A proton exchange membrane fuel cell state monitoring method, device and medium

Abstract

The present invention discloses a proton exchange membrane fuel cell state monitoring method, device and medium, which belongs to the field of fuel cell technology, and is used to solve the technical problem that the existing method of monitoring the state of fuel cells based on electrochemical impedance spectroscopy uses the voltage response of the fuel cell as a one-dimensional signal analysis, resulting in limited ability to accurately identify the performance state of PEMFC. The method includes: obtaining a voltage response signal after the fuel cell is injected with a current interference signal; decomposing the voltage response signal into multiple voltage response components; for any voltage response component of the multiple voltage response components, performing feature extraction to obtain a target feature corresponding to the voltage response signal; inputting the target feature into a classifier model to determine the state of the fuel cell. The present invention obtains multi-scale information of the voltage response signal in EIS through the above method, thereby improving the accuracy of fuel cell state identification.

Description

Technical Field

[0001] The present invention relates to the technical field of fuel cells, and in particular to a proton exchange membrane fuel cell state monitoring method, equipment and medium. Background Art

[0002] In the field of health monitoring and performance recovery of proton exchange membrane fuel cells (PEMFCs), electrochemical impedance spectroscopy (EIS) technology has been applied in fault diagnosis. The state of the fuel cell is determined by analyzing the impedance spectrum of the fuel cell by injecting a current interference signal into the fuel cell to obtain a voltage response. However, existing methods are mostly limited to one-dimensional signal analysis. The voltage response obtained after injecting the current interference signal is analyzed as a one-dimensional signal, which fails to fully utilize the multi-dimensional characteristics of the voltage response data, resulting in limited ability to accurately identify the performance status of PEMFC under complex operating conditions.

[0003] For example, the prior art has accumulated a lot of research on fuel cell fault classification diagnosis, but the classification of gain states is rarely involved. After performance recovery is achieved, the gain state and the normal state are classified into the same state category, and the classification of the normal state and the gain state cannot be achieved. Summary of the invention

[0004] The present invention provides a proton exchange membrane fuel cell state monitoring method, device and medium, which are used to solve the technical problem that the existing method of monitoring the state of fuel cells based on electrochemical impedance spectroscopy uses the voltage response of the fuel cell as a one-dimensional signal analysis, resulting in limited ability to accurately identify the performance state of PEMFC.

[0005] In a first aspect, the present invention provides a proton exchange membrane fuel cell state monitoring method, comprising:

[0006] Acquire a voltage response signal of the fuel cell after a current interference signal is injected;

[0007] decomposing the voltage response signal into a plurality of voltage response components;

[0008] For any voltage response component of the voltage response signal, feature extraction is performed to obtain a target feature corresponding to the voltage response signal;

[0009] The target feature is input into a classifier model to determine the state of the fuel cell.

[0010] In a feasible implementation, before obtaining the voltage response signal after the fuel cell is injected with the current interference signal, the method further includes: performing oxygen-deprivation treatment on the fuel cell so that the fuel cell remains in the oxygen-deprivation state for a preset time period.

[0011] In a feasible implementation manner, the voltage response signal is decomposed into a plurality of voltage response components, including:

[0012] The voltage response signal is decomposed into a plurality of intrinsic mode components by empirical mode decomposition, and the intrinsic mode components are determined as voltage response components.

[0013] In a feasible implementation manner, for any voltage response component of the voltage response signal, feature extraction is performed to obtain a target feature corresponding to the voltage response signal, including:

[0014] Converting any voltage response component of the voltage response signal into a symmetrical point image corresponding to any voltage response component through a symmetrical point pattern;

[0015] Feature extraction is performed on the symmetrical point image to obtain target features.

[0016] In a feasible implementation manner, converting any voltage response component of the voltage response signal into a symmetrical point image corresponding to any voltage response component through a symmetrical point pattern includes:

[0017] Based on at least two different numbers of mirror symmetry planes, any voltage response component is converted into at least two different symmetry point images through a symmetry point pattern.

[0018] In a feasible implementation manner, feature extraction is performed on the symmetrical point image to obtain target features, including:

[0019] Extracting scale-invariant features of the symmetrical point image through scale-invariant feature transformation;

[0020] Copying the symmetrical point image to obtain two identical symmetrical point images, and determining whether the scale-invariant features in the two identical symmetrical point images match based on the distance between the scale-invariant features in the two identical symmetrical point images, to obtain a matching point set;

[0021] The matching point sets are placed on two different planes respectively, and matching points that match each other are connected to obtain matching point vectors corresponding to the matching points;

[0022] Based on the projection of the matching point vector on the reference vector, determining whether the matching point vector overlaps with the reference vector, and determining the matching point vector overlapping with the reference vector as an overlapping vector;

[0023] Constructing a two-dimensional grid consisting of a plurality of cells, determining the cells to which the overlapping vectors belong, and drawing a density map of the overlapping vectors based on the number of overlapping vectors in the cells;

[0024] A GIST feature of the density map is calculated, and the GIST feature is determined as a target feature.

