Modularized power supply health state monitoring method and system based on deep learning

Abstract

The invention discloses a modular power supply health state monitoring method and system based on deep learning, and relates to the field of power supply state monitoring, and the method comprises the steps: collecting the operation characteristics of a plurality of modules in a power supply in a monitoring time period, obtaining a plurality of module operation characteristic sets, and constructing an operation characteristic graph sequence; carrying out adjacency change identification on the operation feature map sequence to obtain an adjacency feature change value sequence, and carrying out feature change coupling analysis on a plurality of modules to obtain a feature change anomaly degree sequence; and according to the feature change abnormity degree sequence and the adjacent feature change value sequence, screening the operation feature map sequence to obtain a screened operation feature map set, and performing identification processing to obtain health state information as a monitoring result. According to the invention, the problems of inaccurate health monitoring and low monitoring efficiency in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of power condition monitoring, and more specifically to a modular power health status monitoring method and system based on deep learning. Background Technology

[0002] With the rapid development of power electronics technology, modular power supplies have been widely used in key fields such as communications, power, industrial control, and new energy. These power supplies are typically composed of multiple functional modules, such as AC / DC modules, DC / DC modules, battery management modules, and monitoring modules. Therefore, performance degradation or failure of any one module can affect the stable operation of the entire system.

[0003] Currently, monitoring methods for power systems mainly focus on monitoring the correctness of real-time operating parameters, failing to consider the data characteristics of modular power supplies and lacking analysis of the characteristic coupling relationships between multiple modules, resulting in inaccurate health monitoring and low monitoring efficiency. Summary of the Invention

[0004] This application provides a modular power supply health status monitoring method and system based on deep learning, which addresses the problems of inaccurate health monitoring and low monitoring efficiency caused by the lack of feature coupling relationship analysis between multiple modules in the prior art.

[0005] In view of the above problems, this application provides a modular power supply health status monitoring method and system based on deep learning.

[0006] In a first aspect, this application provides a modular power supply health status monitoring method based on deep learning, the method comprising: The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets and construct an operating characteristic map sequence; Adjacency change identification is performed on the running feature map sequence to obtain the adjacent feature change value sequence, and feature change coupling analysis of multiple modules is performed to obtain the feature change anomaly sequence; Based on the anomaly sequence of feature changes and the adjacent feature change value sequence, the running feature map sequence is filtered to obtain a filtered running feature map set. The health status information is then identified and processed as the monitoring result.

[0007] Secondly, the present invention provides a modular power supply health status monitoring system based on deep learning, the system comprising: The operation feature acquisition module is used to collect the operation features of multiple modules in the power supply during the monitoring time period, obtain multiple module operation feature sets, and construct an operation feature map sequence. The feature change analysis module is used to identify adjacent changes in the running feature map sequence, obtain the adjacent feature change value sequence, and perform feature change coupling analysis of multiple modules to obtain the feature change anomaly sequence. The health status identification module is used to filter the running feature map sequence based on the feature change anomaly sequence and the adjacent feature change value sequence to obtain a filtered running feature map set, and to identify and process it to obtain health status information as a monitoring result.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first collects the operational characteristics of multiple modules within the power supply during the monitoring period, constructs an operational feature map sequence, and converts the dispersed multi-source monitoring data into a unified image format, laying a high-quality data foundation. Second, it obtains the adjacent feature change value sequence through adjacency change identification, and simultaneously performs feature change coupling analysis to analyze the degree of abnormality in the coupling relationship between modules, obtaining feature changes in two dimensions and improving the accuracy of health status monitoring. Finally, it filters the operational feature map sequence, removes redundant information, and then obtains the final health status information through identification processing. This reduces the computational burden while ensuring monitoring accuracy, providing reliable technical support for the health status monitoring of modular power supplies. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the modular power supply health status monitoring method based on deep learning proposed in this application. Figure 2 This is a schematic diagram of the modular power supply health status monitoring system based on deep learning proposed in this application.

[0010] In the attached diagram, the components represented by each number are as follows: Run the feature acquisition module 11, feature change analysis module 12, and health status recognition module 13. Detailed Implementation

[0011] This application provides a modular power supply health status monitoring method based on deep learning, which specifically solves the problems of inaccurate health monitoring and low monitoring efficiency caused by the lack of feature coupling relationship analysis between multiple modules in the existing technology.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides a modular power supply health status monitoring method based on deep learning, the method comprising: S10: Collect the operating characteristics of multiple modules within the power supply during the monitoring time period, obtain the operating characteristic set of multiple modules, and construct the operating characteristic map sequence; Step S10 in the method provided in this application embodiment includes: The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets. Each module includes multiple operating characteristic categories, and the module operating characteristics of multiple operating characteristic categories are collected. Multiple module running feature sets are used as pixel values ​​of running feature maps to construct multiple running feature maps, resulting in a sequence of running feature maps.

[0014] In this embodiment, the operating characteristics of multiple modules within the power supply are first collected during a monitoring time period. Multiple operating characteristic categories are collected from each module, such as current, voltage, and time delay. Each module is sampled during the monitoring time period to obtain module operating characteristics for multiple operating characteristic categories, thus constructing multiple module operating characteristic sets. The monitoring time period can be one operating cycle of the power supply module.

