generation circuit, a point-like abnormal heat generation state called a hot spot and a linear abnormal heat generation state called a hot stripe appearing in a strip shape are characterized. As feature extraction. Further, the heat generation state of the bypass circuit in the solar cell module 300 and the heat generation state of the power generation circuit may be characterized, and feature extraction is performed by calculating and classifying the feature amount using an analysis algorithm.

  The feature extraction method is not particularly limited, but may be performed by a human using a conventional machine learning method or automatically by deep learning. When humans design feature extraction methods using conventional machine learning techniques, for example, feature quantities such as SIFT feature quantities and SURF feature quantities that are conventionally known are used as feature quantities, and a classification algorithm based on the feature quantities It is possible to classify them over the period. In addition, in the method of automatic feature extraction by deep learning, feature extraction is built into the algorithm, and the selection of features to be extracted itself, such as weighting and feature optimization, is learned by the machine. Data classification can be determined without setting.

  The feedback unit 25 is provided for a feedback process in which the thermal image data of the diagnosis result is used again as learning data, and improves diagnosis accuracy by increasing the number of accumulated learning data (FIG. 2A).

  The thermal image accumulation unit 26 accumulates the thermal image diagnosed by the determination unit 24 as data together with the extracted feature amount and the calculated similarity data (FIG. 2A). In addition, the thermal image of the solar cell module 300 previously captured, extracted and corrected by the module extraction unit 23 is also stored together with the feature amount and similarity data. Note that the cloud-type thermal image storage unit 26 is preferably applied to the present invention. Accurate diagnosis is possible by increasing the accumulation of thermal images. Further, the thermal image stored in the thermal image storage unit 26 is a specific solar power generation facility or a specific type of solar cell module 300, so that the specific solar power generation facility or the specific type of solar cell module 300 is used. It is possible to build a diagnostic system with high diagnostic accuracy specialized in the diagnosis.

  The notification unit 27 displays, on the display unit 21, the degree of similarity with each calculated diagnostic pattern for the thermal image of the solar cell module 300 that has been diagnosed (FIG. 2A).

(Second Embodiment)

  FIG. 3 is a block diagram illustrating a configuration example of the solar cell diagnostic system 100. The thermal image diagnosis unit 2 includes a module extraction unit 23, a temperature scale setting unit 28, a determination unit 24, a feedback unit 25, a thermal image storage unit 26, and a notification unit 27. A duplicate description with the first embodiment is omitted.

  In the second embodiment, a temperature scale setting unit 28 is provided, and optimum temperature scale information is output from the module extraction unit 23 to the thermal image capturing unit via the temperature scale setting unit 28.

  The module extraction unit 23 extracts the solar cell array 3, the string 30, and / or the solar cell module 300 from the captured thermal image, and outputs it to the temperature scale setting unit 28. The temperature scale setting unit 28 automatically sets the temperature scale within a range in which the extracted temperature distribution range of the entire solar cell or a desired part of the solar cell, for example, the temperature distribution range of the solar cell module 300 is comprehensively included. Set. The set temperature scale is output to the thermal image capturing unit 1. The thermal image capturing unit 1 captures a thermal image again based on the set temperature scale.

  Thereby, the thermal image of a high-resolution temperature scale is obtained with respect to a solar cell. Further, it becomes possible to recognize a temperature difference smaller than that in the prior art, and the subsequent determination accuracy can be improved.

  In the first embodiment, the range of coordinates of the outermost four points of the solar cell module 300, which is part of the extraction of the solar cell module 300, and the solar cell of the thermal image displayed on the display unit 21 are used. The number of modules 300 is input by the operator using the input unit 22, but in the second embodiment, the module extraction unit 23 automatically performs the operation without using the operator. That is, the module extraction unit 23 includes a function as the input unit 22.

  When the module extraction unit 23 automatically extracts one solar cell module 300, for example, a straight line detection method used in image processing is employed. As a straight line detection method, for example, a Hough transform (Hough transform or Hough Transform) method is used to detect a straight line of vertical lines m7 and horizontal lines m8 (contour) of the solar cell module 300, and further calculate the intersection of the solar cells Four points at the outermost part of the module 300 are automatically calculated. Here, the four outermost points (outermost points) are the suns that can be cut out as a single solar cell module 300 in the displayed thermal image, as described above in the first embodiment. When one or more battery modules 300 exist, the four apexes of the solar cell module 300 which can be cut out linearly are pointed out (FIG. 2B). Further, the vertical line m7 of the solar cell module 300 refers to a linear gap generated between the solar cell module 300 and the solar cell module 300 arranged in two on the left and right, and the temperature is lowered by being shaded. (FIG. 2B). Further, the horizontal line m8 of the solar cell module 300 refers to a linear gap generated between the solar cell module 300 arranged in two vertically and the solar cell module 300, and the temperature is lowered by being shaded. Part (FIG. 2B). Moreover, the module extraction part 23 calculates automatically the number of the solar cell modules 300 of the thermal image displayed on the display part 21 according to the number of the outermost intersections.