[0025] In a feasible implementation, inputting the target feature into a classifier model to determine the state of the fuel cell includes:

[0026] The target characteristics are input into the ASSDRB to determine the status of the fuel cell.

[0027] In a feasible implementation, inputting the target feature into a classifier model to determine the state of the fuel cell includes:

[0028] The target feature is input into a classifier model to determine whether the fuel cell is in a normal state or a gain state.

[0029] In a second aspect, the present invention provides a proton exchange membrane fuel cell state monitoring device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor so that the at least one processor can execute a proton exchange membrane fuel cell state monitoring method described in any of the above embodiments.

[0030] In a third aspect, the present invention provides a non-volatile computer storage medium, which is a non-volatile computer-readable storage medium, and the non-volatile computer-readable storage medium stores at least one program, each of which includes instructions, and when the instructions are executed by a terminal, the terminal executes a proton exchange membrane fuel cell state monitoring method described in any of the above embodiments.

[0031] The proton exchange membrane fuel cell state monitoring method, device and medium provided by the present invention have the following beneficial technical effects compared with the prior art:

[0032] (1) The present invention decomposes the voltage response signal in the EIS signal to obtain multiple voltage response components, and extracts features from these voltage response components to obtain target features of the voltage response signal. The target features contain multi-scale information of the voltage response signal, which provides richer information for the classifier model to determine the state of the fuel cell, thereby improving the accuracy of fuel cell state identification.

[0033] (2) The symmetrical point image in the present invention intuitively represents the signal amplitude and frequency changes of the voltage response component in a visual manner. By converting the voltage response component into a symmetrical point image through the symmetrical point pattern, subtle changes of the voltage response component in EIS can be captured.

[0034] (3) After converting the voltage response component from a one-dimensional time signal to a two-dimensional symmetrical point image through a symmetrical point model, the present invention determines the scale-invariant features in the symmetrical point image, copies the symmetrical point image, and places the matching point sets (mutually matching scale-invariant features) of the two identical symmetrical point images on different planes for three-dimensional reconstruction, thereby exploring the information of the symmetrical point image in high-dimensional space; then, by projecting the matching point vector on the reference vector, the information of the symmetrical point image is obtained from a new perspective, and a two-dimensional density map of overlapping vectors is obtained; by performing GIST feature extraction on the density map, a one-dimensional global feature of the density map is obtained; and by using the above method, the voltage response information is converted in the order of one-dimensional→two-dimensional→three-dimensional→two-dimensional→one-dimensional, so that the target features obtained are more reasonable, thus breaking through the limitations of previous research on symmetrical point images.

[0035] (4) The classifier model ASSDRB adopted in the present invention can actively learn from unlabeled training data sets in a semi-supervised mode, and continuously update and optimize the performance of the classifier model.

[0036] (5) The present invention performs oxygen-deprivation treatment on the fuel cell, so that the fuel cell remains in an oxygen-deprived state for a preset time. The excessive concentration gradient causes hydrogen ions to move toward the cathode to reduce platinum oxide, which can greatly improve the performance gain recovery of the fuel cell. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0038] Figure 1 A flow chart of a proton exchange membrane fuel cell state monitoring method provided by the present invention;

[0039] Figure 2 A schematic diagram of a voltage response signal of a fuel cell in a normal state provided by the present invention;

[0040] Figure 3 A schematic diagram of a voltage response component of a fuel cell in a normal state provided by the present invention;

[0041] Figure 4 The SDP image of the voltage response component of a fuel cell in a normal state provided by the present invention;

[0042] Figure 5 The SDP images of the six voltage response components of the fuel cell under normal conditions obtained by using different numbers of mirror symmetry planes;

[0043] Figure 6 A flowchart of a method for obtaining a matching point set of an SDP image provided by the present invention;

[0044] Figure 7 A schematic diagram of a matching point set in an SDP image provided by the present invention;

[0045] Figure 8 A three-dimensional schematic diagram of placing the matching point sets in the SDP image on two different planes;

[0046] Fig. 9 A density map of overlapping vectors provided by the present invention;

[0047] Fig.10 A density map with a size of 28 pixels × 28 pixels provided by the present invention;

[0048] Fig.11 A schematic diagram of a GIST feature of a density map provided by the present invention;

[0049] Fig.12 A schematic diagram of classification accuracy of fuel cell status monitoring by different methods provided by the present invention;

