Method and system for battery state of health monitoring of an uninterruptible power supply

Abstract

The application relates to a method for monitoring the state of health of a battery of an uninterruptible power supply, comprising: collecting an infrared image and basic attribute data of the battery of the uninterruptible power supply; performing wavelet packet transformation on voltage and current data in the basic attribute data to generate a time domain graph, and fusing the infrared image; inputting the fused image and the basic attribute data into a bypass free attention classifier and a support vector machine classifier based on active learning for training, and outputting classification results; inputting the classification results of the bypass free attention classifier and the support vector machine classifier based on active learning into a multi-modal fusion classifier for fusion prediction; testing the multi-modal fusion classifier with a test set, taking the multi-modal fusion classifier meeting performance requirements as a real-time battery state of health monitoring deployment model configured by the uninterruptible power supply; obtaining a final battery state of health prediction, and also providing an uninterruptible power supply battery state of health monitoring system using the above method.

Description

Technical Field

[0001] This invention relates to the field of power electronics, and in particular to a method and system for monitoring the battery health status of an uninterruptible power supply. Background Technology

[0002] An uninterruptible power supply (UPS) system mainly consists of batteries, rectifiers, inverters, and control switches. It utilizes batteries to transmit electrical energy and, through rectifier and inverter circuit modules, converts direct current (DC) into alternating current (AC), providing uninterrupted and stable power to critical power electronic equipment. UPS systems are widely used in industries such as manufacturing, transportation, and healthcare; in the era of big data, they also play a vital role in ensuring information security.

[0003] In existing technologies, traditional manual maintenance and post-event operation and maintenance methods have drawbacks such as lag and high cost for UPS battery health status monitoring. They require a large amount of manually labeled data to improve the accuracy of the detectors, and the accuracy performance of single-mode detectors is limited, which cannot meet the needs of large institutions, especially hospitals, for UPS power supply health monitoring. Summary of the Invention

[0004] This application provides a method and system for monitoring the battery health status of an uninterruptible power supply, so as to provide early operation and maintenance support and to monitor the battery health status in real time with high accuracy.

[0005] On the one hand, this application provides a method for monitoring the battery health status of an uninterruptible power supply, including:

[0006] Step 101: Collect infrared images and basic attribute data of the uninterruptible power supply battery, and filter and process the infrared images and basic attribute data.

[0007] Step 102: Perform wavelet packet transform on the voltage and current data in the basic attribute data to generate a time-domain graph, fuse it with the infrared image to generate a fused image, and perform operation division on the fused image;

[0008] Step 103: Input the time-correlated fused images and basic attribute data from the training set into the side-channel free attention classifier and the active learning-based support vector machine classifier for training, and output the prediction results respectively.

[0009] Step 104: Input the prediction results of the side-channel free attention classifier and the active learning-based support vector machine classifier into the multimodal fusion classifier for fusion prediction;

[0010] Step 105: Test the multimodal fusion classifier on the test set. If it meets the performance requirements, stop training and use the multimodal fusion classifier that meets the performance requirements as the deployment model for real-time battery health status monitoring in the uninterruptible power supply configuration.

[0011] Step 106 yields the final battery health status prediction.

[0012] Furthermore, step 105 also includes: if the performance requirements are not met, a query is performed to select suitable unlabeled data, the labels are added to the training set, and the process returns to step 103.

[0013] Furthermore, the infrared images and basic attribute data of the batteries are filtered according to the ratio of "healthy battery: battery to be replaced = 1:1". The infrared images are uniformly cropped to the selected size, and the basic attribute data are associated by time. Data of the same time state are recorded as a group of basic attributes.

[0014] Furthermore, each set of basic attributes is associated with the fused image twice over time and recorded as a set of input data. For multiple sets of input data, they are divided according to the ratio of "initial training data: backup training data: test data = 5:3:2", and the battery health status is labeled for the initial training data and the test data.