[0015] Secondly, the operational features in the module operational feature set are normalized and converted into pixel values. These pixel values ​​can then be used to identify the corresponding operational features. For example, the first pixel in the top left corner of the image can be designated as representing the voltage of a certain module. The converted pixel values ​​are then used to construct the corresponding operational feature map. This same operation is performed on multiple collected module operational feature sets to generate multiple operational feature maps, resulting in a sequence of operational feature maps.

[0016] In step S10 of the method provided in this application embodiment, multiple module running feature sets are used as pixel values ​​of running feature maps to construct multiple running feature maps, resulting in a running feature map sequence, including: Based on multiple operational feature categories from multiple modules, multiple pixel sites within the operational feature map are constructed to obtain the operational feature map template; The running feature sets of multiple modules are normalized, and the normalized running feature value of each module is written into the pixel position corresponding to the running feature map template to obtain multiple running feature maps and obtain a running feature map sequence.

[0017] In this embodiment, firstly, based on multiple operational feature categories of multiple modules, corresponding pixel sites are assigned to each feature category within a module to obtain an operational feature map template. All subsequent monitoring data will be written according to the operational feature map template to obtain an operational feature map.

[0018] Secondly, the feature sets of multiple modules are normalized, and linear scaling is used to map the module features to a fixed numerical range. For example, the ratio of the voltage value to the maximum voltage value can be calculated, multiplied by 50, and rounded to the nearest integer, mapping it to the range of 0-50. All module features are processed in the same way to obtain the converted pixel values. Then, the processed module features are used as pixel values ​​of the image and written into the corresponding pixel locations of the feature map template, resulting in multiple feature maps and a sequence of feature maps.

[0019] In this embodiment, the operating characteristics of multiple modules within the power supply are collected during the monitoring time period. Subsequently, an operating feature map template is constructed, and the normalized module operating feature values ​​are written into the corresponding pixel sites to construct an operating feature map sequence. This process converts the dispersed multi-source monitoring data into a unified format, achieving data standardization and laying a high-quality data foundation for subsequent processing.

[0020] S20: Perform adjacency change identification on the running feature map sequence to obtain the adjacency feature change value sequence, and perform feature change coupling analysis of multiple modules to obtain the feature change anomaly sequence; In this embodiment, adjacency change identification is first performed, and the change values ​​of adjacent running feature maps are calculated to obtain a sequence of adjacent feature change values. Then, using each adjacent feature change value as a baseline, the feature change values ​​of other modules when it appeared in the historical time are retrieved, and then the difference between it and the current adjacent feature change value is calculated as the feature change anomaly degree.

[0021] Step S20 in the method provided in this application embodiment includes: Calculate the change value of all module running feature values ​​in each running feature map within the running feature map sequence and the next adjacent running feature map to obtain multiple sets of adjacent feature change values; Calculate the mean of multiple sets of adjacent feature change values ​​to obtain multiple adjacent feature change values, and arrange them to obtain a sequence of adjacent feature change values.

[0022] In this embodiment, firstly, since the running feature maps are constructed based on running feature map templates, the same running features are represented at the same pixel locations, with only differences in numerical value. Therefore, the change value of all module running feature values ​​in each running feature map within the running feature map sequence is calculated between each running feature map and the first pixel value in the next adjacent running feature map. That is, the change value of the first pixel value in the running feature map is calculated between the first pixel value in the next adjacent running feature map and the first pixel value in the next adjacent running feature map. This process is repeated until all pixel locations have been calculated, resulting in a set of adjacent feature change values. Subsequently, the change values ​​of all running feature maps are calculated to obtain multiple sets of adjacent feature change values.

[0023] For example, suppose there exists a certain operating feature map and its adjacent next operating feature map. There are three modules, each with three operating features: current, voltage, and time delay. The operating feature map is as follows: Module 1: (20,25,40), (23,29,44), (31,32,45). The adjacent next operating feature map sequence is: (23,31,42), (27,37,39), (29,35,42). Through adjacency change calculation, the set of adjacency feature change values ​​is [3,6,2,4,8,5,2,3,3]. Similarly, for N operating feature maps, N-1 sets of adjacency feature change values ​​can be obtained through calculation.

[0024] Secondly, the average value of the adjacent feature change values ​​in the adjacent feature change value sets obtained from two adjacent running feature maps is calculated to obtain the adjacent feature change value. A similar calculation is performed on all adjacent feature change value sets to obtain all adjacent feature change values. Then, they are arranged in chronological order to obtain the adjacent feature change value sequence.

[0025] For example, suppose there exists a certain running feature map and its adjacent next running feature map, with adjacent feature change values ​​of [3,6,2,4,8,5,2,3,3]. The resulting adjacent feature change values ​​are (3+6+2+4+8+5+2+3+3) / 9=4. Similarly, performing similar calculations on all sets of adjacent feature change values ​​yields a sequence of adjacent feature change values, for example: [3,5.6,3,4,4.5,3,4,5,5].

