Method and apparatus for detecting failure of internal critical components of an uninterruptible power supply

Abstract

The application provides a failure detection method for internal key components of an uninterruptible power supply, comprising: judging whether a current operation mode is set as a training stage of a model; if the training stage is entered, historical data is used for data preprocessing and data set division; if a non-training stage is entered, current data is collected for data preprocessing, and the historical data and the current data comprise operation parameter data and image data; the operation parameter data is input into a convolutional neural network model combined with incremental learning, and the image data is input into a tensor network model using multi-dimensional feature sampling and matrix multiplication state operation; the current operation mode is judged again whether it is set as the training stage of the model; if the training stage is entered, model training is performed, and if the requirements are met, the trained model is deployed; if the non-training stage is entered, a failure detection result is output, and a failure detection device using the above method is also provided.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to a fault detection method and apparatus for key internal components of an uninterruptible power supply. Background Technology

[0002] In the information age, uninterruptible power systems (UPS) mainly consist of energy storage modules (typically lead-acid batteries), rectifiers, inverters, and control switches. They utilize batteries to transmit electrical energy, and through rectifier and inverter circuits, convert direct current (DC) to alternating current (AC), providing uninterrupted and stable power to critical power electronic equipment. In the era of big data, UPS plays a vital role in ensuring information security, especially in large hospitals and other locations where UPS systems power medical instruments and surgical equipment, making it crucial to ensure their normal operation.

[0003] Regular maintenance and repair of UPS systems is a specialized technical task, typically performed by professional engineering maintenance personnel. However, manual maintenance has limitations such as heavy reliance on the experience and knowledge of the technicians, significant delays in maintenance and repair that prevent predictive maintenance, and the high cost of human, material, and financial resources. Therefore, proposing intelligent fault diagnosis methods and corresponding devices can effectively overcome the limitations of manual maintenance. However, existing general fault diagnosis methods suffer from poor portability and adaptability across different domains, and the large number of parameters in deep models makes them difficult to deploy on edge devices and perform real-time predictive analysis. Summary of the Invention

[0004] This application provides a robust and highly generalizable method and apparatus for fault detection of critical internal components of an uninterruptible power supply.

[0005] On the one hand, this application provides a fault detection method for key internal components of an uninterruptible power supply, including: Determine whether the current running mode is set to the model training phase: if it is in the training phase, use historical data to perform data preprocessing and divide the dataset; if it is not in the training phase, collect current data and perform data preprocessing. Historical data and current data include running parameter data and image data. A convolutional neural network model that combines incremental learning with inputting parametric data, and a tensor network model that uses multidimensional feature sampling and matrix multiplication state operations as inputting image data. The system then checks again whether the current operating mode is set to the model's training phase: if it is in the training phase, the model is trained, and if the requirements are met, the trained model is deployed; if it is not in the training phase, the fault detection result is output.

[0006] Furthermore, the key internal components of an uninterruptible power supply (UPS) include a battery, rectifier, inverter, and bypass isolation transformer, and its operating parameter data includes the main input voltage value. Bypass input voltage value Inverter output voltage Inverter output current Inverter frequency value Load Battery charging voltage Battery charging current There are a total of 8 operating parameter variables; the image data includes optical detection images at 4 locations: battery, rectifier, inverter, and bypass isolation transformer.

[0007] Furthermore, for the 8 operational parameter variables included in the historical data, data from the previous 30 days to the current moment were collected, and the average value of the parameter per minute was calculated once per hour to form an operational parameter dataset; for the optical detection images at the 4 locations, images from the previous 360 days to the current moment were collected, and images were collected once every 12 hours to form an image dataset.

[0008] Furthermore, "using historical operational parameter data and image data to perform data preprocessing and dataset partitioning" includes: cleaning the operational parameter data by filling in missing values, replacing outliers, and smoothing noise values, and performing data normalization operations; and performing contrast enhancement, noise reduction, and image cropping operations on the image data. "Collecting current operational parameter data and image data for data preprocessing" includes: dividing the operational parameter dataset and image data dataset into training set and validation set according to a selected ratio.