[0050] Fig.13 A schematic diagram of a confusion matrix provided by the present invention;

[0051] Fig.14 A schematic structural diagram of a proton exchange membrane fuel cell state monitoring device provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] In the field of health monitoring and performance recovery of proton exchange membrane fuel cells, although electrochemical impedance spectroscopy technology has been applied in fault diagnosis, existing methods are mostly limited to one-dimensional signal analysis and fail to fully utilize the multidimensional characteristics of data, resulting in limited ability to accurately identify the performance status of PEMFC under complex working conditions. This technical solution aims to overcome the shortcomings of the prior art by introducing an innovative three-dimensional signal processing framework, combining empirical mode decomposition and electrochemical impedance spectroscopy, as well as advanced image processing technology, to achieve a more accurate classification of the performance status of PEMFC. At the same time, a large amount of research has been accumulated on the classification and diagnosis of fuel cell faults, but the classification of gain states has rarely been involved. In previous studies, gain states and normal states were classified in the same state, that is, there was a lack of methods to finely classify normal states and gain states. Therefore, this article will focus on achieving performance recovery while completing the identification of gain states. The technical solution proposed in the present invention is described in detail below through the accompanying drawings.

[0054] Figure 1 The present invention provides a flow chart of a proton exchange membrane fuel cell state monitoring method. Figure 1 As shown, the method includes the following execution steps:

[0055] S101, obtaining a voltage response signal of a fuel cell after a current interference signal is injected into the fuel cell.

[0056] The present invention obtains a voltage response signal of the fuel cell after a current interference signal is injected into the fuel cell.

[0057] As a possible implementation, in order to achieve a gain state for the fuel cell, before obtaining a voltage response signal after the fuel cell is injected with a current interference signal, the method also includes: performing oxygen-deprivation treatment on the fuel cell so that the fuel cell remains in an oxygen-deprived state for a preset time.

[0058] For example, the preset duration can be set to 20s, which means that the fuel cell with performance degradation is cut off from oxygen for 20s while maintaining the supply of hydrogen; in this way, the excessive concentration gradient causes hydrogen ions to move toward the cathode to reduce platinum oxide, and the fuel cell gains 6% performance compared to the initial state.

[0059] S102: Decompose the voltage response signal into multiple voltage response components.

[0060] The present invention decomposes the voltage response signal. After a current interference signal is injected into a fuel cell, the present invention can obtain multiple voltage response components, thereby increasing the multi-scale information of the voltage response signal.

[0061] As a possible implementation manner, S102, decomposing the voltage response signal into multiple voltage response components, includes: decomposing the voltage response signal into multiple intrinsic mode functions (IMFs) by empirical mode decomposition (EMD), and determining the IMFs as voltage response components.

[0062] As the formula As shown, the present invention converts the voltage response signal E ( t ) is decomposed into N indivual IMF n ( t ) and a residual signal e ( t ).

[0063] Each IMF must satisfy (1) the number of local maxima and local minima it contains must be equal or the difference does not exceed 1; (2) the average value of the envelope formed by the local maximum points and local minimum points is zero.

[0064] The IMF extraction process can be carried out through the following steps:

[0065] First, determine E ( t ) and construct the upper envelope using interpolation E upper ( t ) and the lower envelope E lower ( t ).

[0066] Based on the formula , calculate the average value of the upper envelope and the lower envelope, and determine the average value as the average signal E mean ( t ).

[0067] Based on the formula , from the voltage response signal E ( t ) minus the average signal E mean ( t ), get the first residual e 1( t ),Sure e 1( t ) is the IMF.

[0068] If confirmed e 1(t ) is not the IMF, e 1( t ) is considered as a new signal and the above steps are repeated.

[0069] when e 1( t ) meets the IMF conditions, e 1( t ) is determined as IMF 1( t ), and based on the formula , from the voltage response signal E ( t ) minus IMF 1( t ), get a new signal E’ ( t ).

[0070] for E’ ( t ) Repeat the above steps until E ( t ) to extract all IMF and e ( t ). Each IMF can be expressed as follows:

[0071]

[0072] in, a n0 , a nk , ω n , φ n is the coefficient to be determined, K n is the number of cosine terms for each IMF. These coefficients are determined through the above iterative process, which aims to ensure that each IMF satisfies the two basic conditions mentioned above.

[0073] S103 . For any voltage response component of the voltage response signal, perform feature extraction to obtain a target feature corresponding to the voltage response signal.

[0074] The present invention extracts features from each voltage response component obtained by decomposing the voltage response signal to obtain component features corresponding to each voltage response component. The component features corresponding to these voltage response components are also target features corresponding to the voltage response signal.