[0026] Step S20 in the method provided in this application embodiment further includes: Within multiple sets of adjacent feature change values, one adjacent feature change value is randomly selected as the baseline feature change value, resulting in multiple baseline feature change values. Based on multiple baseline feature change values, index the baseline feature change values ​​of other operational feature categories of other modules, and perform feature change coupling analysis of multiple modules to obtain a feature change anomaly sequence.

[0027] In this embodiment of the application, firstly, within a set of multiple adjacent feature change values, an adjacent feature change value is randomly selected with equal probability as a reference feature change value, thereby obtaining multiple reference feature change values.

[0028] Secondly, based on multiple baseline feature change values, the baseline feature change values ​​of other operating feature categories of other modules are indexed, and feature change coupling analysis of multiple modules is performed to obtain a feature change anomaly sequence. Among them, other operating feature categories of other modules are all operating feature categories other than the operating features corresponding to the baseline feature change values. The baseline feature change values ​​are used as the benchmark to determine whether the changes in other operating feature values ​​conform to historical patterns. The index uses the baseline feature change values ​​as keywords to search the historical monitoring database of the power supply to obtain the average change of other operating features when the same baseline feature change value has occurred in the past.

[0029] Coupling is an inherent, physical relationship between different characteristics. For example, under normal circumstances, a decrease in the output voltage of a module does not usually accompany a sharp increase in its own temperature. Coupling analysis determines whether the current coupling relationship has been broken, i.e., whether an anomaly exists, by comparing currently observed changes in other characteristics with changes in other characteristics under the same historical baseline conditions.

[0030] In step S20 of the method provided in this application embodiment, based on multiple baseline feature change values, the baseline feature change value index of other operating feature categories of other modules is performed, and feature change coupling analysis of multiple modules is conducted to obtain a feature change anomaly sequence, including: In the historical monitoring data of the power supply, when multiple baseline feature change values ​​appear, the average feature change values ​​of multiple operating feature categories of other modules are used to obtain multiple sets of index feature change values; Calculate the similarity between multiple sets of index feature change values ​​and the corresponding adjacent feature change values ​​in the multiple sets of adjacent feature change values ​​to obtain multiple feature change similarities; Based on the similarity of multiple feature changes, the anomaly degree of multiple feature changes is calculated and arranged to obtain a sequence of feature change anomaly degrees.

[0031] In this embodiment, firstly, based on the historical monitoring data of the power supply, the relevant historical monitoring data when multiple benchmark feature change values ​​appear are indexed using the benchmark feature change value as the retrieval condition. Each historical monitoring data corresponds to a historical operating feature map. Then, the average feature change values ​​of multiple operating feature categories of multiple modules in a historical operating feature map are calculated to obtain the index feature change value. Next, the average value of the historical operating feature maps corresponding to all historical monitoring data is calculated to obtain multiple index feature change values. All index feature change values ​​are then integrated to obtain a set of index feature change values.

[0032] For example, suppose multiple baseline feature change values ​​are obtained by random selection, and one of the baseline feature change values ​​is used for retrieval to obtain multiple historical monitoring data. The average feature change values ​​of multiple operating feature categories of other modules are calculated to obtain the set of index feature change values, such as [5,5,7.5,4,4,3,4,5,3].

[0033] Secondly, the similarity between the corresponding adjacent feature change values ​​in multiple sets of index feature change values ​​and multiple sets of adjacent feature change values ​​is calculated to obtain multiple feature change similarities. Among them, the feature change similarity represents the degree of consistency between the feature coupling relationship and the historical pattern.

[0034] Specifically, taking modules as the corresponding units, the similarity of adjacent feature change values ​​within two sets is calculated. The calculation formula is: 1-[│ab│ / (a+b)]. The larger the similarity value, the more the actual change pattern matches the historical health pattern.

[0035] For example, suppose there exists a set of index feature change values: [5,5,7.5,4,4,3,4,5,3], and suppose there exists a sequence of adjacent feature change values: [3,5.6,3,4,4.5,3,4,5,5]. Calculate 1-(5-3) / (5+3)≈0.75, 1-(5-5.6) / (5+5.6)=0.94, 1-(7.5-3) / (7.5+3)=0.57; 1- (4-4) / (4+4)=1, 1-(4-4.5) / (4+4.5)=0.94, 1-(3-3) / (3+3)=1; 1-(4-4) / (4+4)=1, 1-(5-5) / (5+5)=1, 1-(3-5) / (3+5)=0.75; Feature change similarity=(0.75+0.94+0.57+1+0.94+1+1+1+0.75) / 9≈0.88.

[0036] Next, based on the similarity of multiple feature changes, multiple feature change anomalies are calculated and arranged to obtain a feature change anomaly sequence. The feature change anomaly is calculated based on similarity and is used to quantify the degree of deviation from historical patterns.

[0037] Specifically, the feature change similarity is calculated as 1 minus the feature change similarity. The higher the anomaly score, the more abnormal the change in the power supply's health status, and the greater the likelihood of a health problem. Similarly, all feature change anomalies are calculated and sorted chronologically to obtain a feature change anomaly score sequence.

[0038] For example, when calculating the feature change anomaly degree for feature similarity, the feature change anomaly degree = 1 - 0.88 = 0.12.