[0009] Furthermore, "collecting current data and performing data preprocessing" includes standardizing the runtime parameter data and cropping the image data.

[0010] Furthermore, "a convolutional neural network model that combines runtime parameter data input with incremental learning" includes: Calculate the matching degree between the original sample and the new sample. The weight parameters of the new convolutional network = the weight parameters of the original convolutional network × + Weight parameters of the new convolutional network × After passing through a convolutional network, the data stream passes through a classifier consisting of three fully connected layers to obtain the fault detection result output, with a matching degree. The calculation formula is as follows: The convolutional network consists of 5 convolutional and pooling blocks, where k is the feature size, n represents the new sample, and o represents the best matching sample of n. and Let n and o represent the i-th features of n and o respectively, and min(·) and max(·) represent the minimum and maximum value operations respectively. The range of its value is between (0,1].

[0011] Furthermore, the tensor network model using multidimensional feature sampling and matrix product state operations consists of a multidimensional sampling layer, a matrix product state network containing 4 MPS operations, another multidimensional sampling layer, a matrix product state network containing 3 MPS operations, and a product state network containing 2 MPS operations connected sequentially. The image is divided into 8 sub-regions. The multidimensional sampling layer includes average region sampling (calculating the average value of each sub-region as the output), maximum region sampling (selecting the maximum value of each sub-region as the output), and correlation region sampling (calculating the local region correlation crs as the output). The local region correlation crs and tensor... The calculation formula is as follows: Where p represents the total number of pixels in each sub-region, (x,y) represents the pixel position in each sub-region, E(·) represents the Euclidean distance, and MPS represents the matrix multiplication state. This represents a low-rank tensor. , ,…, This represents the reference of an N-dimensional tensor.

[0012] Furthermore, accuracy, precision, recall, and task matching metric (TMD) are used to determine whether the trained model meets the deployment requirements. If it does not meet the requirements, the process returns to the step of "using historical data to perform data preprocessing and dataset partitioning". The calculation formula for the task matching metric (TMD) and the model deployment judgment indicator (D) are as follows: in, and Let represent the number of correct predictions and the number of incorrect predictions for the i-th task on the validation set, respectively. n is the number of tasks. apr represents the average of accuracy, precision, and recall. min(·) and avg(·) represent the minimum and average values, respectively.

[0013] On the other hand, this application provides a fault detection device for key internal components of an uninterruptible power supply (UPS), comprising: a data acquisition module, a host computer computing module, and a display and alarm module. The data acquisition module includes an image acquisition device, an electrical signal acquisition device, a data processing unit, and a data transmission line. The data processing unit is used to perform preprocessing operations on image data and operating parameter data. The host computer computing module includes a communication interface, an operating mode determiner, a server-side trainer, an edge-side deployer, and a data storage unit. The communication interface is used for data flow transmission and information communication with other modules. The operating mode determiner is used to set whether the model is in the training phase or the non-training phase. The server-side trainer is used to complete the training process of the multi-task learning model. The edge-side deployer implements lightweight edge device deployment for the model that has completed training and meets the deployment requirements. The calculation results output by the algorithm model are stored in the data storage unit. The display and alarm module includes a communication interface, a result display unit, a result analysis unit, a fault alarm unit, and a data storage unit. The communication interface is used for data transmission and information communication between different modules. The result display unit visualizes the calculation results of the model. The result analysis unit performs logical judgment and analysis based on the output results and sends instructions to the fault alarm unit. The data storage unit stores the result analysis and fault alarm records.