[0075] As a possible implementation manner, S301 performs feature extraction on any voltage response component among the multiple voltage response components to obtain a target feature corresponding to any voltage response component, including:

[0076] S301A, converting any voltage response component among the plurality of voltage response components into a symmetrical point image through a symmetrical point pattern.

[0077] The symmetric dot image in the present invention intuitively represents the signal amplitude and frequency changes of the voltage response component in a visual manner. The voltage response component is represented based on the polar coordinate system through the symmetric dot pattern (SDP), and the voltage response component is converted into a mirror-symmetrical snowflake image, that is, the symmetric dot image, SDP image, to capture the subtle changes of the voltage response component in EIS.

[0078] In some embodiments, the radius and the deflection angle of each voltage response component in a polar coordinate system may be analyzed to convert the voltage response component into an SDP image.

[0079] For example, for the voltage response component IMF obtained by EMD, the signal sampling point is point T, and the radius of point T is determined by the following formula: r （T） , counterclockwise rotation angle α （T） and clockwise rotation angle β （T） .

[0080]

[0081]

[0082]

[0083] NMSP

[0084] in, and It's time point t and t+m Place n indivual IMF The amplitude of m is the length of the interval; and Indicates n indivual IMF The minimum and maximum values ​​of θ NMSP is the rotation angle of the mirror's symmetry plane, NMSP represents the number of symmetry planes of the mirror, μ is the angular magnification factor.

[0085] In this way, the voltage response component is transformed from a one-dimensional time signal to a two-dimensional symmetrical point image through the symmetrical point pattern.

[0086] As a possible implementation method, in order to further improve the information dimension, any voltage response component of at least one voltage response component can be converted into at least two different symmetry point images through a symmetry point pattern based on at least two different numbers of mirror symmetry planes.

[0087] For example, each voltage response component is converted into six different symmetrical point images whose mirror symmetry planes are 1, 2, 3, 4, 5, and 6 respectively.

[0088] S301B, extract features from the symmetrical point image to obtain target features.

[0089] In order to facilitate the processing of the classifier model, after the voltage response components are converted into symmetrical point images, the features of each voltage response component are extracted as the target features of the voltage response signal and used as the input of the classifier model.

[0090] As a possible implementation method, S301B extracts features from the symmetrical point image to obtain target features, including:

[0091] SB1. Extract the scale-invariant features of the symmetrical point image through scale-invariant feature transformation.

[0092] The present invention can extract scale-invariant features of symmetrical point images through scale-invariant feature transform (SIFT).

[0093] In some embodiments, before extracting the scale-invariant features of the symmetric point image, the symmetric point image is first converted into a grayscale image. I , and then preprocess the symmetrical point image to obtain the preprocessed symmetrical point image I’ ,The preprocessing includes contrast enhancement, median filtering for noise reduction, and Laplacian sharpening, as shown in the following equation.

[0094]

[0095] in, M represents the median filter, K is the sharpening strength factor, L represents the Laplace operator, ρ is the contrast control factor, σ is an image I The average brightness.

[0096] SB2. Copy the symmetrical point image to obtain two identical symmetrical point images, and determine whether the scale-invariant features in the two identical symmetrical point images match based on the distance between the scale-invariant features in the two identical symmetrical point images to obtain a matching point set.

[0097] In some embodiments, D 1 and D 2 represents the scale-invariant features in two identical symmetrical point images, and the Euclidean distance between the scale-invariant features of the two images is calculated. d ( D 1 ,D 2), as shown in the following formula.

[0098]

[0099] in, d 1i yes D 1 in i The value of the dimension, d 2i yes D 2 in i The value of the dimension.

[0100] For one of the images, the scale-invariant features D 1. Find the scale-invariant feature closest to the scale-invariant feature in another image D nn and the second closest scale-invariant feature D nnn .

[0101] Then, the closest scale-invariant feature D nn and the second closest scale-invariant feature D nnn Apply a ratio test between them to compute the scale-invariant features of one of the images D 1 and the closest scale-invariant feature in the other image D nn The first distance between, and one of the scale-invariant features of the image D 1 and the second closest scale-invariant feature in the other image D nnn If the ratio of the first distance to the second distance is less than a preset ratio threshold, the match is determined to be valid, and the scale-invariant feature of one of the images is converted into D 1 and scale-invariant features in another image D nn Determined as a set of matching points.

[0102] It can be understood that it is necessary to ensure that each scale-invariant feature is matched only once to obtain a set of matching points for two identical symmetrical point images.

[0103] SB3. Place the matching point sets on two different planes, connect the matching points that match each other, and obtain the matching point vectors corresponding to the matching points.

[0104] In order to explore the characteristics of voltage response components in three-dimensional space, the present invention places the matching points of one image in two identical symmetrical point images on a first plane, and the matching points of the other image on a second plane, the first plane and the second plane are different, and the matching points of one image are connected with the corresponding matching points in the other image to obtain a matching point vector.