[0039] In this embodiment, the average change value of all features between adjacent operating feature maps is calculated to obtain a sequence of adjacent feature change values, quantifying the drastic change in the overall state of the power system. Subsequently, a baseline feature change value is selected for indexing to obtain the coupling relationship between feature parameters of different modules. By calculating the similarity between other feature change values ​​and historical expected patterns, a feature change anomaly sequence is obtained, and the consistency of coupling relationships between features is analyzed and identified, improving the sensitivity and foresight of health status monitoring.

[0040] S30: Based on the anomaly sequence of feature changes and the adjacent feature change value sequence, the running feature map sequence is filtered to obtain a filtered running feature map set, which is then processed to obtain health status information as the monitoring result.

[0041] In this embodiment, the operating feature map is screened by combining the adjacent feature change value and the feature change anomaly degree, and the selected operating feature map is used to construct a screened operating feature map set. Among them, the operating feature map with a larger feature change anomaly degree has greater reference value for analyzing the health status. Finally, the health status information of the continuous operation data of multiple modules in the power supply is identified, and the monitoring results are obtained.

[0042] Step S30 in the method provided in this application embodiment includes: Based on the anomaly degree sequence of the feature changes and the adjacent feature change value sequence, the selected reference value sequence is calculated; Based on the selected reference value sequence, the running feature map corresponding to the largest preset number of selected reference values ​​is selected to obtain the filtered running feature map set.

[0043] In this embodiment of the application, the selected reference value sequence is first calculated based on the feature change anomaly degree sequence and the adjacent feature change value sequence. The selected reference value sequence is a sequence formed by arranging all the selected reference values ​​in chronological order.

[0044] Specifically, the anomaly degree of feature change in the feature change anomaly degree sequence and the adjacent change value in the adjacent feature change value sequence are matched one-to-one. Then, the mean of the two is calculated to obtain multiple corresponding selection reference values, which are also arranged in chronological order to obtain the corresponding selection reference value sequence. That is, the selection reference value = (feature change anomaly degree + adjacent change value) / 2. The size of the selection reference value can determine whether the running feature map is filtered.

[0045] The degree of abnormality and magnitude of change in power supply health status can be reflected by the anomaly degree of characteristic change and the adjacent change value. The larger the value, the more significant the change in power supply status can be reflected, the greater the value for health assessment, and the larger the reference value should be selected.

[0046] Secondly, a preset selection quantity is set based on computing resources, monitoring accuracy requirements, and experience; this preset quantity is the screening quantity. Then, all values ​​in the selected reference value sequence are sorted from largest to smallest, and the K largest values ​​are identified. For each of the K largest values, the corresponding operational feature map is obtained. For example, if the largest reference value corresponds to the change between operational feature map A and operational feature map B, then operational feature map A is selected. The selected K operational feature maps are then compiled to obtain the final set of filtered operational feature maps. Each feature map in this set integrates the magnitude of change and the degree of anomaly, representing the most valuable data that best reflects the evolution of the power supply's health status.

[0047] In step S30 of the method provided in this application embodiment, the process of identifying and processing to obtain health status information includes: Within the selected set of running feature maps, running feature maps are selected according to the first step length to obtain the first set of selected running feature maps; Within the selected set of running feature maps, running feature maps are selected according to a second step size to obtain a second set of selected running feature maps, wherein the second step size is greater than the first step size; The first set of selected running feature maps and the second set of selected running feature maps are input into the battery change health identifier, and the battery change health information is output. Based on the aforementioned feature change anomaly sequence, the average feature change anomaly is calculated, and the battery coupling health information is obtained. By combining the battery coupling health information and the battery change health information, the health status information is calculated and used as the monitoring result.

[0048] In this embodiment, firstly, in the selected running feature map set, sampling is performed starting from the beginning of the sequence according to a pre-set first step length. One running feature map is selected every step length -1, thus obtaining the first selected running feature map set. Here, the first step length is usually small in order to perform dense sampling and capture rapid and detailed changes in the state. For example, if the first step length is set to 2, for the selected running feature map set: [selected running feature map C, selected running feature map D, selected running feature map E, selected running feature map F], then the first selected running feature map set is [selected running feature map C, selected running feature map E].

[0049] Similarly, a second step size is pre-defined. Within the selected running feature map set, sampling is performed starting from the beginning of the sequence according to the second step size. One running feature map is selected every step size - 1, thus obtaining the second selected running feature map set. The second step size is larger than the first step size, and the second step size is usually larger to achieve sparsity and capture macroscopic, slow changes in the state. For example, if the second step size is set to 3, for the selected running feature map set: [selected running feature map C, selected running feature map D, selected running feature map E, selected running feature map F], then the first selected running feature map set is [selected running feature map C, selected running feature map F].

[0050] Secondly, the first set of selected operating feature maps and the second set of selected operating feature maps are input into the battery change health identifier in parallel. The battery change health identifier extracts and quantifies the power health status, and outputs the battery change health information.