  In the second embodiment, Hough transformation is used. By detecting straight lines of the vertical line m7 and horizontal line m8 of the solar cell module 300, and further calculating intersections thereof, four points of the outermost contour (contour) of the solar cell module 300 are automatically calculated. Note that, as in the first embodiment, the outermost contour is not necessarily limited, and automatic calculation of the apex of the solar cell module 300 (for example, m5 and m6 shown in FIG. 2B) may be performed. As shown in FIG. 3, the straight line extraction unit H11 includes an outermost shell extraction unit H11, an intersection detection unit H12, and a Hough transform unit H13, and the module extraction unit 23 extracts the feature amount through these processes. The determination unit 24 determines this. A straight line passing through an arbitrary point of the image having the XY coordinate system is converted into coordinates (θ, ρ) in the Hough space. The number of straight lines converted into the coordinates (θ, ρ) of the Hough space is totaled using a two-dimensional array [θ] [ρ]. Then, a straight line in the XY coordinate system is detected from coordinates (θ, ρ) having a total value exceeding a predetermined threshold in the two-dimensional array [θ] [ρ]. When each point in the thermal image is converted into coordinates (θ, ρ) in the Hough space, a value weighted by the size of the feature amount of each point in the target image is added to the total value. The In addition to the detection of the temperature distribution range of the solar cell module by the temperature scale setting unit 28, it is also possible to automatically calculate four points of the outermost contour (contour) of the solar cell module 300 (FIG. 3).

  In the captured thermal image, the temperature corresponding to the frame of the solar cell module 300 or the temperature between the solar cell modules 300 is often different from the temperature of the solar cell surface and is displayed linearly. It is possible. Since the solar cell module 300 to be diagnosed can be automatically cut out and the number of the solar cell modules 300 can be automatically calculated, it is possible to omit the process involving the operator and achieve automation of fault diagnosis. This can greatly contribute to speeding up the diagnosis.

(Third embodiment)
In the solar cell diagnosis system 100 of the present embodiment, the determination unit 24 determines the state of the solar cell module 300, that is, the “feature amount” (including similarity) using artificial intelligence. Further, the configuration described below makes it possible to separate the abnormality and normal state of the bypass diode 300b, which could not be easily detected in the past, and to show the classification performance of f1-score 0.97. In addition, regarding the configuration related to the solar cell diagnosis system 100 of the present embodiment, the same configuration as the first and second embodiments is not described because it is redundantly described. In the present embodiment, the determination unit 24 classifies and / or determines the state of the solar cell module 300 as six patterns as in the first and second embodiments, but is not limited to this. Classification and / or determination may be made as the number of patterns equal to or greater than 6 patterns or less.

“Artificial intelligence” generally refers to an attempt to artificially realize intelligence similar to that of a human using a computer or the like and / or its technology. One type of artificial intelligence is machine learning, and “machine learning” generally refers to an attempt to give a computer the ability to learn without explicit programming and / or techniques thereof. One type of machine learning is “deep learning” using a multi-layered neural network. In conventional machine learning, a feature amount is extracted by human design, but in deep learning, a machine automatically extracts and acquires feature amounts from data. Here, the “feature amount” refers to a numerical value indicating what kind of features the learning data has in the field of artificial intelligence. A neural network called AutoEncoder is used for automatic feature extraction and acquisition in deep learning. “AutoEncoder” is composed of an input layer, an intermediate layer, and an output layer, and is used for supervised learning using the same data for the input layer and the output layer. In AutoEncoder, the input data is restored by output, and the input data is compressed and the features are condensed in the intermediate layer, so that the feature amount is automatically extracted. It should be noted that “similarity” to be described later may be included in the “feature amount” to determine an abnormal heat generation state such as a short circuit and an open circuit of the bypass circuit or a disconnection of the power generation circuit.
In the solar cell diagnostic method of the present invention, the thermal image capturing step (first step, S11) for capturing the surface of the solar cell as a thermal image, and the heat generation state of the entire or a part of the bypass circuit of the solar cell module or the solar cell array Judgment of determining the similarity between the feature amount extraction step (second step, S12) for extracting the feature amount of the heat generation state of the power generation circuit and the feature amount of the abnormal heat generation state that appears in the thermal image Steps (third step, S13) are sequentially performed. In the determination step (third step, S13), the similarity of the failure (abnormal heat generation state) is determined, the distortion of the thermal image is determined, the combination of these is determined, and further Is determined after artificial intelligence processing.

