A method and device for uninterruptible power supply fault diagnosis of multi-source information

Abstract

A kind of uninterrupted power supply fault diagnosis method and device of multi-source information, method includes the following steps: obtaining the operating data of multiple modalities of uninterrupted power supply, and the operating data of multiple modalities is converted into target modal characteristic vector group;Causal relationship diagram between each variable of operating data is constructed, and target modal characteristic vector group is adjusted according to causal relationship diagram, and optimized modal characteristic vector group is obtained;Optimized modal characteristic vector group is obtained through cross-modal attention network to obtain fusion feature vector and attention weight matrix;Path reasoning is carried out through fault knowledge graph, and diagnostic path pointing to fault cause is generated;Integrate the fault classification result corresponding to fusion feature vector and diagnostic path, and generate fault diagnosis report.The technical scheme provided in the application is implemented, and the accuracy of uninterrupted power supply fault diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of electronic power technology, and in particular to a method and apparatus for diagnosing faults in an uninterruptible power supply with multi-source information. Background Technology

[0002] Uninterruptible power supplies (UPS) are widely used in data centers, communication base stations, industrial automation and other fields to provide temporary power support during power outages and ensure the normal operation of equipment.

[0003] Existing uninterruptible power supply (UPS) fault diagnosis solutions typically rely on real-time data from a single sensor or human experience to identify faults by monitoring changes in parameters such as voltage or current. However, UPS systems are complex (involving multiple components such as rectifiers, inverters, and battery packs). Relying solely on a single data source or human experience for fault diagnosis may result in missed or incorrect diagnoses when encountering complex or latent faults, posing a risk of low fault diagnosis accuracy. Summary of the Invention

[0004] In view of this, this application provides a method and apparatus for diagnosing uninterruptible power supply faults using multi-source information, in order to solve the above-mentioned problems.

[0005] Firstly, a multi-source information-based uninterruptible power supply (UPS) fault diagnosis method is provided, applied to an UPS fault diagnosis platform. This method includes:

[0006] The system acquires operational data from multiple modes of an uninterruptible power supply (UPS) and converts this data into a target mode feature vector set using multiple preset feature extraction models.

[0007] The operational data is analyzed by causal discovery algorithm to construct a causal relationship diagram between the variables in the operational data, and the target modal feature vector group is adjusted according to the causal relationship diagram to obtain the optimized modal feature vector group;

[0008] The optimized modal feature vector set is input into a pre-defined cross-modal attention network to obtain the fused feature vector and attention weight matrix;

[0009] By using a pre-defined fault knowledge graph, path reasoning is performed on target features with weight values ​​greater than a pre-defined threshold within the attention weight matrix to generate diagnostic paths pointing to the cause of the fault.

[0010] The fault classification results and diagnostic paths corresponding to the feature vectors are integrated and fused to generate a fault diagnosis report.

[0011] The above technical solutions, by acquiring multi-modal operational data and converting it into target modal feature vector sets, overcome the limitations of traditional single-source data, providing a multi-dimensional characterization of UPS operating status (including real-time status, event records, historical experience, etc.) and offering a comprehensive data foundation for diagnosis. Introducing a causal discovery algorithm to construct a variable causal relationship graph and adjust the target modal feature vector sets can identify core correlated variables, reduce redundant interference, and improve feature quality. A cross-modal attention network quantifies the importance of correlations between features from different modalities, achieving effective fusion while retaining unique information from each modality and highlighting key cross-modal correlations, solving the multi-modal fusion problem and improving fault classification accuracy. Based on a fault knowledge graph, path reasoning is performed on high-weight features to generate a diagnostic path from phenomenon to root cause, ensuring the results include both fault type and cause explanation. Integrating the results to generate a report provides complete evidence for operation and maintenance, significantly improving the accuracy of UPS fault diagnosis.