[0015] Furthermore, the side-channel free attention classifier includes a side-channel resolution fusion-backbone feature extraction network and a channel fusion-head feature utilization network. In step 103, the fused image is input into the side-channel free attention classifier to perform the following steps:

[0016] The fused image is input into the side-channel resolution fusion-backbone feature extraction network. The backbone network extracts feature maps through multiple convolution-regularization-nonlinear modules, and then combines them with the resolution feature maps introduced from the side-channel network through convolution-regularization-nonlinear modules and fusion modules. After being concatenated and fused according to a specific dimension, the image is input into the side-channel resolution fusion free attention module for resolution fusion, then through the max pooling convolution module for feature compression and extraction, and then through the image feature information enhancement module for further feature extraction. Finally, the image is input into the location embedding module and the self-attention encoder module for feature refinement.

[0017] The refined feature map is input into the channel fusion-head feature utilization network, which first performs resolution enlargement and dimension reduction, and then performs resolution reduction and dimension enlargement operations. At the same time, the feature maps with different numbers of channels are stitched and fused, and then input into the channel fusion free attention module for free attention, and input into the location embedding module and self-attention encoder module for feature refinement.

[0018] The feature map after further feature refinement is input into the reparameterized convolution module for further feature refinement.

[0019] Output the classification results.

[0020] Further, in step 103, the basic attribute data is input into an active learning-based support vector machine classifier to perform the following steps:

[0021] Hyperparameters are optimized using a grid search algorithm, and classification results are output.

[0022] Furthermore, in step 105, the multimodal fusion classifier is a convolutional neural network self-learning weight fusion classifier, which obtains the final battery state prediction by autonomously learning the most suitable weight allocation from the classification results of the side-channel free attention classifier and the active learning-based support vector machine classifier.

[0023] Furthermore, in step 102, the wavelet packet transform involves selecting a cubic spline wavelet as the wavelet basis function to transform the voltage and current data X in the basic attribute data. m The frequency band is decomposed into P layers, and then reconstructed. The formula for calculating the time-frequency matrix composed of all sub-bands belonging to the P layers is as follows:

[0024]

[0025] in, For the voltage and current data X in the basic attribute data m The time-frequency matrix is ​​given by m, where m is the number of sampling points and P is the number of wavelet packet decomposition layers. For the second layer of layer P P The value of each sub-band at the m-th sampling point

[0026] The time-frequency matrices of voltage and current in the basic attributes are arranged from left to right to form a new matrix M:

[0027]

[0028] in, The time-frequency matrix represents the voltage U in the basic properties. The time-frequency matrix represents the current I in the basic properties.

[0029] Transform matrix M into a time-domain graph.

[0030] Furthermore, in step 105, if the performance requirements are not met, the unlabeled battery infrared images and basic attribute data are calculated using the joint maximum uncertainty algorithm. The joint maximum uncertainty values ​​are then sorted, and the top Q data points are labeled and added to the training set for training.

[0031] The formula for calculating the uncertainty value of the basic battery attribute data is as follows:

[0032]

[0033] Wherein, H(x) i The ) represents the uncertain value of the basic attribute. The possible label for each group of battery basic attribute data is predicted by an active learning-based support vector machine classifier from all backup training basic attribute data. i N refers to a set of backup training basic attribute data, w refers to the specific category of battery health status predicted by the active learning-based support vector machine classifier, and N refers to the battery health status. i (1≤N i ≤2) is the backup training basic attribute data x i The corresponding class predicted by the active learning-based support vector machine classifier has two possibilities: "healthy" and "to be replaced". Represents the basic attribute data x for predicting backup training. i belong The possibility, It is the backup training basic attribute data x i The predicted value,

[0034] For battery infrared images, the formula for calculating image uncertainty is as follows:

[0035]

[0036]

[0037] Where, m ij It is a pixel value in the infrared image of the backup training battery. Represents N picture Number of images, N for each image pixel Sum of pixel values This represents the average value of its pixels. This represents the value m of each pixel in an infrared image of a backup training battery. i and The sum of the offset values ​​is denoted as the infrared image uncertainty value H(m). i ),

[0038] The calculation formula for the joint-maximum uncertainty algorithm is as follows:

[0039]

[0040] in, Indicates the joint maximum uncertainty. This represents sorting the backup training basic attribute data according to uncertain values, sort{H(m i )} represents sorting the infrared image data of the backup training battery according to the uncertainty value. This represents selecting a set of input data with the largest joint uncertainty.