  A neural network having a plurality of intermediate layers (two or more layers) in deep learning is called a “deep neural network”. A convolutional neural network is a model that performs supervised learning well with a deep neural network structure. In the “convolutional neural network”, means for extracting a feature amount by convolution filter processing is used, and there is a convolutional neural network. In the learning of convolutional neural networks, learning is performed by a gradient descent optimization method using backpropagation (error back propagation method) on the premise of supervised learning. When there are a plurality of intermediate layers, it is known that backpropagation generally converges to an inappropriate minimum solution. Therefore, when the intermediate layer is made into a plurality of layers, only one intermediate layer is formed first, then the output layer is removed, the intermediate layer is regarded as the input layer, and another intermediate layer is formed and stacked. Such a method of making repeatedly is called a Stacked AutoEncoder. The Stacked AutoEncoder is a stack of AutoEncoders, and is one of unsupervised learning models that performs pre-learning in deep learning to extract feature quantities.

(Analysis of thermal images by deep learning)
A thermal image acquisition method, a data expansion method essential for accuracy improvement, and a convolutional neural network structure as a proposed classifier are shown.

  In this embodiment, six types of modules with different bypass circuit states (normal, open, short circuit) and power generation circuit states (normal, broken) are prepared for realization of fault diagnosis using deep learning. In between the thermal images were accumulated. Among the accumulated data, 1400 thermal images judged to be suitable for learning were used as the material of the present embodiment.

  In the present embodiment, deep learning is selected as the configuration of the determination unit 24, and a convolutional neural network is adopted among them. In other words, a convolutional neural network, such as elimination of position dependency using windows and convolutions, pooling, etc., and using feature values generated based on data and error signals was applied.

  In the case where the difference cannot be easily detected with the naked eye as in the states 01, 03, and 06 in FIG. 11 or in the case where there are classes in which the degree of similarity is increased by affine transformation as in the states 04 and 05, etc. It is difficult to achieve high accuracy only with this image processing. In addition to the above problems, the image classification problem in the thermal image is difficult to define the position dependency of the feature and the feature amount. However, in the present embodiment, these can be solved by using the convolution neural network in the determination unit 24.

  The total number of thermal images accumulated for learning is 1400, which is insufficient for learning of the deep neural network. For this reason, in this embodiment, P.I. Vincent, et al. , “Stacked denoising autoencoders: Learning useful representations in a deep network with a local de- ducing crea- tion,“ Journal of Machining Learning. 3371-3408, Dec. 2010. The Stacked Auto-encoder method as shown in Srivastava, et al. , "Dropout: a simple way to present neural networks from overfitting," Journal of Machine Learning Research, vol. 15 (1), pp. 1929-1958, 2014. The dropout method as shown in the above document was introduced to alleviate the problems of over-learning and gradient loss. In particular, to solve the problem of over-learning, a method for extending data in a timely manner was also examined.

  Based on the above, the following Thermal image data accumulation, 2. 2. expansion of learning data; The details of the convolutional neural network will be described.

(1. Apparatus and method related to thermal image data storage)
In order to increase the generalization capability, it is necessary to shoot the solar cell module 300 in different states under various shooting conditions and accumulate thermal image data. However, in taking a thermal image of a solar cell, the heat generation characteristics of the solar cell may not appear at all depending on the shooting condition, and it is necessary to examine an acceptable shooting condition.

  FIG. 4 shows differences in thermal images depending on the shooting angle α. A glass wafer is attached to the surface of the polycrystalline cell module made of silicon, which is the object of the present embodiment, for cell protection, and specular reflection is performed. Therefore, depending on the shooting angle α, the reflectivity becomes high and heat generation cannot be captured, and there is a possibility that shadows of the photographer and surrounding buildings may be reflected.