[0051] Next, based on the anomaly sequence of feature changes, the average anomaly degree of feature change is calculated, and the battery coupling health information is obtained. The average anomaly degree of feature change is the average of the anomalies across all time periods within the entire monitoring period. For example, for the anomaly sequence of feature changes: [0.12, 0.2], the average value is calculated as (0.12 + 0.2) / 2 = 0.11. The average anomaly degree of feature change reflects the average degree of anomaly in the coupling relationship between power modules throughout the entire monitoring period.

[0052] Calculate the average value of the anomaly rate of feature changes, and then calculate the battery coupled health information based on the average value of the feature change anomaly rate. The battery coupled health information can be calculated as 1 - the average feature change anomaly rate, i.e., Battery coupled health information = 1 - average feature change anomaly rate. For example, Battery coupled health information = 1 - 0.11 = 0.89.

[0053] Finally, by combining battery coupling health information and battery change health information, the health status information is calculated and used as the monitoring result.

[0054] Specifically, the battery-coupled health information and battery change health information are fused using a weighted average to calculate the health status information, i.e., health status information = w1 × battery change health information + w2 × battery-coupled health information, where w1 and w2 are the weights of battery change health information and battery-coupled health information, respectively, and can be set according to the degree of influence of battery-coupled health information and battery change health information on the health status information. The final calculated health status information is output as the monitoring result.

[0055] For example, if w1 and w2 are set to 0.6 and 0.4 respectively, the battery change health information output by the battery change health identifier is 0.8, the battery coupling health information is 0.89, and the health status information is calculated as 0.6 × 0.8 + 0.4 × 0.89 = 0.836.

[0056] In step S30 of the method provided in this application embodiment, the configuration step of the battery change health identifier includes: Based on the health monitoring data of the battery over a historical period, multiple sample operation feature maps and sample battery change health information sets are collected. The multiple sample operation feature maps are processed according to the first step length and the second step length to obtain multiple sample first selected operation feature map sets and multiple sample second selected operation feature map sets. The battery change health information of each sample includes the labeled power health percentage value. A battery health change identifier is constructed based on the slowfast network, wherein the battery health change identifier includes a dynamic identifier network and a static identifier network. The battery change health identifier is trained and validated using the first set of selected running feature maps of multiple samples and the second set of selected running feature maps of multiple samples, respectively, and the battery change health information set of the samples is used as the supervision data. The configuration is completed after convergence.

[0057] In this embodiment, the sample running feature map set is a running feature map set extracted from historical data; the sample battery change health information set is the historical power health status information corresponding to the sample running feature map; the first sample selected running feature map set and the second sample selected running feature map set are two sets of images obtained by sampling each sample running feature map set according to the first step length and the second step length, respectively, and will be used as two parallel inputs to the subsequent deep learning model.

[0058] First, based on the health monitoring data of the battery over a historical period, multiple sample operation feature maps and sample battery change health information sets are collected. Then, according to the preset first step length and second step length, the collected multiple sample operation feature maps are subjected to the above-mentioned filtering process to construct multiple sample first selected operation feature map sets and multiple sample second selected operation feature map sets. Among them, the battery change health information of each sample includes a labeled power health percentage value, which is usually a power health percentage value labeled by experts based on historical fault records, maintenance records or laboratory test results.

[0059] Secondly, a battery change health identifier is constructed based on the slowfast network, which includes a dynamic identification network and a static identification network.

[0060] Specifically, the SlowFast network is a common deep learning network architecture in the field of video action recognition. It uses two parallel pathways to process the video: a slow pathway, which samples at a lower frame rate to capture spatial semantics and slowly changing background information; and a fast pathway, which samples at a higher frame rate to capture details of fast motion. The two pathways are then fused to process features, thus simultaneously handling both slow-changing trends and fast-changing details.

[0061] A dynamic recognition network and a static recognition network were constructed respectively. The structure of the dynamic recognition network and the static recognition network consists of an input layer, a convolutional layer, and a pooling layer. The input layer of the dynamic recognition network is used to receive the first set of selected running feature maps of the samples to capture the rapid changes and instantaneous fluctuations of the power supply state. The convolutional layer is used to extract the deep layers that can reflect the rapid fluctuations and instantaneous changes of the power supply parameters. The convolutional layer uses a 5×7×7 convolutional kernel to capture rapid changes. The pooling layer is a 1×3×3 max pooling layer to downsample the feature maps and compress the spatial size to retain fine temporal change information.

[0062] The input layer of the static recognition network receives a set of sample second-selected running feature maps and extracts deep features that reflect the overall state of the power supply and its slow aging trend. The convolutional layer uses 1×7×7 convolutions to extract local features; the pooling layer uses 1×3×3 max pooling to downsample the feature maps and compress their spatial size. The fused feature vector obtained by concatenating the two layers is then passed through fully connected layers and output via the Softmax activation function.

[0063] Finally, the training dataset was divided into training and validation sets in an 8:2 ratio. The model training parameters were then set as follows: batch size was set to 16, initial learning rate to 0.01, stochastic gradient descent optimizer was used, momentum coefficient was set to 0.9, and weight decay was set to 1e. -4 To prevent overfitting, a cosine annealing strategy is used for learning rate scheduling, reducing the learning rate to 0.1 times its original value every 10 training epochs. The cross-entropy loss function, suitable for multi-class classification tasks, is used.