[0012] Optionally, the system acquires operational data for multiple modes of the uninterruptible power supply (UPS), and converts this operational data into a target mode feature vector set using multiple preset feature extraction models. Specifically, this includes:

[0013] Acquire sensor time-series data collected by a multi-type sensor group pre-deployed on an uninterruptible power supply, and process the sensor time-series data through a pre-set convolutional neural network model to generate sensor time-series feature vectors;

[0014] Acquire the operation log text data of the uninterruptible power supply output recording device status and events, and encode the operation log text data through a preset natural language processing model to generate semantic feature vectors;

[0015] Acquire historical fault record data of uninterruptible power supply, construct a fault association graph based on the historical fault record data, and process the fault association graph through a preset graph neural network model to generate fault map feature vectors;

[0016] The sensor time-series feature vector, semantic feature vector, and fault map feature vector are time-aligned to form the target modal feature vector group.

[0017] The above technical solutions address the structural differences between sensor time-series data, log text data, and historical fault record data by employing convolutional neural networks, natural language processing models, and graph neural networks for feature extraction, respectively. This maximizes the preservation of unique information in each modality (such as temporal dynamics, semantic meaning, and fault correlation patterns). Time alignment ensures the consistency of features across different modalities in the time dimension, avoiding correlation distortion caused by temporal misalignment, and providing a high-quality, temporally unified feature foundation for subsequent causal analysis and cross-modal fusion.

[0018] Optionally, the operational data is analyzed using a causal discovery algorithm to construct a causal relationship diagram among the variables in the operational data. Based on this diagram, the target modal feature vector set is adjusted to obtain an optimized modal feature vector set, specifically including:

[0019] In a causal relationship graph, the variables in the running data corresponding to nodes with an in-degree of zero are identified as target variables;

[0020] In the target modal feature vector set, locate the modal feature vector to be optimized generated from the running data containing the target variable;

[0021] Using a pre-defined attribution analysis method, the attribution score of each dimension of the modality feature vector to be optimized to the target variable is calculated.

[0022] All dimensions with attribution scores greater than a preset attribution threshold are used as target dimensions;

[0023] For each target dimension, the weights are adjusted using preset weight coefficients to generate an optimized modal feature vector;

[0024] Replace the feature vector of the modality to be optimized with the feature vector of the optimized modality to obtain the set of optimized modal feature vectors.

[0025] The above technical solution refines the process of adjusting the feature vector group based on the causal relationship graph. It generates an optimized modal feature vector group through steps such as identifying the target variable, locating the vector to be optimized, calculating attribution scores, and adjusting the weights of the target dimensions. By leveraging the causal relationship graph to accurately identify the target variable with core driving force (nodes with an in-degree of zero), and quantifying the influence of feature dimensions on the target variable through attribution analysis, it achieves targeted weighting of key feature dimensions while weakening interference from irrelevant dimensions. This adjustment makes the feature vector group more focused on the causal features that play a decisive role in fault diagnosis, improving the effectiveness and relevance of the features and reducing the interference of redundant information in the subsequent diagnostic process.

[0026] Optionally, the optimized modal feature vector set is input into a pre-defined cross-modal attention network to obtain a fused feature vector and an attention weight matrix, specifically including:

[0027] Matrix multiplication is performed on the target feature vector in the optimized modal feature vector group through preset query linear projection layer, key linear projection layer and value linear projection layer respectively, to generate corresponding query vector, key vector and value vector for the target feature vector. The target feature vector is any feature vector in the optimized modal feature vector group.

[0028] By using a cross-modal attention network to exchange information between the query vector and the key vector, the correlation importance between any two different modalities can be calculated. Any two different modalities represent any two feature vectors in the optimized modal feature vector group.

[0029] Based on the importance of association, the value vectors of each optimized modality feature vector are weighted and fused to generate a fused feature vector;

[0030] An attention weight matrix is ​​constructed based on the importance of association.

[0031] The above technical solution clarifies the specific processing flow of the cross-modal attention network, including generating query, key, and value vectors, calculating the importance of intermodal associations, and fusing to generate fusion features and weight matrices. A linear projection layer provides a unified interactive computational foundation for features from different modalities. The cross-modal attention mechanism quantifies the importance of associations between any two modalities (such as the association strength between sensor features and log features), enabling the fusion process to dynamically focus on key association information and avoid information loss caused by simple splicing. The generated fusion feature vector integrates multimodal core information to improve classification accuracy, while the attention weight matrix provides a basis for feature importance in subsequent knowledge graph reasoning, achieving effective connection between feature fusion and reasoning processes.