  Experimentally, the range of the shooting angle α in which the emissivity is almost constant and the range of the shooting angle α in which the emissivity rapidly decreases were evaluated. In both the vertical direction and the horizontal direction, the emissivity is maintained substantially constant up to 50 ° from the normal line of the installation surface of the solar cell module 300, but the emissivity suddenly decreased when it exceeded about 60 °. In photography such as infrared thermography, glass reflection affects the emissivity of sunlight. For this reason, when photographing a thermal image (infrared image), the deviation angle from the normal (imaging angle α) is defined to be 50 ° or less. Also, the image of the adjacent building and cloudy shadows were excluded, and the images were manually selected so that only the data considered to be the same pattern as the image taken from the top of the solar cell module 300 was learned. . The thermal image capturing conditions are summarized in Table 3 below. Table 4 shows the number of accumulated thermal images for each class (state) as the thermal image data details.

(2. Apparatus and method related to expansion of learning data: Data Augmentation)
The number of learning data can be increased by data expansion. However, if the characteristics of the original data are not fully taken into account, the generalization performance and accuracy of the model will be reduced. In this embodiment, A. Krizhevsky, et al. , "Imagenet classification with deep evolutional neural networks," Advanced in Neural Information Processing Systems, vol. 25, pp. 1106-1114, 2012. The data is extended by referring to the structure of AlexNet as shown in the above document and making adjustments so as not to lose the original meaning of the solar cell thermal image.

  An example of data expansion performed by the inventors in this embodiment is shown in FIG. In the present embodiment, inversion (upside down) S1, random crop S2, and rotation transformation S3 are used.

  First, the data expansion work by inversion S1 will be described. The solar cell module 300 installed this time has a power generation circuit and a bypass circuit. However, since one solar cell module 300 includes three clusters 300A, the same failure occurs when heat is generated in units of the clusters 300A. There may be cases where only the position of heat generation differs depending on the state. On the other hand, since there is only one junction box 300c for storing the bypass diode 300b for each solar cell module 300, the hot spot caused by the heat generation is very highly position dependent. For these reasons, only the upside down S1 is used in the data expansion work by the inversion S1, and the left / right inversion is not used.

  Next, random crop S2 and rotation transformation S3 will be described. In the thermal image of each solar cell module 300, the edge portion of the thermal image is likely to include a boundary line (boundary region) between the solar cell modules 300, and even if not included, it is affected by the adjacent solar cell module 300. Therefore, it is desirable to ignore it when diagnosing internal conditions. Random crop S2 and rotational transformation S3 both contribute to reducing the significance of the edge portion of the thermal image (possibly considered a meaningful portion in the thermal image). Here, the data is expanded by random cropping S2 up to 10% from the edge of the input image, and the data is further expanded by rotating each image by ± 2.5 ° S3.

  From the above work, 1120 data excluding 280 test sets were expanded to 5600 from 1400 data.

(3. Structure and method related to introduction of convolutional neural network)
FIG. 6A shows the structure of the designed convolutional neural network 240. This network has an input layer P1 and an output layer P4, and five convolution layers P2 and two full coupling layers P3 between them. The number of neurons in the input layer P1 and the output layer P4 is 3 × 32 × 64 = 6144 for the input layer P1 and 1 × 6 × 1 = 6 for the output layer P4. The number of neurons in the output layer P4 corresponds to the number of states in Table 4 (states 01 to 06). The activation function of the output layer P4 employs a softmax function, and indicates which state of the states 01 to 06 the input image corresponds to.

  The first convolution P2a in the convolution layer P2 is located next to the input layer P1 (FIG. 6A). Here, convolution is performed using a 5 × 5 size window to generate a 16-channel filter. Considering the above-mentioned problem of the boundary line and the characteristic of mixing hot spots and hot stripes, the window slide unit is set to 1, and zero-padding is not adopted. Next to the first convolution layer P2a, a maxpooling layer P2b using a 2 × 2 size window is placed. Further, a convolution layer P2c for generating a 32-channel filter and a maxpooling layer P2d having a 2 × 2 size window are placed. Thereafter, the output through the Dropout layer P2e with p = 0.5 is passed to the 1024-dimensional full coupling P3, and Dropout (P3a) with p = 0.5 is applied. All activation functions other than the output layer P4 are unified by ReLU.