[0064] During model training, the first set of selected running feature maps is input into the dynamic recognition network, and the second set of selected running feature maps is input into the static recognition network. Forward propagation is performed to obtain the feature outputs of the two paths. After feature fusion, global average pooling and fully connected layers are used to obtain the predicted battery health information. The cross-entropy loss between the predicted probability and the true label is calculated, and backpropagation is performed to calculate the gradient. Gradient clipping is applied to prevent gradient explosion. The optimizer is called to update the network parameters. The gradient is cleared to zero after each training batch is completed.

[0065] After each training epoch, the model performance is evaluated on the validation set, and classification accuracy, precision, and recall are calculated, while the validation loss value is monitored. Training is performed for a total of 80 epochs. Training is stopped early when the validation set accuracy no longer improves for 10 consecutive epochs, and the model weights are saved to obtain the battery change health identifier.

[0066] For example, the first selected set of running feature maps and the second selected set of running feature maps are input into the trained battery change health identifier, and the output battery change health information is, for example, 0.8.

[0067] In this embodiment, a first set of selected operating feature maps and a second set of selected operating feature maps are obtained through first-step and second-step sampling, providing input data for capturing power state evolution characteristics. Subsequently, a battery change health identifier is constructed based on a Slowfast network, enabling the model to process high-temporal-resolution fast-change information and low-temporal-resolution slow-change information in parallel, improving the ability to identify complex health patterns. Furthermore, by fusing battery coupled health information and battery change health information, a comprehensive assessment of the power state from different dimensions is performed, improving the accuracy of monitoring results.

[0068] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the operating characteristics of multiple modules within the power supply are first collected during the monitoring period. Then, an operating feature map template is constructed and the normalized module operating feature values ​​are written into the corresponding pixel sites to construct an operating feature map sequence. This process converts the dispersed multi-source monitoring data into a unified format, achieving data standardization and laying a high-quality data foundation for subsequent processing.

[0069] Secondly, the average change value of all features between adjacent operating feature maps is calculated to obtain a sequence of adjacent feature change values, quantifying the drastic change in the overall state of the power system. Subsequently, a baseline feature change value is selected for indexing to obtain the coupling relationship between feature parameters of different modules. By calculating the similarity between other feature change values ​​and historical expected patterns, a feature change anomaly sequence is obtained. This allows for analysis and identification of the consistency of coupling relationships between features, improving the sensitivity and foresight of health status monitoring. Finally, by sampling with first and second step lengths, a first set of selected running feature maps and a second set of selected running feature maps are obtained, providing input data for capturing power state evolution characteristics. Subsequently, a battery change health identifier is constructed based on a Slowfast network, enabling the model to process high-temporal-resolution fast-changing information and low-temporal-resolution slow-changing information in parallel, improving the ability to identify complex health patterns. Furthermore, by fusing battery coupled health information and battery change health information, a comprehensive assessment of the power state from different dimensions is performed, improving the accuracy of monitoring results.

[0070] Example 2, as Figure 2 As shown, based on the same inventive concept as the deep learning-based modular power supply health status monitoring method provided in Embodiment 1, this embodiment of the invention also provides a deep learning-based modular power supply health status monitoring system, including: The feature acquisition module 11 is used to acquire the operating features of multiple modules in the power supply during the monitoring time period, obtain multiple module operating feature sets, and construct an operating feature map sequence. The feature change analysis module 12 is used to identify adjacent changes in the running feature map sequence, obtain the adjacent feature change value sequence, and perform feature change coupling analysis of multiple modules to obtain the feature change anomaly sequence. The health status identification module 13 is used to filter the running feature map sequence based on the feature change anomaly degree sequence and the adjacent feature change value sequence to obtain a filtered running feature map set, identify and process it to obtain health status information as a monitoring result.

[0071] In one embodiment, the feature acquisition module 11 is used to: The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets. Each module includes multiple operating characteristic categories, and the module operating characteristics of multiple operating characteristic categories are collected. Multiple module running feature sets are used as pixel values ​​of running feature maps to construct multiple running feature maps, resulting in a sequence of running feature maps.

[0072] This involves using the pixel values ​​of multiple module operation feature sets as the pixel values ​​of the operation feature maps to construct multiple operation feature maps, resulting in an operation feature map sequence, including: Based on multiple operational feature categories from multiple modules, multiple pixel sites within the operational feature map are constructed to obtain the operational feature map template; The running feature sets of multiple modules are normalized, and the normalized running feature value of each module is written into the pixel position corresponding to the running feature map template to obtain multiple running feature maps and obtain a running feature map sequence.

[0073] In one embodiment, the feature change analysis module 12 is used for: Calculate the change value of all module running feature values ​​in each running feature map within the running feature map sequence and the next adjacent running feature map to obtain multiple sets of adjacent feature change values; Calculate the mean of multiple sets of adjacent feature change values ​​to obtain multiple adjacent feature change values, and arrange them to obtain a sequence of adjacent feature change values.