[0041] On the other hand, this application provides a system for monitoring the battery health status of an uninterruptible power supply (UPS), comprising: a data acquisition device for real-time acquisition of infrared images and basic attribute data of the UPS battery; a memory for real-time storage of the acquired infrared images and basic attribute data of the battery, as well as a battery health status monitoring program; a processor for executing the battery health status monitoring program; a cloud database for storing battery health status predictions obtained by the processor; and an alarm device for activating an alarm when the battery health status is detected as "to be replaced".

[0042] As can be seen from the technical solution provided in this application, a self-designed bypass free attention classifier and a support vector machine classifier based on active learning are trained using fused images and basic battery attributes, respectively. Multimodal classifier fusion is performed, and active learning is introduced. Data is selected by calculating according to the joint maximum uncertainty algorithm, which makes the classifier perform well, supports real-time data input, realizes real-time battery health status monitoring, and overcomes the defects of existing uninterruptible power supply battery health status monitoring. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a method for monitoring the battery health status of an uninterruptible power supply provided in an embodiment of this application;

[0045] Figure 2 This application provides an embodiment for... Figure 1 A schematic diagram of the active learning method in the battery health status monitoring method of uninterruptible power supply;

[0046] Figure 3 This application provides an embodiment for... Figure 2 A schematic diagram of the bypass free attention classifier network structure for a battery health status monitoring method for uninterruptible power supplies.

[0047] Figure 4 This application provides an embodiment for... Figure 2 A schematic diagram of the structure of a support vector machine classifier based on active learning for monitoring the battery health status of an uninterruptible power supply.

[0048] Figure 5This application provides an embodiment for... Figure 2 A schematic diagram of the structure of a multimodal classification fusion device for monitoring the battery health status of an uninterruptible power supply.

[0049] Figure 6 This is a schematic diagram of the structure of the uninterruptible power supply battery health status monitoring system provided in the embodiments of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] This application proposes a method for monitoring the battery health status of an uninterruptible power supply, as shown in the attached figure. Figure 1 As shown, the main steps are as follows, which are described in detail below:

[0052] Step S101: Collect infrared images and basic attribute data of the batteries in the hospital's uninterruptible power supply configuration, and perform relevant data screening and processing.

[0053] Infrared images of the batteries in the hospital's uninterruptible power supply (UPS) configuration are captured in real-time by a battery infrared detection device placed inside the UPS. Basic attribute data includes battery model, current, voltage, temperature, internal resistance, and usage time; this data is captured in real-time. Related data screening and processing involves filtering battery infrared images and basic attribute data according to a "healthy battery: battery to be replaced = 1:1" ratio. The infrared images are uniformly cropped to 640×640 pixels. Basic attribute data are correlated by time; data from the same time period are grouped into one set of basic attributes. "Batteries to be replaced" includes any scenario requiring battery replacement, such as decreased battery capacity or battery malfunction. The "1:1" ratio refers to the data selection proportion, meaning the amount of healthy battery data and the amount of battery data to be replaced are the same after screening. This aims to ensure a sufficient number of negative samples, increasing the model's generalization ability and ensuring that negative sample data and normal sample data have the same order of magnitude.

[0054] Step S102: Perform wavelet packet transform on the voltage and current data in the basic attribute data to generate a time-domain graph, fuse it with the infrared image, and then perform relevant data operations to divide it.

[0055] Wavelet packet transform is performed on the voltage and current data in each set of basic attributes to generate a 640×640 small-domain image, which is then fused with the battery infrared image from the same time. Here, "fusion" means that the images are directly superimposed according to the dimensions to retain information to the greatest extent. Each set of basic attributes and the fused image are then associated twice over time and recorded as a set of input data. For multiple sets of input data, based on practical experimental experience, the data is divided according to the ratio of "initial training data: backup training data: test data = 5:3:2". Battery health status is labeled on the initial training data and the test data. This division ratio ensures both sufficient training data and sufficient fair representativeness of the test data.