  The details of this structure are almost the same as the structure of AlexNet of Krizhevsky et al., But the structure is relatively simple. However, the feature extraction in the first convolutional layer P2a does not have an ensemble, and the class is larger than the data dimension. Considering the fact that the number of is less, the difference is that Dropout (P3a) is also applied to the entire coupling layer P3. In addition, considering the complexity of the task and the number of pixels, a structure having a filter (5 × 5, Small size filter) having an area of 0.4% of the original image is adopted. Is adopted because the size of the hot spot that appears as heat generation in cell units is approximately 5 × 5 and the size of the cluster is approximately 10 × 64.

(3-2. Determination Method Using Convolutional Neural Network)
A method for identifying and diagnosing the state of the thermal image of the solar cell as states 01 to 06 using the convolutional neural network 240 will be described with reference to FIG. 6A. The determination unit 24 of the solar cell diagnostic system 100 includes a convolutional neural network 240, and a thermal image to be diagnosed is input as data to the input layer P1. The thermal image of the input layer P1 is extracted through the convolutional layer P2 and the 1024-dimensional feature quantity is extracted from the entire coupling layer P3 (FIG. 6B). The output layer P4 is composed of the same number of 6 units as the class to be classified (states 01 to 06), and the output layer P4 uses the softmax function of the activation function. Can be learned as representing the probability of belonging to a certain class (any one of states 01 to 06). In other words, the 1024-dimensional feature value extracted in all the coupling layers P3 indicates which state of the states 01 to 06 the input image to be diagnosed corresponds to within the numerical value range of 0 to 1 in the output layer P4. It is expressed probabilistically and output to the graph P5. The horizontal axis of the graph P5 is six classes (states), and the vertical axis is the value of the Softmax function.

  In the present embodiment, the “similarity” calculated by the determination unit 24 is a value calculated by the Softmax function (the numerical value on the vertical axis of the graph P5). The calculated value by the Softmax function is generally known, but an example of deriving the calculated value by the Softmax function is shown below.

In general, an artificial neuron receives one or more inputs and passes a weighted sum u to an activation function φ to generate an output y. First, the sum u weighted to each node for all inputs to the artificial neuron is expressed by the following equation (1), where n is the number of inputs, w is a weighting vector, and x is an input vector.

Usually, the sum u weighted by each node is passed to the activation function φ, and the output y is calculated. For example, as shown in the following formula (2), the output y is calculated through the activation function φ by adding the bias value b to the sum u.

Here, when the activation function φ is the Softmax function, the activation function φ is expressed by the following formula (3), where K is the number of classes. Assuming classification into K classes, the output is K, the sum is 1, and the output y can be interpreted as the probability of belonging to that class.

  The output y calculated in this way is a calculated value by the Softmax function, and is the “similarity” calculated by the determination unit 24 in the present embodiment.

  Note that, using the similarity of the present embodiment, the similarity can be displayed on the diagnosis history screen as shown in FIG. 2C. At this time, the value displayed in the similarity 1 column in FIG. 2C is the probability of belonging to the class of state 01 (similarity of state 01), and the value displayed in the similarity 2 column is state 02. Is the probability of belonging to the class (state 02 similarity), and the value displayed in the similarity 3 column is the probability of belonging to the class of state 03 (similarity of state 03). The value displayed in the column is the probability of belonging to the class of state 04 (similarity of state 04), and the value displayed in the column of similarity 5 is a value indicating the probability of belonging to the class of state 05 ( The value displayed in the similarity 6 column is a value indicating the probability of belonging to the class of state 06 (similarity of state 06), and these are all in the range of 0 to 1. Takes the value of

  The determination unit 24 includes a convolution filter (a filter that has been subjected to optimal weighting and / or feature extraction) that has been preliminarily learned by the learning process and has an optimized filter coefficient, so that it is reflected in the input thermal image. Image recognition processing of the solar cell can be performed.

(About learning process)
In the convolutional neural network, the convolution filter coefficient used in the recognition process included in the determination unit 24 is optimized by performing learning processing using a learning data set with a label (true value), and optimal weighting and / or feature extraction. Is made.

The learning principle in the neural network is to minimize the error function that is a measure of correctness. A set of input vectors {x _n } (n = 1, 2,..., N), a set of output vectors {y _n } (n = 1, 2,..., N), a target vector (true value) ) Is defined as {t _n } (n = 1, 2,..., N). Assuming that, for example, a cross-entropy function is used as the error function E (w), the error function E (w) is represented by the following equation (4), and weighting that minimizes E (w) for learning. (Weight coefficient) w is obtained.