[0074] This involves performing a coupled analysis of feature changes across multiple modules to obtain a sequence of feature change anomalies, including: Within multiple sets of adjacent feature change values, one adjacent feature change value is randomly selected as the baseline feature change value, resulting in multiple baseline feature change values. Based on multiple baseline feature change values, index the baseline feature change values ​​of other operational feature categories of other modules, and perform feature change coupling analysis of multiple modules to obtain a feature change anomaly sequence.

[0075] Specifically, based on multiple baseline feature change values, baseline feature change value indexes for other operational feature categories of other modules are performed, and feature change coupling analysis of multiple modules is conducted to obtain a feature change anomaly sequence, including: In the historical monitoring data of the power supply, when multiple baseline feature change values ​​appear, the average feature change values ​​of multiple operating feature categories of other modules are used to obtain multiple sets of index feature change values; Calculate the similarity between multiple sets of index feature change values ​​and the corresponding adjacent feature change values ​​in the multiple sets of adjacent feature change values ​​to obtain multiple feature change similarities; Based on the similarity of multiple feature changes, the anomaly degree of multiple feature changes is calculated and arranged to obtain a sequence of feature change anomaly degrees.

[0076] In one embodiment, the health status identification module 13 is used for: Based on the anomaly degree sequence of the feature changes and the adjacent feature change value sequence, the selected reference value sequence is calculated; Based on the selected reference value sequence, the running feature map corresponding to the largest preset number of selected reference values ​​is selected to obtain the filtered running feature map set.

[0077] The process of identifying and processing health status information includes: Within the selected set of running feature maps, running feature maps are selected according to the first step length to obtain the first set of selected running feature maps; Within the selected set of running feature maps, running feature maps are selected according to a second step size to obtain a second set of selected running feature maps, wherein the second step size is greater than the first step size; The first set of selected running feature maps and the second set of selected running feature maps are input into the battery change health identifier, and the battery change health information is output. Based on the aforementioned feature change anomaly sequence, the average feature change anomaly is calculated, and the battery coupling health information is obtained. By combining the battery coupling health information and the battery change health information, the health status information is calculated and used as the monitoring result.

[0078] The configuration steps of the battery change health identifier include: Based on the health monitoring data of the battery over a historical period, multiple sample operation feature maps and sample battery change health information sets are collected. The multiple sample operation feature maps are processed according to the first step length and the second step length to obtain multiple sample first selected operation feature map sets and multiple sample second selected operation feature map sets. The battery change health information of each sample includes the labeled power health percentage value. A battery health change identifier is constructed based on the slowfast network, wherein the battery health change identifier includes a dynamic identifier network and a static identifier network. The battery change health identifier is trained and validated using the first set of selected running feature maps of multiple samples and the second set of selected running feature maps of multiple samples, respectively, and the battery change health information set of the samples is used as the supervision data. The configuration is completed after convergence.

[0079] Compared to existing technologies, this application first collects the operating characteristics of multiple modules within the power supply during the monitoring time period; then, it constructs an operating feature map template and writes the normalized module operating feature values ​​into the corresponding pixel sites to construct an operating feature map sequence, thereby standardizing the data by uniformly converting the scattered multi-source monitoring data into a unified format, laying a high-quality data foundation for subsequent applications.

[0080] Secondly, the average change value of all features between adjacent operating feature maps is calculated to obtain a sequence of adjacent feature change values, quantifying the drastic change in the overall state of the power system. Subsequently, a baseline feature change value is selected for indexing to obtain the coupling relationship between feature parameters of different modules. By calculating the similarity between other feature change values ​​and historical expected patterns, a feature change anomaly sequence is obtained. This allows for analysis and identification of the consistency of coupling relationships between features, improving the sensitivity and foresight of health status monitoring. Finally, by sampling with first and second step lengths, a first set of selected running feature maps and a second set of selected running feature maps are obtained, providing input data for capturing power state evolution characteristics. Subsequently, a battery change health identifier is constructed based on a Slowfast network, enabling the model to process high-temporal-resolution fast-changing information and low-temporal-resolution slow-changing information in parallel, improving the ability to identify complex health patterns. Furthermore, by fusing battery coupled health information and battery change health information, a comprehensive assessment of the power state from different dimensions is performed, improving the accuracy of monitoring results.

Claims

1. A modular power supply health status monitoring method based on deep learning, characterized in that, The method includes: The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets and construct an operating characteristic map sequence; Adjacency change identification is performed on the running feature map sequence to obtain the adjacent feature change value sequence, and feature change coupling analysis of multiple modules is performed to obtain the feature change anomaly sequence; Based on the anomaly sequence of feature changes and the adjacent feature change value sequence, the running feature map sequence is filtered to obtain a filtered running feature map set. The health status information is then identified and processed as the monitoring result.

2. The modular power supply health status monitoring method based on deep learning according to claim 1, characterized in that, The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets, and an operating characteristic map sequence is constructed, including: The operating characteristics of multiple modules within the power supply are collected during the monitoring time period to obtain multiple module operating characteristic sets. Each module includes multiple operating characteristic categories, and the module operating characteristics of multiple operating characteristic categories are collected. Multiple module running feature sets are used as pixel values ​​of running feature maps to construct multiple running feature maps, resulting in a sequence of running feature maps.