[0105] In this way, the voltage response component is converted from a two-dimensional symmetrical point image to a three-dimensional matching point set on two different planes through the above method.

[0106] SB4. Based on the projection of the matching point vector on the reference vector, determine whether the matching point vector overlaps with the reference vector, and determine the matching point vector that overlaps with the reference vector as the overlapping vector.

[0107] The present invention can obtain information of symmetrical point images from a new perspective by matching the relationship between the point vector and the reference vector.

[0108] In some embodiments, the projection length of the matching point vector on the reference vector is calculated, and when the projection length is greater than a preset length threshold, it is determined that the matching point vector overlaps with the reference vector, and the matching point vector is determined as an overlapping vector.

[0109] In SB3, the matching points of one of the two identical symmetrical point images are placed on the first plane, and the matching points of the other image are placed on the second plane, thus forming a three-dimensional space.

[0110] P ni ( x i , y i , z i ) represents the first i The first point in the set of matching points of the image n The coordinates of the points, where i is 1 or 2, i =1 means the point is located in the first plane, i =2 means that the point is located in the second plane.

[0111] Exemplarily, the first plane and the second plane may be z =0 andz =1, that is, z 1=0 and z 2=1.

[0112] is a set of matching points on two planes that match each other P n1 and P n2 The vector formed is the matching point vector. is the reference vector.

[0113] Among them, the reference vector Can be with xy The angle between the planes is a vector of 45 degrees, or a vector of other angles, which is not limited in the present invention.

[0114] The absolute value of the dot product of two vectors represents the length of the projection, so we can find the length of the projection at the matching point vector. With reference vector When the absolute value of the dot product is greater than a preset length threshold, it is determined that the matching point vector overlaps with the reference vector.

[0115] SB5. Construct a two-dimensional grid consisting of multiple cells, determine the cells to which the overlapping vectors belong, and draw a density map of the overlapping vectors based on the number of overlapping vectors in the cells.

[0116] In some embodiments, the two-dimensional grid may be xy Plane-parallel, overlapping vectors of matching points xy If the coordinates are within the cell range, it is determined that the overlap vector belongs to the cell.

[0117] In some embodiments, the color of a cell may be determined by the number of overlapping vectors in the cell, and a density map of the overlapping vectors may be drawn.

[0118] In this way, two matching point sets placed on different planes in three-dimensional space are converted into a density map of overlapping vectors in two-dimensional space. Compared with the symmetrical point image, the different density maps corresponding to fuel cells in different states have stronger regularity and higher state separability.

[0119] SB6. Calculate the GIST features of the density map and determine the GIST features as target features.

[0120] The present invention converts a two-dimensional density map into a one-dimensional GIST feature by calculating the global information feature (Global Image Spatial Texture, GIST) of the density map, while retaining the effective information in the density map, making it easier for the classifier model to process the target feature and accelerating the processing speed of the classifier model.

[0121] In some embodiments, first, the density map is converted into a grayscale image. I '' ( x , y ), and then for the grayscale image I '' ( x , y ) Apply the Gabor filter to get the image I ''' ( x , y ), as shown in the following formula.

[0122]

[0123] Among them, G( x , y ) is the Gabor operator.

[0124] Getting the image I ''' ( x , y ), the image is subsampled to reduce the amount of data and extract key features, the image is divided into smaller, non-overlapping blocks, the average value of each block is calculated to achieve feature dimensionality reduction, and pixel compression is applied to obtain an image of size 28 pixels x 28 pixels.

[0125] Finally, the 2D image is converted into a feature vector of 1 row and 512 columns by calculating the GIST features of an image of size 28 pixels x 28 pixels.

[0126] S104: Input the target feature into the classifier model to determine the state of the fuel cell.

[0127] The present invention performs feature extraction on multiple voltage response components obtained after decomposing the voltage response signal in the EIS signal to obtain the target feature of the voltage response signal. The target feature contains multi-scale information of the voltage response signal, provides richer information for the classifier model to determine the state of the fuel cell, and improves the accuracy of fuel cell state identification.

[0128] As a possible implementation, the target features are input into an active semi-supervised deep rule classifier (ASSDRB) to determine the state of the fuel cell.

[0129] In the prior art, the semi-supervised deep rule-based classifier (SSDRB) can actively learn new categories from unlabeled images. In industrial scenarios with fewer labeled images, it can start from a small number of labeled training images and autonomously evolve through the concept of "pseudo-labels" to process real-time data online. In view of the remarkable achievements of SSDRB in processing texture images, the present invention further studies and analyzes SSDRB.