[0014] This invention provides a fault detection method and apparatus for key internal components of an uninterruptible power supply (UPS). It employs a multimodal, multi-task analysis method to adapt to the intelligent analysis needs of various faults and components under different detection environments, ensuring the accuracy and reliability of the results. This includes intelligent analysis of collected operational parameter data using a convolutional neural network model with incremental learning, effectively solving the problem of model domain adaptation. It also utilizes a tensor network with multidimensional feature sampling and matrix multiplication state operations to address the high redundancy parameter challenges of general deep image recognition models. Furthermore, it provides an integrated system comprising a data acquisition module, a host computer processing module, and a display and alarm module to achieve full-process deployment of the fault diagnosis algorithm model on the UPS. This invention's fault detection method and apparatus for key internal components of an uninterruptible power supply improves the automation and intelligence level of UPS fault diagnosis while efficiently ensuring the adaptive learning and lightweight deployment capabilities of the algorithm model. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a fault detection method for key internal components of an uninterruptible power supply provided by the present invention. Figure 2 yes Figure 1 Schematic diagram of the fault diagnosis model for key internal components of an uninterruptible power supply; Figure 3 This is a structural diagram of the fault detection device for key internal components of an uninterruptible power supply provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are preferred embodiments of the present invention and should not be considered as excluding other embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] This invention provides a fault detection method for critical internal components of an uninterruptible power supply (UPS), which is used to monitor critical internal components of the UPS in real time and output fault diagnosis results.

[0019] Key internal components of an uninterruptible power supply (UPS) include a battery, rectifier, inverter, and bypass isolation transformer. It also collects operational parameter data and image data through multi-source sensors. Operational parameter data includes the main input voltage value. Bypass input voltage value Inverter output voltage Inverter output current Inverter frequency value Load Battery charging voltage Battery charging current There are a total of 8 operating parameter variables; the image data includes optical detection images at four locations: battery, rectifier, inverter, and bypass isolation transformer. The fault attributes include fault domain (including appearance defects and mechanism faults), fault type (covering battery faults, rectifier faults, inverter faults, and bypass isolation transformer faults), fault level (divided into four levels: I / II / III / IV, with the severity of the fault increasing progressively), and confidence probability (i.e., how confident the model is in determining the reliability of the output results).

[0020] For the eight operational parameter variables included in the above-mentioned operational parameter data, data from the previous 30 days to the current moment will be collected. The average value of the parameter per minute will be calculated once every hour, meaning each operational parameter variable will have 720 data values, for a total of 720 × 8 = 5760 data values. For the optical detection images at the four locations mentioned above, images from the previous 360 days to the current moment will be collected once every 12 hours, meaning each location will have 720 images, for a total of 720 × 4 = 2880 images. The 5760 data values ​​and 2880 images mentioned above will be divided into training and validation sets in a 4:1 ratio. The eight operational parameter variable values ​​at each collection moment will form one set, and the optical detection images at the four locations at each collection moment will form another set. Fault attribute labels (including fault domain, fault type, and fault level) will be manually labeled, containing a total of 720 labels corresponding to the operational parameters and 720 labels corresponding to the images.

[0021] like Figure 1 As shown in the figure, the specific process of a fault detection method for key internal components of an uninterruptible power supply provided by an embodiment of the present invention is described as follows: Step 1: Power on the system and start running the program.

[0022] Step 2: Determine the current operating mode and check if it is set to the model training phase.

[0023] Step 3: If in the training phase, utilize historical data and perform data preprocessing as required: For the runtime parameter data, perform data cleaning by filling missing values, replacing outliers, and smoothing noise values, and then perform data normalization; for the image data, perform contrast enhancement, noise reduction, and image cropping. Divide the dataset (i.e., 2880 images and 5760 runtime parameter variables) into training and validation sets at a ratio of 4:1 (this ratio is a suggested value based on experience; alternatively, 5:1, 3:1, etc., can also be chosen). Step 4: If the system is not in the training phase (prediction phase), collect the current operating parameter data or image data and perform data preprocessing as required: perform normalization operations on the operating parameter data; and perform image cropping operations on the image data.

[0024] Step 5: Input the parameter data into the convolutional neural network model that incorporates incremental learning, and input the image data into the tensor network model that uses multidimensional feature sampling and matrix multiplication state operations.