[0056] The aforementioned wavelet packet transform: After multiple performance tests, cubic spline wavelet was selected as the wavelet basis function for the voltage and current data X in the basic attributes. m The frequency band is decomposed into P layers, and then reconstructed. The formula for calculating the time-frequency matrix composed of all sub-bands belonging to the P layers is as follows:

[0057]

[0058] in, For the voltage and current data X in the basic attributes m The time-frequency matrix is ​​given by m, where m is the number of sampling points and P is the number of wavelet packet decomposition layers. For the second layer of layer P P The value of each sub-band at the m-th sampling point.

[0059] The time-frequency matrices of voltage and current in the basic attributes are arranged from left to right to form a new matrix M:

[0060]

[0061] in, The time-frequency matrix represents the voltage components of the basic properties. The time-frequency matrix represents the current in the basic properties.

[0062] Convert matrix M into a 640×640 small-scale grayscale image.

[0063] Step S103: Input the fused images and basic attributes from the training set into the side-channel free-attention classifier and the active learning-Support vector machine (AL-SVM) classifier, respectively.

[0064] The fused images and basic attributes associated by time are respectively input into an off-path free-attention classifier designed by an autonomous structure and an active learning-based support vector machine (AL-SVM) classifier for training.

[0065] As attached Figure 3 As shown, the self-designed bypass free-attention classifier has the following characteristics:

[0066] The network is divided into two parts: a bypass resolution fusion-backbone feature extraction network and a channel fusion-head feature utilization network.

[0067] The traditional backbone feature extraction network is supplemented by a CBS bypass network. First, the CBS module (CBS: Con2d-BatchNorm-SiLu) performs convolution-regularization-nonlinear operation to extract and expand the dimensions of features. Then, the max-pooling convolution MPConv (Max-Pooling Conv) module performs feature compression and extraction. Next, the CBS module and the concat module perform concatenation and fusion operations to concatenate the features along specific dimensions. After resolution fusion, the features are input into the bypass resolution fusion free-attention module. This module increases the overall feature dimension by allowing features to be biased towards three aggregation dimensions, thereby improving feature extraction capability. Finally, the features are input into the image feature information enhancement SPP-CSP (Spatial Pyramid Pooling-Cross Stage Partial) module. This module enhances cross-stage information by performing spatial pyramid pooling on the features.

[0068] A position embedding module and a self-attention encoder module are added to the upper layers of the backbone feature extraction network and the head feature utilization network, respectively. The position embedding module feeds position information into the network, increasing the network's sensitivity to feature positions, while the self-attention encoder module makes the network pay more attention to the "points it should pay attention to", thereby improving the network's feature utilization ability.

[0069] A channel fusion free attention module was added to the head feature utilization network layer, and then the features were input into the reparameterized convolution (RepConv) module for further refinement of feature effects. Considering the subtle image changes in battery health status, related experiments showed that a reasonable balance between network accuracy and processing speed was achieved when the task mapping output head—"Battery Health Status Monitoring"—was set to 4. Therefore, the output head of the head feature utilization network layer was set to 4 to further improve the battery health status monitoring capability.

[0070] like Figure 4 The diagram shown illustrates the structure of an Active Learning-based Support Vector Machine (AL-SVM) classifier. The structure of the Active Learning "AL" can be found in the reference diagram. Figure 2 . Figure 4 After the basic properties of the battery are input into the active learning-based support vector machine (AL-SVM), the active learning-based support vector machine optimizes the hyperparameters through a grid search algorithm and then outputs the results. Figure 2 In the process, if the results of the support vector machine based on active learning are fused with the results of the side-channel resolution free-attention classifier through convolutional self-learning weights (i.e., multimodal classifier fusion) for prediction, and the performance requirements are not met, active learning is performed (i.e., through the joint-maximum uncertainty algorithm, see below) to further improve the classifier performance.

[0071] like Figure 5 The diagram shows a multimodal fusion classifier. The classification results of the side-channel free-attention classifier and the active learning-based support vector machine (AL-SVM) classifier are input into this structure. This structure uses a convolutional self-learning weight fusion method to obtain the most suitable weights and thus the final prediction result.