[0130] First, use a labeled training dataset D labeled Initialize the deep rule-based classifier (DRB) to obtain the initialized DRB; then, DRB is used in the unlabeled training dataset D (1) Next, we use the ASSDRB function to enable SSDRB to actively learn the remaining unlabeled training data set D in offline mode. (2) Finally, we can use the unlabeled test dataset D unlabeled Test the classification performance of ASSDRB.

[0131] Through the above steps, ASSDRB can continuously update and evolve itself without retraining from scratch, thereby improving learning efficiency and adapting to new data distributions.

[0132] As a possible implementation method, the target feature is input into a classifier model to determine whether the fuel cell is in a normal state or a gain state.

[0133] The fuel cell of the present invention has better performance in the gain state than in the normal state.

[0134] The present invention provides a proton exchange membrane fuel cell state monitoring method, which is formed by electrochemical impedance spectroscopy, signal decomposition, symmetric point pattern transformation, three-dimensional reconstruction and perspective conversion, and visualizes the data processing workflow. The method provided by the present invention is described below through a specific example.

[0135] Under certain conditions, a current interference signal is introduced into the fuel cell, and the corresponding voltage response signal is collected over time. The voltage response signal is decomposed by EMD, and all IMFs are iteratively obtained. The injected current interference signal triggers the multi-scale IMF response of the fuel cell system, revealing the signal behavior characteristics at various frequency scales.

[0136] All obtained IMFs are transformed using SDP, and the signal behavior characteristics of the voltage response component at each scale are converted into geometric space to obtain the SDP image, which intuitively represents the state of the fuel cell system in the geometric space.

[0137] Figure 2 A schematic diagram of a voltage response signal of a fuel cell in a normal state provided by the present invention, Figure 3 A schematic diagram of a voltage response component of a fuel cell under normal conditions provided by the present invention, such as Figure 3 As shown, the voltage response signal of the fuel cell in a normal state can be decomposed into six IMFs, namely IMF1, IMF2, IMF3, IMF4, IMF5 and IMF6, and a residual.

[0138] Figure 4 The present invention provides a SDP image of a voltage response component of a fuel cell in a normal state, such as Figure 4 As shown, IMF1-SDP, IMF2-SDP, IMF3-SDP, IMF4-SDP, IMF5-SDP and IMF6-SDP are the SDP images corresponding to IMF1, IMF2, IMF3, IMF4, IMF5 and IMF6 respectively.

[0139] The multiple IMFs obtained by decomposing the voltage response signal can reflect the voltage response of the fuel cell under different states from multiple scales.

[0140] Figure 5 The SDP images of the six voltage response components of the fuel cell under normal conditions obtained by using different numbers of mirror symmetry planes are shown in Figure 2. Figure 5 Shown are six SDP images of the voltage response component IMF1 when the mirror symmetry plane NMSP is 1, 2, 3, 4, 5 and 6, respectively, and SDP images of other voltage response components IMF2, IMF3, IMF4, IMF5 and IMF6 when the mirror symmetry plane NMSP is 1, 2, 3, 4, 5 and 6, respectively.

[0141] Figure 6 A flowchart of a method for obtaining a matching point set of an SDP image provided by the present invention is shown in FIG. Figure 6 As shown, the SDP image is converted into a grayscale image, and the grayscale image is subjected to contrast enhancement, medium filtering and noise reduction, and sharpening using the Laplace operator. The feature points in the image are detected through SIFT, and the matching point set in the SDP image is obtained by comparing feature symbols, nearest neighbor search, ratio test and uniqueness constraint.

[0142] Figure 7 A schematic diagram of a matching point set in an SDP image provided by the present invention, such as Figure 7As shown in FIG. 1 , taking the SDP image of IMF1 when NMSP=1 as an example, the left and right sides are two identical SDP images, the points in the image are SIFT feature points, and the two feature points connected by a straight line are a set of feature points that match each other.

[0143] Figure 8 A three-dimensional schematic diagram of placing the matching point sets in the SDP image on two different planes, as shown in Figure 8 As shown, the matching point set of one of the two identical SDP images is placed on a first plane, and the matching point set of the other image is placed on a second plane. The first plane and the second plane are different and are both parallel to the XY plane.

[0144] like Figure 8 As shown, a group of matching point lines that match each other constitute a matching point vector, the reference vector is a vector with an angle of 45 degrees with the XY plane, and the absolute value of the dot product of the matching point vector and the reference vector is calculated. When the absolute value is greater than a preset length threshold, the matching point vector is determined to be an overlapping vector.

[0145] Fig. 9 A density map of overlapping vectors provided by the present invention, such as Fig. 9 As shown, a two-dimensional grid parallel to the XY plane is constructed, the cell to which the overlapping vector belongs is determined according to the XY coordinates of a matching point in the overlapping vector, the color of the cell is determined according to the number of overlapping vectors in the cell, and a density map is generated.