3. The modular power supply health status monitoring method based on deep learning according to claim 2, characterized in that, Multiple module running feature sets are used as pixel values ​​to construct multiple running feature maps, resulting in a sequence of running feature maps, including: Based on multiple operational feature categories from multiple modules, multiple pixel sites within the operational feature map are constructed to obtain the operational feature map template; The running feature sets of multiple modules are normalized, and the normalized running feature value of each module is written into the pixel position corresponding to the running feature map template to obtain multiple running feature maps and obtain a running feature map sequence.

4. The modular power supply health status monitoring method based on deep learning according to claim 1, characterized in that, The adjacency change identification is performed on the running feature map sequence to obtain an adjacency feature change value sequence, including: Calculate the change value of all module running feature values ​​in each running feature map within the running feature map sequence and the next adjacent running feature map to obtain multiple sets of adjacent feature change values; Calculate the mean of multiple sets of adjacent feature change values ​​to obtain multiple adjacent feature change values, and arrange them to obtain a sequence of adjacent feature change values.

5. The modular power supply health status monitoring method based on deep learning according to claim 4, characterized in that, A feature change coupling analysis of multiple modules is performed to obtain a feature change anomaly sequence, including: Within multiple sets of adjacent feature change values, one adjacent feature change value is randomly selected as the baseline feature change value, resulting in multiple baseline feature change values. Based on multiple baseline feature change values, index the baseline feature change values ​​of other operational feature categories of other modules, and perform feature change coupling analysis of multiple modules to obtain a feature change anomaly sequence.

6. The modular power supply health status monitoring method based on deep learning according to claim 5, characterized in that, Based on multiple baseline feature change values, the baseline feature change values ​​of other operational feature categories of other modules are indexed, and feature change coupling analysis of multiple modules is performed to obtain a feature change anomaly sequence, including: In the historical monitoring data of the power supply, when multiple baseline feature change values ​​appear, the average feature change values ​​of multiple operating feature categories of other modules are used to obtain multiple sets of index feature change values; Calculate the similarity between multiple sets of index feature change values ​​and the corresponding adjacent feature change values ​​in the multiple sets of adjacent feature change values ​​to obtain multiple feature change similarities; Based on the similarity of multiple feature changes, the anomaly degree of multiple feature changes is calculated and arranged to obtain a sequence of feature change anomaly degrees.

7. The modular power supply health status monitoring method based on deep learning according to claim 1, characterized in that, Based on the aforementioned feature change anomaly sequence and adjacent feature change value sequence, the running feature map sequence is filtered to obtain a filtered running feature map set, including: Based on the anomaly degree sequence of the feature changes and the adjacent feature change value sequence, the selected reference value sequence is calculated; Based on the selected reference value sequence, the running feature map corresponding to the largest preset number of selected reference values ​​is selected to obtain the filtered running feature map set.

8. The modular power supply health status monitoring method based on deep learning according to claim 1, characterized in that, The process of identifying and processing health status information includes: Within the selected set of running feature maps, running feature maps are selected according to the first step length to obtain the first set of selected running feature maps; Within the selected set of running feature maps, running feature maps are selected according to a second step size to obtain a second set of selected running feature maps, wherein the second step size is greater than the first step size; The first set of selected running feature maps and the second set of selected running feature maps are input into the battery change health identifier, and the battery change health information is output. Based on the aforementioned feature change anomaly sequence, the average feature change anomaly is calculated, and the battery coupling health information is obtained. By combining the battery coupling health information and the battery change health information, the health status information is calculated and used as the monitoring result.

9. The modular power supply health status monitoring method based on deep learning according to claim 8, characterized in that, The configuration steps for the battery change health detector include: Based on the health monitoring data of the battery over a historical period, multiple sample operation feature maps and sample battery change health information sets are collected. The multiple sample operation feature maps are processed according to the first step length and the second step length to obtain multiple sample first selected operation feature map sets and multiple sample second selected operation feature map sets. The battery change health information of each sample includes the labeled power health percentage value. A battery health change identifier is constructed based on the slowfast network, wherein the battery health change identifier includes a dynamic identifier network and a static identifier network. The battery change health identifier is trained and validated using the first set of selected running feature maps of multiple samples and the second set of selected running feature maps of multiple samples, respectively, and the battery change health information set of the samples is used as the supervision data. The configuration is completed after convergence.

10. A modular power supply health status monitoring system based on deep learning, characterized in that, The system is used to implement the modular power supply health status monitoring method based on deep learning according to any one of claims 1-9, the system comprising: The operation feature acquisition module is used to collect the operation features of multiple modules in the power supply during the monitoring time period, obtain multiple module operation feature sets, and construct an operation feature map sequence. The feature change analysis module is used to identify adjacent changes in the running feature map sequence, obtain the adjacent feature change value sequence, and perform feature change coupling analysis of multiple modules to obtain the feature change anomaly sequence. The health status identification module is used to filter the running feature map sequence based on the feature change anomaly sequence and the adjacent feature change value sequence to obtain a filtered running feature map set, and to identify and process it to obtain health status information as a monitoring result.

Images