[0025] Step 6: Determine the current operating mode again to see if it is set to the model training phase.

[0026] Step 7: If in the training phase, train the model and perform a comprehensive analysis using accuracy, precision, recall, and the Task Match Degree (TMD) metric proposed in this embodiment of the invention to fully evaluate whether the trained model meets the deployment requirements. For models that meet the deployment requirements, deploy the trained model to the subsequent uninterruptible power supply fault detection device; for models that do not meet the deployment requirements, return to step 3. Step 8: If the system is in the non-training phase (prediction phase), output the fault detection results, including the fault domain (appearance defects or mechanism faults), fault type (covering battery faults, rectifier faults, inverter faults or bypass isolation transformer faults), fault level (divided into four levels: I / II / III / IV, with the severity of the fault increasing step by step), and confidence probability (i.e. how confident the model is in determining the reliability of the output results).

[0027] The calculation formula for the task matching evaluation index TMD and the model deployment judgment indicator D are as follows: in, and represents the number of correct predictions and the number of incorrect predictions made by the i-th task on the validation set, respectively. n is the number of tasks (in this embodiment of the invention, n=2, including analysis models for running parameter data and analysis models for image data). apr represents the average of accuracy, precision, and recall. min(·) and avg(·) represent the minimum and average values, respectively.

[0028] like Figure 2 As shown, it illustrates the principles of a convolutional neural network model that combines incremental learning and a tensor network model that uses multidimensional feature sampling and matrix multiplication state operations.

[0029] On the one hand, for the runtime parameter data, the input to the convolutional neural network model combined with incremental learning first calculates the matching degree between the original samples and the newly added samples. and according to The value of the weights determines how many weights the original convolutional network has. A new convolutional network can be obtained by copying the parameters; that is, the weights of the new convolutional network = the weights of the original convolutional network × + Weight parameters of the newly added sample convolutional network × The convolutional network consists of 5 convolutional and pooling blocks. The kernel size (kernel_size), stride, number of filters (num_filters), and pooling size (padding_size) and stride are set in the figure. After passing through the convolutional network, the data stream passes through a classifier consisting of three fully connected layers to obtain the fault detection result output. The matching degree... The calculation formula is as follows, where k is the size of the feature, n represents the new sample, and o represents the best matching sample of n. and Let n and o represent the i-th features of n and o respectively, and min(·) and max(·) represent the minimum and maximum value operations respectively. The range of its value is between (0,1].

[0030] On the other hand, the detected image data is input into a tensor network model that uses multidimensional feature sampling and matrix product state operations. This model consists of a multidimensional sampling layer and a matrix product state network. The image is divided into eight sub-regions. The multidimensional sampling layer includes average region sampling (calculating the average value of each sub-region as the output), maximum region sampling (selecting the maximum value of each sub-region as the output), and correlation region sampling (calculating the correlation of region sampling (crs) of the local region as the output, where p is the total number of pixels in each sub-region, (x,y) represents the pixel position of each sub-region, and E(·) represents the Euclidean distance). The formula for calculating the correlation of the local region (crs) is as follows: Matrix product state (MPS) is a type of tensor network state that approximates a higher-order tensor by using a network of shrinking lower-order tensors. This represents the Nth tensor node of the MPS operation; specifically, an index is... , ,…, The N-dimensional tensor is obtained through the low-rank tensor. By approximating, we obtain the approximate tensor. for: The tensor network model proposed in this embodiment, which uses multidimensional feature sampling and matrix product state operation, consists of a multidimensional sampling layer, a matrix product state network containing 4 MPS operations, a multidimensional sampling layer, a matrix product state network containing 3 MPS operations, and a product state network containing 2 MPS operations connected sequentially. The final output is the fault detection result (including fault domain, fault type, fault level, and confidence probability).

[0031] like Figure 3 As shown in the figure, this embodiment of the invention proposes a fault detection device for key internal components of an uninterruptible power supply based on incremental learning and tensor networks, including a data acquisition module, a host computer computing module, and a display and alarm module.