[0032] Optionally, matrix multiplication is performed on the target feature vector in the optimized modality feature vector group through preset query linear projection layers, key linear projection layers, and value linear projection layers, respectively, to generate corresponding query vectors, key vectors, and value vectors for the target feature vectors, specifically including:

[0033] The target feature vector is multiplied by the query weight matrix of the query linear projection layer to generate the query vector.

[0034] The target feature vector is multiplied by the key weight matrix of the key linear projection layer to generate the key vector.

[0035] The target feature vector is multiplied by the value weight matrix of the linear projection layer to generate a value vector.

[0036] The above technical solution achieves a linear transformation of the feature vector through matrix multiplication with the preset weight matrix, mapping the original features to a higher-dimensional space that is more suitable for calculating the importance of intermodal associations. This transformation can highlight the key information in the features that is related to the association calculation, suppress noise, and ensure that the generated query, key, and value vectors accurately capture the essential attributes of each modality feature. This provides a reliable vector representation for the accurate calculation of the importance of associations in cross-modal attention mechanisms and improves the effectiveness of intermodal interactions.

[0037] Optionally, a cross-modal attention network can be used to interact with the query vector and the key vector to calculate the association importance between any two different modalities, specifically including:

[0038] For any two optimized modal feature vectors of different modalities, the query vector of the first modality and the key vector of the second modality are calculated by a preset similarity function through the cross-attention mechanism of the cross-modal attention network to obtain the initial attention score. Here, any two different modalities include the first modality and the second modality.

[0039] The initial attention scores are normalized to obtain attention weights, which are then used as the correlation importance.

[0040] The above technical solutions utilize a cross-attention mechanism to enable direct interaction between features from different modalities (such as the query vector of the first modality and the key vector of the second modality), and combine this with a similarity function to effectively measure the degree of correlation between modalities. The normalization process transforms the initial scores into comparable attention weights, accurately quantifies the importance of different modal features to each other, ensures the rationality of weight allocation during cross-modal fusion, and prioritizes the retention of closely related key information in the fused features, thereby improving the quality of fused features and the accuracy of subsequent diagnosis.

[0041] Optionally, a pre-defined fault knowledge graph is used to perform path reasoning on target features with weight values ​​greater than a preset threshold within the attention weight matrix, generating diagnostic paths pointing to the cause of the fault. Specifically, this includes:

[0042] By using preset feature mapping rules, each target feature is mapped to a starting node in the fault knowledge graph;

[0043] Based on the fault knowledge graph, a graph search algorithm is used to search for a reasoning path that connects the starting node and points to the root node representing the cause of the fault, starting from the starting node, and then uses the reasoning path as the diagnostic path.

[0044] The above technical solutions establish a bridge between data features and domain knowledge by associating high-weight target features with knowledge graph nodes through feature mapping rules. The graph search algorithm traces from the starting node (phenomenon feature) to the root node (fault cause), generating a complete reasoning path of "phenomenon-intermediate factor-root cause", which upgrades the diagnostic results from a simple "fault classification" to an interpretable "fault cause chain". This path reasoning solves the problem of traditional diagnostic results "knowing what but not why", providing clear clues for operation and maintenance personnel and significantly improving the efficiency of fault location.

[0045] Secondly, a multi-source information uninterruptible power supply fault diagnosis system is provided. The system includes a data acquisition module, a data processing module, and a fault diagnosis module, wherein:

[0046] The data acquisition module is configured to acquire operational data of multiple modes of uninterruptible power supply, and convert the operational data of multiple modes into a target mode feature vector group through multiple preset feature extraction models;

[0047] The data processing module is configured to analyze the running data using a causal discovery algorithm, construct a causal relationship diagram between the variables in the running data, and adjust the target modal feature vector group based on the causal relationship diagram to obtain an optimized modal feature vector group.