The optimization method used for learning utilizes a stochastic gradient descendant (SGD). In the present embodiment, Y. Nesterov, “A method for unconstrained convex minimization probe with the rate of convergence O (1 / k2),“ Doklad an SSSR, Vol. 269 (3), 1983. Nestrov momentum (Nesterov accelerated gradient) as shown in the above document was adopted, the learning rate was set to 0.01, and maxepoch was set to 1000 after applying earlystop. As the learning data, 5600 sheets that have been expanded as described above are used.

The stochastic gradient descent method is an improvement of the steepest descent method, and the steepest descent method is represented by the following equation (5). In the steepest descent method to calculate the slope ∇E of current weight w _t (w _t), to update the weight w _t to the weight w _{t + 1} in multiplied by the learning rate η the gradient ∇E (w _t), The iteration is repeated until the error function E (w _t ) is minimized.

If the error function E (w _t ) has N pieces of training data, it can be decomposed into N functions, and the error function E (w _t ) can be expressed as a sum of errors obtained from the respective pieces of training data by the equation It can be expressed as (6).

  The output label (true value) is given in advance by the user, and the probabilistic gradient descent method is based on the error between the output value subjected to discrimination processing and the true value, and moves forward from the filter coefficient close to the output layer. The filter coefficients are corrected in order. In the probabilistic gradient descent method, the filter coefficient is corrected based on the accumulated error of recognition processing for multiple images. Therefore, compared with the case of correcting the filter coefficient based on the error in a single image, Bias and vibration are reduced.

Here, the case of updating the position with a moment is called momentum. The momentum is expressed by the following equation (7), and in addition to the gradient ∇E (w _t ) at the current weight w _t multiplied by the learning rate η, the attenuation rate γ with respect to the previous update amount The error function E (w _t ) is minimized by repeatedly updating the weight using a value multiplied by (for example, 0.9).

Further, the nesterov momentum (Nesterov accelerated gradient) is expressed as the following equation (8). For brevity, it is effective to use the approximate value of the parameter at the next position instead of the gradient calculation for the current parameter. It is a way to proceed.

  As described above, the output result (the result of identifying the input image) is compared with the true value, and the filter coefficient is optimized by repeating the update of the filter coefficient (optimum weighting). And / or feature extraction).

  By configuring the determination unit 24 as described above, the solar cell diagnostic system 100 and the method thereof according to the present embodiment are configured. In the following, by evaluating and considering the results of the simulation using the test set, The effect of the solar cell diagnostic system 100 and the method of the present embodiment will be described.

(Effects of the apparatus and method of the present embodiment)
FIG. 7 shows error function value transitions in the learning data (train) and the test data (valid). The X axis is the Epoch, and the Y axis is the logarithmic value of the error function. The term “epoch” refers to a cycle in which the learning data used is updated, and learning is generally performed by repeating the “epoch” several times. Since the error due to the learning data is dominant, it can be seen that some over-learning has occurred, but since both the learning data and the test data show a tendency to decrease, significant learning has progressed. I understand that.

FIG. 8 shows the result of simulation using test data. Judgment accuracy is evaluated by Precision (relevance ratio), Recall (reproducibility), and F1 value (F value). The Precision, the Recall, and the F1 value are calculated by the following formula (9). Support is the number of correct answers in the test data.
(Equation 9)
Precision (rate of precision) = number of correct answers of system / number of outputs of system Recall (recall rate) = number of correct answers of system / number of correct answers in test data (F value) = 2 × Precision × Recall / (Precision + Recall)

  In general, there is a trade-off relationship between Precision and Recall, and when one is increased, the other tends to decrease. In machine learning, an attempt is made to increase both Precision and Recall. Therefore, an evaluation scale called F1 value, which is obtained by combining Precision and Recall by harmonic averaging, is used. From the result of FIG. 8, it can be confirmed that both the average values of Precision and Recall are both recorded as 0.97. On the other hand, only “Recall” in the normal state is 0.87, which is evaluated relatively lower than the other states. This indicates that the network is excessively sensitive to heat generation, distortion, and the like, and it is considered that the normal state thermal image used for learning is fewer than the abnormal state image. This suggests that further performance improvement of this network can be expected by performing additional learning in the direction to compensate for unequal data.