[0146] After retaining local features and performing perspective transformation to obtain a density map, we focus on global features. After converting the density map into a grayscale image of size 28 pixels × 28 pixels, we extract the GIST features of the density map, such as Fig.10 The density map with a size of 28 pixels × 28 pixels provided by the present invention is shown as follows: Fig.11 Shown is a schematic diagram of a GIST feature of a density map provided by the present invention.

[0147] The GIST features corresponding to each IMF of the voltage response signal obtained through the above steps are input into ASSDRB for classification to determine whether the fuel cell is in a gain state or a normal state.

[0148] ASSDRB can be trained and optimized by following the steps below:

[0149] S1. Initialize the DRB classifier using the labeled training dataset: set (Mode='Learning') and execute the DRB function to train the DRB classifier.

[0150] S2. Apply ASSDRB classifier for active learning: Set and adjust the semi-supervised learning threshold ( Φ) and the active learning threshold ( Γ ), which enables the classifier to actively learn from unlabeled training datasets in a semi-supervised mode, continuously updating and optimizing the performance of the classifier.

[0151] Fig.12 A schematic diagram of classification accuracy of fuel cell status monitoring by different methods provided by the present invention, such as Fig.12 As shown in the figure, in the experimental phenomenon that the fuel cell is cut off from oxygen for 20s and the performance gain recovers by 6%, the IMF of the voltage response signal is obtained by EIS and EMD, and the one-dimensional IMF of the voltage response signal is classified by decision tree, with a classification accuracy of 61.22%; the IMF of the voltage response signal is obtained by EIS and EMD, and the one-dimensional IMF of the voltage response signal is classified by support vector machine (SVM), with a classification accuracy of 61.43%; the IMF of the voltage response signal is obtained by EIS and EMD, and the one-dimensional IMF of the voltage response signal is classified by k-nearest neighbor (KNN), with a classification accuracy of 61.20%; the IMF of the voltage response signal is obtained by EIS and EMD, and the IMF of the voltage response signal is classified by logistic regression (Logistic Regression). The IMF of the one-dimensional voltage response signal was classified by the Naive Bayes algorithm (NaiveBayes), and the classification accuracy was 61.03%.

[0152] On the basis of obtaining the IMF of the voltage response signal through EIS and EMD, the target features of the voltage response signal are obtained by obtaining the SDP image through the above-mentioned symmetrical point pattern, determining the scale-invariant features in the SDP image, determining the set of matching points in the scale-invariant features, determining the overlapping vectors in the matching point vectors, drawing the density map and extracting the GIST features of the density map, that is, obtaining the target features through the Framework method, and obtaining the target features through the Decision tree classifies the target features with an accuracy of 84.31%; after obtaining the target features through the Framework method, the target features are classified by SVM with an accuracy of 86.10%; after obtaining the target features through the Framework method, the target features are classified by KNN with an accuracy of 85.20%; after obtaining the target features through the Framework method, the target features are classified by logistic regression with an accuracy of 84.28%; after obtaining the target features through the Framework method, the target features are classified by naive Bayes algorithm with an accuracy of 84.92%; after obtaining the target features through the Framework method, the target features are classified by SSDRB with an accuracy of 87.5%; after obtaining the target features through the Framework method, the target features are classified by ASSDRB with an accuracy of 96.88%.

[0153] It can be seen that multiple IMFs are obtained by combining electrochemical impedance spectroscopy with signal decomposition, which can reflect the multi-scale information of the voltage response signal. The target features are obtained using the above-mentioned Framework method, and combined with ASSDRB, a higher classification accuracy can be achieved.

[0154] Fig.13 A confusion matrix diagram provided by the present invention is as follows: Fig.13 As shown in the figure, after the target features are obtained by the Framework method, the target features are classified by ASSDRB. The predicted value corresponds to the determination result of ASSDRB, and the actual value corresponds to the actual state of the fuel cell. The actual value and the predicted value of 1 represent the normal state, and the actual value and the predicted value of 2 represent the gain state. The normal state is positive, and the gain state is negative. The true positive rate is 95.5%, the true negative rate is 100%, the false positive rate is 0%, and the false negative rate is 4.5%.

[0155] The present invention adopts the EMD method of EIS to analyze the voltage response component, provides a proton exchange membrane fuel cell state monitoring, improves the accuracy of fuel cell state identification, and provides more accurate data support for health monitoring and fault diagnosis of proton exchange membrane fuel cells.