[0048] The data processing module is also configured to input the optimized modal feature vector group into a preset cross-modal attention network to obtain the fused feature vector and attention weight matrix;

[0049] The fault diagnosis module is configured to use a preset fault knowledge graph to perform path reasoning on target features with weight values ​​greater than a preset threshold in the attention weight matrix, and generate a diagnostic path pointing to the cause of the fault.

[0050] The fault diagnosis module is also configured to integrate and fuse the fault classification results and diagnostic paths corresponding to the feature vectors to generate a fault diagnosis report.

[0051] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the foregoing.

[0052] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed, perform the method described in any of the preceding descriptions.

[0053] In summary, implementing one or more technical solutions provided in this application has at least the following technical effects or advantages:

[0054] By acquiring multi-modal operational data and converting it into feature vector sets, this approach overcomes the limitations of traditional single-source data, providing a multi-dimensional characterization of UPS operational status (including real-time status, event logs, and historical experience), thus offering a comprehensive data foundation for diagnosis. A causal discovery algorithm is introduced to construct a variable causal relationship graph and adjust the feature vector sets, identifying core correlated variables, reducing redundant interference, and improving feature quality. A cross-modal attention network quantifies the importance of correlations between features from different modalities, achieving effective fusion. This retains unique information from each modality while highlighting key cross-modal correlations, solving the multi-modal fusion challenge and improving fault classification accuracy. Based on a fault knowledge graph, path reasoning is performed on high-weight features to generate a diagnostic path from phenomenon to root cause, ensuring the results include both fault type and cause explanation. The integrated results generate a report, providing complete evidence for operation and maintenance, significantly improving the accuracy of UPS fault diagnosis. Attached Figure Description

[0055] Figure 1 This is an exemplary system architecture diagram of an uninterruptible power supply fault diagnosis method or a multi-source information uninterruptible power supply fault diagnosis system that applies the present application.

[0056] Figure 2 This is a flowchart illustrating a multi-source information uninterruptible power supply fault diagnosis method according to an embodiment of this application.

[0057] Figure 3 This is a schematic diagram of a multi-source information uninterruptible power supply fault diagnosis system according to an embodiment of this application;

[0058] Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in the application embodiment.

[0059] Explanation of reference numerals in the attached figures: 100, System architecture; 101, First terminal device; 102, Second terminal device; 103, Third terminal device; 104, Network; 105, Server; 301, Data acquisition module; 302, Data processing module; 303, Fault diagnosis module; 401, Processor; 402, Communication bus; 403, User interface; 404, Network interface; 405, Memory. Detailed Implementation

[0060] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0061] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0062] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0063] Figure 1 This paper illustrates an exemplary system architecture diagram of an embodiment of an uninterruptible power supply (UPS) fault diagnosis method or a UPS fault diagnosis system based on multi-source information, which can be applied according to this application.

[0064] like Figure 1 As shown, the system architecture 100 may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium to provide communication links between the terminal devices 101, 102, 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0065] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as model training applications, video recognition applications, web browser applications, social platform software, etc.

[0066] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, laptops, and desktop computers, etc. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0067] When terminals 101, 102, and 103 are hardware devices, video capture devices can also be installed on them. These video capture devices can be various devices capable of capturing video, such as cameras, sensors, etc. Users can use the video capture devices on terminals 101, 102, and 103 to capture video.

[0068] Server 105 can be a server that provides various services, such as a backend server for processing data displayed on terminal devices 101, 102, and 103. The backend server can analyze and process the received data and can feed back the processing results (such as recognition results) to the terminal devices.

[0069] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0070] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. In particular, if the target data does not need to be obtained remotely, the above system architecture may exclude the network and include only terminal devices or servers.

[0071] Figure 2This is a flowchart illustrating a multi-source information uninterruptible power supply (UPS) fault diagnosis method according to an embodiment of this application. This method can be implemented using a computer program, a microcontroller, or run on a multi-source information UPS fault diagnosis system. The computer program can be integrated into the application or run as a standalone utility application. The specific steps of the multi-source information UPS fault diagnosis method are described in detail below.