  FIG. 9 shows a filter located in the convolution layer as an image. In the deep learning model used in object recognition, the filters located in the first convolution layer include Edge (border where color and shade change) and Blob (the same gray scale aggregate existing in the image). Edge and Blob patterns often appear as filters that extract and react primitive information such as pixels of the same density). However, from the result of FIG. 9, the shape of Edge is hardly seen in the first-layer filter.

  When solving classification problems with small-scale data, a transfer learning approach using a learned network is the mainstream. Transfer learning is a technology that adapts a model that has already been learned to another area, such as adapting a model that has been learned in an area where much data can be collected to another area where data collection is difficult. Made. However, the results of FIGS. 8 and 9 suggest that the base domain necessary to solve this problem is different from the network used in general object recognition and classification problems. From the above, it can be seen that the generation of a full scratch classifier performed in the present embodiment is useful.

  In the present embodiment, a deep learning structure for identifying anomalous factors from a solar cell thermal image is proposed, and the results of simulations using data accumulation and learning techniques necessary for the learning and a test set are shown. By defining six classes from the bypass circuit and the power generation circuit, constructing a convolutional neural network that solves the classification problem, and expanding and learning the collected 1400 thermal images in a timely manner, it is possible to use f1-score in the test data. It became possible to show a classification performance of 0.97. Conventionally, classification accuracy such as the open state of the bypass circuit that cannot be easily classified has been improved.

  In the present embodiment, the extracted / corrected solar cell modules are divided into the above six types, and their features are extracted. Then, we worked on a technique that treats the output of all connected layers of convolutional neural networks as a feature.

  In this embodiment, focusing on the fact that classifying the six states of the thermal image will require a different ability from the domain for distinguishing objects, the method of performing a full scratch build by increasing the collected images in a timely manner. A convolutional neural network was constructed. In the model configuration, the window shape, Pooling layer, and DropOut layer were matched to the physical structure of the solar cell. In particular, since each solar cell module 300 is configured in three clusters 300A and each abnormal state appears independently for each cluster 300A, the classification performance is improved by setting the windows to be long. From these contrivances, a Convolutional Neural Network having an f1-score of 0.94 to 0.95 could be constructed by cross-validation. In addition, a Ranom Forest structure using the output of all of the Convolutional Neural Networks as a feature quantity was introduced. This was introduced because diagnosing a non-failure as non-failure is more fatal than diagnosing a non-failure as failure, and at the expense of false positives by diluting the accuracy of the neural network. And has the effect of alleviating false negatives. By introducing these devices, it was confirmed that f1-score increased to about 0.97. By adopting the present embodiment based on these ideas, it is possible to improve the discrimination ability for the open state of the bypass circuit in which no hot spot or hot stripe occurs, compared to the method using transfer learning.

(Fourth embodiment)
In general, convolutional neural networks that are well trained on large-scale data have an abstract discriminating ability for domains. Recent research has shown that the discriminatory ability shows high accuracy even if it is not a provisional task when constructing the model, if the target domain has some commonalities with the original domain. ing. Such a learning method is called transfer learning. In the present invention, a method using transfer learning was also examined. The convolutional neural network generated from the large-scale data set including various things such as ImageNet and Cifar-10 was transferred, and the discriminator generation of the solar cell module was worked on. The discriminating ability for the six states was up to about 0.80 based on f1-score. As described above, this embodiment can be changed and expanded without departing from the spirit of the present embodiment.

1 Thermal image capture unit,
2 Thermal imaging diagnostic department,
3 Solar cell array,
23 module extractor,
24 determination unit,
25 Feedback section,
26 Thermal image storage unit,
27 Notification section,
28 Temperature scale setting section,
30 strings,
100 solar cell diagnostic system,
101 solar power generation system,
300 Solar cell module

Claims (14)