[0072] The following is in conjunction with the appendix Figure 3 The process of processing fused images by the self-designed bypass free attention classifier is described in detail:

[0073] The fused image first passes through a side-channel resolution fusion-backbone feature extraction network: the image is processed into two branches. The backbone network first extracts a 160×160×128 feature map through four CBS layers. This feature map is then CBS-operated with the resolution feature map from the side-channel network and concatenated by the fusion module. Finally, it is input into the side-channel resolution fusion free-attention module for free attention. The side-channel resolution fusion free-attention module is a variant of the efficient self-attention Transformer module. Traditional Transformer modules make the network focus more on the "points it should focus on," making feature aggregation manifest in dimensions that have practical significance. However, the spatial dimension of its focus points is too complex, and stacking modules multiple times will make the network too large and impractical. The free-attention mechanism allows the network to freely learn the points it wants to focus on. By allowing features to be biased towards three aggregation dimensions, it increases the overall feature dimension and reduces the spatial dimension of the focus points, while reducing the computational cost and meeting real-time requirements.

[0074] As the backbone network deepens, the feature map resolution continuously decreases while the channel dimension continuously increases. After passing through four bypass resolution fusion free attention modules, it is connected to the image feature information enhancement SPP-CSP (Spatial PyramidPooling Cross Stage Partial) module for further feature extraction. Then, the feature map is input into the position embedding and self-attention encoder (Transformer Encoder) modules for feature refinement.

[0075] The refined feature maps are input into the channel fusion-head feature utilization network: the feature maps are processed through a pyramid-shaped network structure, first undergoing resolution enlargement and dimensionality reduction, then resolution reduction and dimensionality enlargement. Simultaneously, feature maps with different numbers of channels are fused by the fusion module (Concat) and then input into the channel fusion free-attention module for free attention. By allowing features to be biased towards three aggregation dimensions, the overall feature dimension is increased. At the top of the pyramid-shaped network structure, position embedding and self-attention encoder modules are added for feature refinement.

[0076] Considering the subtle image changes in battery health status, the head features are set to 4 in the network layer output head - "Battery Health Status Monitoring" to further improve the battery health status monitoring capability.

[0077] Step S104: Fuse the classification results of the two classifiers for learning, referring to... Figure 5 A schematic diagram of a multimodal fusion classifier, including a self-learning weight fusion unit of a Convolutional Neural Network (CNN) performing the corresponding... Figure 2 The fusion operation of the multimodal fusion classifier combines the classification results of the side-channel free attention classifier and the active learning-based support vector machine AL-SVM classifier with the most suitable weight allocation through autonomous learning to obtain the final battery health status prediction.

[0078] Step S105: Test the trained multimodal fusion classifier on a test set. If the performance requirements are met, stop training. If not, calculate the joint maximum uncertainty using a self-designed algorithm for unlabeled battery infrared images and basic attribute data. Sort the joint maximum uncertainty values, select the top 20 data points, label them, and add them to the training set for training. Skip to step S103 and repeat the process until the performance requirements are met. Here, "self-designed" refers to a newly proposed image uncertainty calculation formula (described in detail below) based on actual project needs, and its joint calculation with other existing uncertainty algorithms. For basic attribute data, the uncertainty calculation formula is as follows:

[0079]

[0080] The fusion classifier is obtained in step S104, and the possible labels for each group of battery basic attribute data are predicted by the multimodal fusion classifier from all backup training basic attribute data. i N refers to a set of backup training basic attribute data, w refers to the specific category of battery health status predicted by the fusion classifier, and N refers to the battery health status. i (1≤N i ≤2) is the backup training basic attribute data x i The corresponding class predicted by the fusion classifier has two possibilities: "healthy" and "to be replaced". Represents the basic attribute data x for predicting backup training. i belong The possibility, It is the backup training basic attribute data x i The predicted value, H(x) i ) represents an uncertain value of a basic attribute.