[0072] S201: Acquire the operating data of the uninterruptible power supply in multiple modes, and convert the operating data of multiple modes into a target mode feature vector group through multiple preset feature extraction models.

[0073] For example, the fault diagnosis platform acquires multi-modal operating data of the UPS. Multi-modal operating data refers to heterogeneous, multi-source, and semantically related data sets acquired through different acquisition methods in the same system or scenario, such as time-series data collected by multiple types of sensors, equipment operation log text data, and historical fault record data. These data are processed by multiple preset feature extraction models. The preset feature extraction models are pre-designed, trained, and weighted models. Their core function is to automatically convert raw data (such as images, text, time-series signals, etc.) into low-dimensional feature vectors with high discriminative power. For example, a convolutional neural network model is used to process sensor time-series data to generate sensor time-series feature vectors, a natural language processing model is used to process operation log text data to generate semantic feature vectors, and a graph neural network model is used to generate fault map feature vectors based on fault association graphs constructed from historical fault records. Finally, these three feature vectors are time-aligned to form a target modality feature vector group.

[0074] In one possible implementation, the operation data of the uninterruptible power supply (UPS) in multiple modes are acquired, and the operation data of multiple modes are converted into a target mode feature vector set through multiple preset feature extraction models. Specifically, this includes: acquiring sensor time-series data collected by a multi-type sensor group pre-deployed on the UPS, and processing the sensor time-series data through a preset convolutional neural network model to generate sensor time-series feature vectors; acquiring operation log text data of the UPS outputting the recording device status and events, and encoding the operation log text data through a preset natural language processing model to generate semantic feature vectors; acquiring historical fault record data of the UPS, constructing a fault association graph based on the historical fault record data, and processing the fault association graph through a preset graph neural network model to generate fault map feature vectors; and aligning the sensor time-series feature vectors, semantic feature vectors, and fault map feature vectors in time to form a target mode feature vector set.

[0075] Specifically, the fault diagnosis platform first acquires sensor timing data pre-deployed in the UPS using multiple sensor groups. These sensor groups include voltage sensors (deployed at the input, output, and DC bus terminals to collect A-phase / B-phase input voltage, output voltage of each phase, and DC bus voltage), current sensors (deployed in the input and output circuits to collect input and output current of each phase), temperature sensors (attached to IGBT (Insulated Gate Bipolar Transistor) modules, heat sinks, and other key components to monitor operating temperature), and speed sensors (installed on the fan unit to record fan speed). The platform then processes this timing data using a pre-defined convolutional neural network model to generate sensor timing feature vectors. For example, a 500kVA three-phase UPS deployed in a data center may have multiple high-precision sensors deployed at key nodes (such as the input, output, DC bus, IGBT modules, and fan units). The platform collects 32 key physical quantities in real time at a sampling rate of 10kHz, including A-phase input voltage, B-phase output current, DC bus voltage, IGBT module temperature, and fan speed. During diagnostics, the platform extracts data from the past second as an analysis window, thus obtaining a time-series data set containing 10,000 sampling points for each physical quantity. These 32 time-series data sets form a 32×10000 two-dimensional data matrix. This matrix is ​​input into a pre-defined one-dimensional convolutional neural network model. This model contains three convolutional layers (using 64, 128, and 256 convolutional kernels respectively), three max-pooling layers, and one fully connected layer. Through convolution and pooling operations, the model can automatically learn and extract deep patterns such as periodicity, instantaneous impacts, and waveform distortion hidden in the time-series data. Finally, the fully connected layer maps the extracted features into a 128-dimensional sensor time-series feature vector, which highly condenses the electrical and physical operating status of the UPS within that time window.