  In a solar cell diagnostic method for photographing a solar cell surface as a thermal image and diagnosing a failure of a solar cell, the solar cell surface is photographed as a thermal image and the entire solar cell module or solar cell array or part of a bypass circuit The sun is characterized by separately detecting the heat generation state and the heat generation state of the power generation circuit, detecting the temperature distribution of these heat generation states, and combining them to determine as a characteristic of the abnormal heat generation state appearing in the thermal image Battery diagnostic method.   In a solar cell diagnostic method for photographing a solar cell surface as a thermal image and diagnosing a failure of the solar cell, a thermal image photographing step for photographing the solar cell surface as a thermal image, and a whole or a part of the solar cell module or solar cell array A feature amount extraction step of extracting a feature amount from the heat generation state of the bypass circuit and the heat generation state of the power generation circuit, and the similarity between the state of the thermal image captured based on the feature amount and the abnormal heat generation state appearing in the thermal image A solar cell diagnostic method comprising: a determination step for determining. As a feature of the abnormal heat generation state, the temperature distribution of the heat generation state is equal to or higher than a predetermined temperature, and a linear abnormal heat generation state called a hot stripe appearing in a strip shape is displayed, and the state of the bypass circuit is short-circuited and opened. The solar cell diagnosis method according to claim 1, wherein the solar cell diagnosis method is determined as either one of a short circuit and an open state of the state of the power generation circuit. As the characteristic of the abnormal heat generation state, a point-like abnormal heat generation state called a hot spot and / or a linear abnormal heat generation state called a hot stripe appearing in a strip shape are displayed together with a temperature distribution to determine, or The feature value is extracted by an analysis algorithm from the heat generation state of the bypass circuit in the battery module and the heat generation state of the power generation circuit, and the identification result by the analysis algorithm is displayed and determined as an abnormal heat generation state. 2. The solar cell diagnostic method according to 2.   The temperature distribution of the heat generation state is determined by automatically or manually setting a temperature scale within a range in which the temperature distribution of the whole or a part of the solar cell module or the solar cell array extracted by photographing is comprehensively included. The solar cell diagnosis method according to claim 1, wherein a heat generation state is detected.   The thermal image diagnosis unit includes a module extraction unit that extracts and distorts an image corresponding to the solar cell module or the solar cell array from the captured thermal image, and the module extraction unit includes the solar cell module. 6. The solar cell diagnostic method according to claim 1, wherein a predetermined outermost point that is an outermost contour is calculated, and distortion captured by the thermal image diagnostic unit is corrected.   The solar cell diagnosis method according to claim 6, wherein the module extraction unit corrects distortion captured by the thermal image diagnosis unit using a straight line detection method of image processing.   8. The convolutional neural network, which is deep learning of artificial intelligence, is used for the similarity of the abnormal heat generation state of the determination unit (determination step) and / or the distortion correction of the module extraction unit. The solar cell diagnostic method according to any one of the above.   In the solar cell diagnostic system based on thermal image analysis of the surface of the solar cell, the solar cell diagnostic system divides the heat generation state of the entire solar cell module or solar cell array or a part of the bypass circuit and the heat generation state of the power generation circuit. A thermal image capturing unit configured to detect the surface of the solar cell as a thermal image, and a thermal image diagnostic unit for diagnosing the captured thermal image, wherein the thermal image diagnostic unit includes the thermal image diagnostic unit, A solar cell diagnostic system comprising: a determination unit configured to determine the thermal image by dividing a diagnostic pattern based on a state of a bypass circuit and a state of a power generation circuit.   The thermal image diagnosis unit includes a module extraction unit that extracts and distorts an image corresponding to the solar cell module or the solar cell array from the captured thermal image, and the module extraction unit includes the solar cell module. The solar cell diagnostic system according to claim 9, wherein a predetermined outermost point serving as an outermost contour is calculated to correct distortion captured by the thermal image diagnostic unit.   The solar cell diagnosis system according to claim 9, wherein the module extraction unit corrects the distortion photographed by the thermal image diagnosis unit using a straight line detection method of image processing.   The determination unit divides the state of the bypass circuit into three types of normal, short circuit, and open, divides the state of the power generation circuit into two types of normal and open, and states the state of the bypass circuit and the state of the power generation circuit. 12. The sun according to claim 9, wherein the diagnostic pattern is divided into six types from combinations, and feature amounts are extracted based on a distribution and shape of a heat source in each failure state. Battery diagnostic system.   The module extraction unit detects a vertical line and a horizontal line that are the outermost contour of the solar cell module using a straight line detection method of image processing, calculates an intersection of the vertical line and the horizontal line, and calculates the outermost contour The solar cell diagnostic system according to any one of 9 to 12, wherein four points are automatically calculated. The thermal image diagnosis unit includes a temperature scale setting unit that automatically or manually sets a temperature scale within a range in which the temperature distribution of the whole or a part of the solar cell module or the solar cell array is comprehensively included, and the heat The solar cell diagnostic system according to any one of claims 9 to 13, wherein the image capturing unit captures a thermal image again based on the set temperature scale.