[0081] For battery infrared images, the self-designed formula for calculating image uncertainty is as follows:

[0082]

[0083]

[0084] Where m ij It is a pixel value in the infrared image of the backup training battery. Represents N picture = 100 images, N for each image pixel = Sum of 640 x 640 pixel values This represents the average value of its pixels. This represents the value m of each pixel in an infrared image of a backup training battery. i and The sum of the offset values ​​is denoted as the infrared image uncertainty value H(m). i ).

[0085] The calculation formula of the self-designed joint-maximum uncertainty algorithm is as follows:

[0086]

[0087] in Indicates the joint maximum uncertainty. This represents sorting the backup training basic attribute data according to uncertain values, sort{H(m i )} represents sorting the infrared image data of the backup training battery according to the uncertainty value. This represents the selection of a set of input data with the largest joint uncertainty (i.e., backup training basic attribute data and backup training battery infrared image data).

[0088] Step S106: Use the fusion classifier that meets the performance requirements as the deployment model for real-time battery health status monitoring in hospital uninterruptible power supply configuration. Here, "real-time" means that the model acquires battery infrared images and basic attribute data in real time and provides real-time battery health status monitoring.

[0089] The present invention also provides a battery health status monitoring system for hospital uninterruptible power supply (UPS) configurations, as shown in the attached figure. Figure 6 As shown, the system mainly includes: a data acquisition device (such as an infrared detection device), a memory, a processor, a computer program stored in the memory and executable on the processor, a cloud database, and an alarm device, such as a program for monitoring the battery health status of an uninterruptible power supply (UPS). When the processor executes the computer program, it implements the steps for monitoring the battery health status of the UPS, for example... Figure 1 Steps S101 to S105 are shown below. The memory collects and stores infrared image data and basic attribute data of the uninterruptible power supply (UPS) battery in real time. The processor runs a computer program to determine the battery's health status. The cloud database stores the health status obtained by the processor in real time. When the battery's health status is detected as "needs replacement," an alarm device is activated.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this invention.

Claims

1. A method for monitoring the battery health status of an uninterruptible power supply, characterized in that, The method includes: Step 101: Collect the battery infrared image and basic attribute data of the uninterruptible power supply, and filter and process the infrared image and the basic attribute data; Step 102: Perform wavelet packet transform on the voltage and current data in the basic attribute data to generate a time-domain graph, fuse it with the infrared image to generate a fused image, and perform operation division on the fused image; Step 103: The fused image and the basic attribute data associated by time in the training set are input into a side-channel free attention classifier and an active learning-based support vector machine classifier for training, respectively, and the classification results are output. The side-channel free attention classifier includes a side-channel resolution fusion-backbone feature extraction network and a channel fusion-head feature utilization network. When the fused image is input into the side-channel free attention classifier, the following steps are performed: The fused image is input into the side-channel resolution fusion-backbone feature extraction network. The backbone feature extraction network extracts feature maps through multiple convolution-regularization-nonlinear modules, and then combines them with the resolution feature maps introduced from the side-channel network through convolution-regularization-nonlinear modules and a fusion module. After being concatenated and fused according to a specific dimension, the result is output. The features are fed into a bypass resolution fusion free attention module for free attention, then into a max pooling convolution module for feature compression and extraction, followed by image feature enhancement for further feature extraction. The extracted features are then fed into a location embedding module and a self-attention encoder module for feature refinement. The refined feature map is then fed into a channel fusion-head feature utilization network, where resolution is first increased and dimensionality reduced, followed by a second resolution reduction and dimensionality increase. Simultaneously, feature maps with different channel numbers are concatenated and fused by a fusion module, then fed into a channel fusion free attention module for free attention, followed by a second input to the location embedding module and self-attention encoder module for feature refinement. The further refined feature map is then fed into a reparameterized convolution module for further feature refinement. The classification result is then output. Step 104: Input the classification results of the side-channel free attention classifier and the active learning-based support vector machine classifier into the multimodal fusion classifier for fusion prediction. The multimodal fusion classifier is a convolutional neural network self-learning weight fusion classifier. The classification results of the side-channel free attention classifier and the active learning-based support vector machine classifier are assigned the most suitable weights through self-learning. Step 105: Test the multimodal fusion classifier on a test set. If the performance requirements are met, stop training and use the multimodal fusion classifier that meets the performance requirements as the deployment model for real-time battery health status monitoring of the uninterruptible power supply configuration. Step 106 yields the final battery health status prediction.

2. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 1, characterized in that, Step 105 also includes: if the performance requirements are not met, a query is performed to select suitable unlabeled data, the labels are added to the training set, and the process returns to step 103.

3. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 2, characterized in that, The infrared images and basic attribute data of the batteries are filtered according to the ratio of "healthy battery: battery to be replaced = 1:1". The infrared images are uniformly cropped to the selected size. The basic attribute data are associated by time. Data of the same time state are recorded as a group of basic attributes.

4. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 3, characterized in that, Each set of basic attributes is associated with the fused image twice over time and recorded as a set of input data. For multiple sets of input data, they are divided according to the formula "initial training data: backup training data: test data = 5:3:2". Battery health status is labeled for the initial training data and the test data.

5. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 4, characterized in that, In step 103, the basic attribute data is input into the active learning-based support vector machine classifier to perform the following steps: Hyperparameters are optimized using a grid search algorithm, and classification results are output.

6. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 1, characterized in that, The wavelet packet transform described in step 102: A cubic spline wavelet is selected as the wavelet basis function to transform the voltage and current data in the basic attribute data. The frequency band is decomposed into P layers, and then reconstructed. The formula for calculating the time-frequency matrix composed of all sub-bands belonging to the P layers is as follows: in, The voltage and current data in the basic attribute data. The time-frequency matrix, The number of sampling points. The wavelet packet decomposition level is [number of layers]. For the first Layer The sub-band is in the first The value of each sampling point Arrange the time-frequency matrices of voltage and current from left to right in the aforementioned basic attributes to form a new matrix. for: in, Voltage representing the basic attribute The time-frequency matrix formed Current representing the basic properties The time-frequency matrix formed matrix Convert it into the time domain diagram.

7. The method for monitoring the battery health status of an uninterruptible power supply as described in claim 1, characterized in that, In step 105, if the performance requirements are not met, the unlabeled battery infrared images and basic attribute data are calculated using the joint maximum uncertainty algorithm. The joint maximum uncertainty values ​​are then sorted, and the top Q data points are selected for labeling and added to the training set for training. The formula for calculating the uncertainty value of the battery's basic attribute data is as follows: in, Representing the uncertain values ​​of basic attributes, the possible labels for each group of battery basic attribute data are predicted by the active learning-based support vector machine classifier from all backup training basic attribute data. Refers to a set of basic attribute data for backup training. This refers to the specific category of battery health status predicted by the active learning-based support vector machine classifier. , It is the basic attribute data for backup training. The corresponding class predicted by the active learning-based support vector machine classifier has two possibilities: "healthy" and "to be replaced". This represents the basic attribute data for predicting backup training. belong The possibility, It is the basic attribute data for backup training. The predicted value, The formula for calculating the image uncertainty value of the battery infrared image is as follows: in, It is a pixel value in the infrared image of the backup training battery. Representative to Images, each image Sum of pixel values This represents the average value of its pixels. This represents the pixel value of an infrared image of a backup training battery. and The sum of the offset values ​​is denoted as the infrared image uncertainty value. , The calculation formula for the joint-maximum uncertainty algorithm is as follows: in, Indicates the joint maximum uncertainty. This represents sorting the backup training basic attribute data according to the uncertain values. This represents sorting the infrared image data of the backup training battery according to the uncertainty value. This represents selecting a set of input data with the largest joint uncertainty.

8. A system for monitoring the battery health status of an uninterruptible power supply, characterized in that, include: A data acquisition device is used to acquire infrared images and basic attribute data of the battery of the uninterruptible power supply in real time. The memory is used to store the acquired infrared images of the battery and the basic attribute data in real time, as well as the battery health status monitoring program. The processor executes the battery health status monitoring program to implement the steps of the method according to any one of claims 1 to 7; A cloud-based database stores the battery health status predictions obtained by the processor. The alarm device is activated when the battery health status is detected as "to be replaced".

Images