[0032] The data acquisition module includes an image acquisition unit, an electrical signal acquisition device, a data processing unit, and data transmission lines. The image acquisition unit collects optical inspection images from four locations: the battery, rectifier, inverter, and bypass isolation transformer. The electrical signal acquisition device collects operating parameter data from the battery, rectifier, inverter, and bypass isolation transformer, including the main input voltage value. Bypass input voltage value Inverter output voltage Inverter output current Inverter frequency value Load Battery charging voltage Battery charging current There are a total of 8 electrical parameters. The data processing unit is used to perform preprocessing operations on image data and operational parameter data. The data transmission line is used for data stream transmission and information communication with other modules.

[0033] The host computer computing module includes a communication interface, a runtime mode determiner, a server-side trainer, an edge-side deployer, and a data storage unit. The communication interface is used for data flow transmission and information communication with other modules. The runtime mode determiner is used to set whether the model is in the training or non-training phase. The server-side trainer completes the training process of the multi-task learning model. The edge-side deployer enables lightweight deployment of the trained model to edge devices that meet deployment requirements. The computational results output by the algorithm model are stored in the data storage unit for convenient data management and on-demand retrieval.

[0034] The alarm display module includes a communication interface, a result display unit, a result analysis unit, a fault alarm unit, and a data storage unit. The communication interface is used for data transmission and communication between different modules. The result display unit visualizes the model's calculation results, including the fault domain (appearance defects or mechanism faults), fault type (covering battery faults, rectifier faults, inverter faults, or bypass isolation transformer faults), fault level (FL, divided into four levels: I / II / III / IV, with progressively increasing fault severity), and confidence probability (prob, i.e., how confident the model is in determining the reliability of the output results). The result analysis unit performs logical judgment and analysis based on the output results, sending instructions to the fault alarm unit, which then executes different fault alarms (DA) according to the instructions. Specifically, R / Y / B / G represent red, yellow, blue, and green indicator light alarms, respectively, with color indicating progressively decreasing severity. Indicator lights at different locations represent the fault domain and fault type. Related result analyses and fault alarm records are stored in the data storage unit for easy retrieval and analysis of the results data.

[0035] This invention provides a fault detection method for key internal components of an uninterruptible power supply (UPS). It designs a multi-task learning model to complete intelligent fault detection tasks for different modal data, including operational parameters and images. Specifically, for operational parameter data, a convolutional neural network model combining incremental learning is designed for fault attribute prediction, effectively addressing the adaptive learning problem when new fault categories are introduced. For image data, a tensor network using multidimensional feature sampling and matrix multiplication state operations is proposed for image-based fault identification, effectively overcoming the high redundancy and large-scale parameter problems of general deep image recognition models. The training process for both the convolutional neural network model combining incremental learning and the tensor network using multidimensional feature sampling and matrix multiplication state operations is completed using training set data. The performance of these two models is evaluated using validation set data, including three conventional metrics: accuracy, precision, and recall, as well as the Task Matching Degree (TMD) metric proposed in this application. Models meeting the performance requirements are then deployed.

[0036] As described above, this invention constructs a novel multimodal, multi-task analysis method to adapt to the intelligent analysis needs of various faults and components under different detection environments, ensuring the accuracy and reliability of the results. This includes designing a convolutional neural network model combining incremental learning to intelligently analyze the collected operational parameter data, effectively solving the problem of model domain adaptation; and designing a tensor network using multidimensional feature sampling and matrix multiplication state operations to address the high redundancy parameter challenge of general deep image recognition models. This achieves domain adaptation and accurate, rapid fault diagnosis, improving the automation and intelligence level of uninterruptible power supply (UPS) fault diagnosis while ensuring the strong robustness and high generalization of the fault detection model, guaranteeing the algorithm model's adaptive learning and lightweight deployment capabilities.