[0156] Corresponding to the above-mentioned embodiment, the present invention further provides a proton exchange membrane fuel cell state monitoring device. Fig.14 A structural schematic diagram of a proton exchange membrane fuel cell state monitoring device provided in an embodiment of the present invention, the device may include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, so that the at least one processor can execute a proton exchange membrane fuel cell state monitoring method as described in the above embodiment.

[0157] In a possible implementation of the present invention, the at least one processor is capable of executing the following steps: obtaining a voltage response signal after a current interference signal is injected into the fuel cell; decomposing the voltage response signal into a plurality of voltage response components; performing feature extraction on any voltage response component of the voltage response signal to obtain a target feature corresponding to the voltage response signal; and inputting the target feature into a classifier model to determine the state of the fuel cell.

[0158] In a specific implementation, the present invention can also be provided as a non-volatile computer storage medium, wherein the storage medium is a non-volatile computer-readable storage medium, and the non-volatile computer-readable storage medium stores at least one program, each program includes instructions, and when the instructions are executed by a terminal, the terminal executes a proton exchange membrane fuel cell state monitoring method as described in the above embodiment.

[0159] In a possible implementation of the present invention, the aforementioned terminal executes, obtaining a voltage response signal after the fuel cell is injected with a current interference signal; decomposing the voltage response signal into multiple voltage response components; performing feature extraction on any voltage response component of the voltage response signal to obtain a target feature corresponding to the voltage response signal; and inputting the target feature into a classifier model to determine the state of the fuel cell.

[0160] The various embodiments of the present invention are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0161] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0162] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A proton exchange membrane fuel cell state monitoring method, characterized in that: The method comprises: Acquire a voltage response signal of the fuel cell after the current interference signal is injected; decomposing the voltage response signal into a plurality of voltage response components; For any voltage response component of the voltage response signal, feature extraction is performed to obtain a target feature corresponding to the voltage response signal, including: converting any voltage response component of the voltage response signal into a symmetrical point image corresponding to any voltage response component through a symmetrical point pattern; extracting the scale-invariant feature of the symmetrical point image through a scale-invariant feature transformation; copying the symmetrical point image to obtain two identical symmetrical point images, and determining whether the scale-invariant features in the two identical symmetrical point images match based on the distance between the scale-invariant features in the two identical symmetrical point images to obtain a matching point set; placing the matching point sets on two different planes respectively, connecting the matching points that match each other, and obtaining a matching point vector corresponding to the matching point; determining whether the matching point vector overlaps with the reference vector based on the projection of the matching point vector on the reference vector, and determining the matching point vector that overlaps with the reference vector as an overlapping vector; constructing a two-dimensional grid composed of multiple cells, determining the cell to which the overlapping vector belongs, and drawing a density map of the overlapping vector based on the number of overlapping vectors in the cell; calculating the GIST feature of the density map, and determining the GIST feature as a target feature; The target feature is input into a classifier model to determine the state of the fuel cell.

2. According to the method for monitoring the state of a proton exchange membrane fuel cell in claim 1, before obtaining the voltage response signal of the fuel cell after the current interference signal is injected, the method further comprises: The fuel cell is subjected to oxygen deprivation treatment so that the fuel cell remains in an oxygen deprivation state for a preset time period.

3. A proton exchange membrane fuel cell state monitoring method according to claim 1, characterized in that: Decomposing the voltage response signal into a plurality of voltage response components, including: The voltage response signal is decomposed into a plurality of intrinsic mode components by empirical mode decomposition, and the intrinsic mode components are determined as voltage response components.

4. A proton exchange membrane fuel cell state monitoring method according to claim 1, characterized in that: Converting any voltage response component of the voltage response signal into a symmetrical point image corresponding to any voltage response component through a symmetrical point pattern, including: Based on at least two different numbers of mirror symmetry planes, any voltage response component is converted into at least two different symmetry point images through a symmetry point pattern.

5. A proton exchange membrane fuel cell state monitoring method according to claim 1, characterized in that: Inputting the target feature into a classifier model to determine the state of the fuel cell includes: The target characteristics are input into the ASSDRB to determine the status of the fuel cell.

6. A proton exchange membrane fuel cell state monitoring method according to claim 1, characterized in that: Inputting the target feature into a classifier model to determine the state of the fuel cell includes: The target feature is input into a classifier model to determine whether the fuel cell is in a normal state or a gain state.

7. A proton exchange membrane fuel cell state monitoring device, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, so that the at least one processor can execute a proton exchange membrane fuel cell state monitoring method according to any one of claims 1-6.

8. A non-volatile computer storage medium, characterized in that: The storage medium is a non-volatile computer-readable storage medium, which stores at least one program, each of which includes instructions, and when the instructions are executed by the terminal, the terminal executes a proton exchange membrane fuel cell state monitoring method according to any one of claims 1-6.

Images