[0037] The descriptions of the foregoing specification and embodiments are used to explain the scope of protection of the present invention, but do not constitute a limitation on the scope of protection of the present invention. Modifications, equivalent substitutions, or other improvements to the embodiments of the present invention or a portion thereof that can be obtained by those skilled in the art through logical analysis, reasoning, or limited experimentation, based on the teachings of the present invention or the foregoing embodiments, in conjunction with common general knowledge, ordinary technical knowledge in the art, and / or existing technology, should all be included within the scope of protection of the present invention.

Claims

1. A method for fault detection of key internal components of an uninterruptible power supply, characterized in that, The method includes: Determine whether the current operating mode is set to the training phase of the model: if it is in the training phase, then use historical data to perform data preprocessing and divide the dataset; if it is not in the training phase, then collect current data and perform data preprocessing. The historical data and current data include operating parameter data and image data, and the image data is optical detection image data. The runtime parameter data is input into a convolutional neural network model that combines incremental learning, and the image data is input into a tensor network model that uses multidimensional feature sampling and matrix multiplication state operations. The tensor network model using multidimensional feature sampling and matrix multiplication state operations consists of a multidimensional sampling layer, a matrix multiplication state network containing 4 MPS operations, another multidimensional sampling layer, a matrix multiplication state network containing 3 MPS operations, and a product state network containing 2 MPS operations, connected sequentially. The image is divided into 8 sub-regions. The multidimensional sampling layer includes average region sampling, maximum region sampling, and relevant region sampling. The correlation coefficient (crs) and tensor of the relevant regions are... The calculation formula is as follows: Where p represents the total number of pixels in each sub-region, (x,y) represents the pixel position in each sub-region, E(·) represents the Euclidean distance, and mps represents the matrix multiplication state. This represents a low-rank tensor. , ,…, represents the index of an N-dimensional tensor. , ,…, Represents the Nth tensor node of the mps operation; and The system then checks again whether the current operating mode is set to the model training phase: if it is in the training phase, the model is trained, and if the requirements are met, the trained model is deployed; if it is not in the training phase, the fault detection result is output.

2. The fault detection method for key internal components of an uninterruptible power supply as described in claim 1, characterized in that, The key internal components of the uninterruptible power supply include a battery, rectifier, inverter, and bypass isolation transformer. The operating parameter data includes the main input voltage value. Bypass input voltage value Inverter output voltage Inverter output current Inverter frequency value Load Battery charging voltage Battery charging current There are a total of 8 operating parameter variables; the image data includes optical detection images at four locations: battery, rectifier, inverter, and bypass isolation transformer.

3. The fault detection method for key internal components of an uninterruptible power supply as described in claim 2, characterized in that, For the eight operational parameter variables included in the historical data, data from the previous 30 days to the current moment are collected, and the average value of the parameter per minute is calculated once per hour to form an operational parameter dataset; for the optical detection images at the four locations, images from the previous 360 days to the current moment are collected, and images are collected once every 12 hours to form an image dataset.

4. The fault detection method for key internal components of an uninterruptible power supply as described in claim 3, characterized in that, The "using historical operational parameter data and image data to perform data preprocessing and dataset partitioning" includes: cleaning the operational parameter data by filling in missing values, replacing outliers, and smoothing noise values, and performing data normalization operations; and performing contrast enhancement, noise reduction, and image cropping operations on the image data. The "collecting current operational parameter data and image data and performing data preprocessing" includes: dividing the operational parameter dataset and image data dataset into training set and validation set according to a selected ratio.

5. The fault detection method for key internal components of an uninterruptible power supply as described in claim 2, characterized in that, The "collecting current data and performing data preprocessing" includes normalizing the running parameter data and cropping the image data.

6. The fault detection method for key internal components of an uninterruptible power supply as described in claim 2, characterized in that, "A convolutional neural network model that inputs the aforementioned operational parameter data and combines it with incremental learning" includes: Calculate the matching degree between the original sample and the new sample. Based on the value of Mi, it is determined how many weight parameters of the original convolutional network can be used to obtain a new convolutional network through parameter copying, i.e., the weight parameters of the new convolutional network = the weight parameters of the original convolutional network × + Weight parameters of the newly added sample convolutional network × After passing through a convolutional network, the data stream passes through a classifier consisting of three fully connected layers to obtain the fault detection result output, the matching degree. The calculation formula is as follows: The convolutional network consists of 5 convolutional and pooling blocks, where k is the feature size, n represents the new sample, and o represents the best matching sample of n. and Let n and o represent the i-th features of n and o respectively, and min(·) and max(·) represent the minimum and maximum value operations respectively. The range of its value is between (0,1].

7. The fault detection method for key internal components of an uninterruptible power supply as described in claim 6, characterized in that, The accuracy, precision, recall, and task suitability (TMD) metrics are used to determine whether the trained model meets the deployment requirements. If it does not meet the requirements, the process returns to the step of "using historical data for data preprocessing and dataset partitioning." The calculation formula for the task suitability (TMD) metric and the model deployment judgment indicator D are as follows: in, and Let represent the number of correct predictions and the number of incorrect predictions for the i-th task on the validation set, respectively. n is the number of tasks. apr represents the average of accuracy, precision, and recall. min(·) and avg(·) represent the minimum and average values, respectively.

8. A fault detection device for key internal components of an uninterruptible power supply, characterized in that, The fault detection device includes a data acquisition module, a host computer processing module, and a display and alarm module. The data acquisition module includes an image acquisition device, an electrical signal acquisition device, a data processing unit, and a data transmission line. The data processing unit is used to perform preprocessing operations on image data and operating parameter data. The host computer processing module includes a communication interface, an operating mode determiner, a server-side trainer, a side-end deployer, and a data storage unit. The communication interface is used for data stream transmission and information communication with other modules. The operating mode determiner is used to set whether the model is in the training phase or the non-training phase. The server-side trainer is used to complete the training process of the multi-task learning model. The edge-end deployer enables lightweight deployment of edge devices for trained models that meet deployment requirements. The computation results output by the algorithm model are stored in the data storage unit. The display and alarm module includes a communication interface, a result display unit, a result analysis unit, a fault alarm unit, and a data storage unit. The communication interface is used for data transmission and communication between different modules. The result display unit visualizes the model's computation results. The result analysis unit performs logical judgment and analysis based on the output results and sends instructions to the fault alarm unit. The data storage unit stores the result analysis and fault alarm records. The operating mode determiner determines whether the current operating mode is set to model. During the training phase: If in the training phase, the data processing unit uses historical data to perform data preprocessing and partition the dataset; if in the non-training phase, the data acquisition module collects current data, which is then preprocessed by the data processing unit. The historical and current data include runtime parameter data and image data, where the image data is optical detection image data. The runtime parameter data is input into a convolutional neural network model combining incremental learning, and the image data is input into a tensor network model using multidimensional feature sampling and matrix multiplication state operations. The mode determiner then determines whether the current running mode is set to the model's training phase: if in the training phase, the... The server-side trainer trains the model. If the training meets the requirements, the edge-side deployer deploys the trained model. If the model is not in the training phase, the result display unit outputs the fault detection result. The tensor network model using multidimensional feature sampling and matrix product state operations consists of a multidimensional sampling layer, a matrix product state network with 4 MPS operations, a multidimensional sampling layer, a matrix product state network with 3 MPS operations, and a product state network with 2 MPS operations connected sequentially. The image is divided into 8 sub-regions. The multidimensional sampling layer includes average region sampling, maximum region sampling, and relevant region sampling. The correlation crs and tensor of the relevant regions are... The calculation formula is as follows: Where p represents the total number of pixels in each sub-region, (x,y) represents the pixel position in each sub-region, E(·) represents the Euclidean distance, and mps represents the matrix multiplication state. This represents a low-rank tensor. , ,…, Indices representing the index of an N-dimensional tensor. , ,…, This represents the Nth tensor node of the mps operation.

Images