[0076] Furthermore, the platform acquires the operational log text data from the uninterruptible power supply (UPS), which records device status and events. This operational log text data is then encoded using a pre-defined natural language processing model to generate semantic feature vectors. Within the same time window of the UPS, its internal monitoring and management unit may generate several operational logs. For example, the platform acquires the following two log texts: "2025-07-07 12:40:15 [WARNING] Mains input A-phase voltage drops below threshold" and "2025-07-07 12:40:16 [INFO] System switches to battery power mode". The platform integrates these log texts generated within the time window. Subsequently, the integrated text is input into a pre-trained natural language processing model based on the BERT (Bidirectional Encoder Representations from Transformers) architecture. The model first segments the text and adds special CLS (Classification) and SEP (Separator) tags (CLS tags represent the main idea or central theme of a sentence, and SEP tags separate different sentences or sequences). Then, it performs deep semantic understanding through its internal multi-layer Transformer encoder. The model's final output, a hidden state vector corresponding to the CLS tags, is used as the aggregate semantic representation of the entire text segment. This hidden state vector is a 768-dimensional semantic feature vector that not only captures the literal meaning of the log but also provides a deeper understanding of the event correlations, alarm levels, and potential fault symptoms contained within.

[0077] Furthermore, the platform acquires historical fault records of the uninterruptible power supply (UPS), constructs a fault association graph based on the historical fault data, and processes the fault association graph using a pre-defined graph neural network model to generate fault map feature vectors. For example, the platform retrieves all historical fault records of the UPS from the operation and maintenance database. Each record contains fields such as fault time, fault code, faulty component, and associated alarms. Based on this data, the platform constructs a fault association graph. In this graph, each node represents a faulty component (such as "IGBT module A1" or "fan unit F3") or a fault event (such as "inverter over-temperature"). If two events or components frequently occur together within a short period of time in historical faults, an edge is established between their corresponding nodes, and the weight of the edge represents their co-occurrence frequency. This constructs a knowledge network that reflects the fault propagation path and component correlation. Subsequently, a pre-defined graph convolutional network (GCN) model is applied to this fault association graph. GCN learns the embedding representation of each node in the graph by aggregating information from neighboring nodes. Finally, by performing graph-level pooling operations (such as average pooling) on ​​the embedding vectors of all nodes in the graph, a 256-dimensional fault map feature vector that can represent an overview of the entire UPS historical fault modes is generated.

[0078] Furthermore, the platform aligns the sensor time-series feature vectors, semantic feature vectors, and fault map feature vectors generated in the preceding steps to form a target modal feature vector set. For example, for the diagnostic time point "12:40:16 on August 7, 2025", the platform has generated three corresponding feature vectors: a 128-dimensional sensor time-series feature vector representing the electrical state within the second from 12:40:15 to 12:40:16, a 768-dimensional semantic feature vector representing log information near that time, and a 256-dimensional fault map feature vector representing the historical fault background of the current device. The platform organizes these three feature vectors from different data sources but describing the same diagnostic time point as a set, thus ultimately forming the target modal feature vector set for the next step of analysis.

[0079] S202: Analyze the running data using a causal discovery algorithm, construct a causal relationship diagram between the variables in the running data, and adjust the target modal feature vector group based on the causal relationship diagram to obtain an optimized modal feature vector group.

[0080] For example, the platform first applies a causal discovery algorithm (such as the PC (Peter-Clark) algorithm) to analyze the variables of various modal operational data and constructs a directed acyclic graph representing the causal relationships between variables. Then, the variables corresponding to the nodes with an in-degree of zero in the graph are identified as target variables, and their corresponding modal feature vectors to be optimized are located. Next, the attribution score of each variable in the vector is calculated by the attribution analysis method (such as the SHAP (SHapley Additive exPlanations) algorithm), and target variables with scores higher than a preset threshold are selected. Finally, these target variables are enhanced with preset weight coefficients to generate optimized modal feature vectors, which replace the original vectors to be optimized, resulting in the final optimized modal feature vector set.

[0081] In one possible implementation, the operational data is analyzed using a causal discovery algorithm to construct a causal relationship graph among the variables in the operational data. Based on this graph, the target modal feature vector set is adjusted to obtain an optimized modal feature vector set. Specifically, this includes: identifying the variables in the operational data corresponding to nodes with zero in-degree in the causal relationship graph as target variables; locating the modal feature vectors to be optimized generated from the operational data containing the target variables within the target modal feature vector set; calculating the attribution score of each dimension of the modal feature vectors to be optimized relative to the target variables using a preset attribution analysis method; designating all dimensions with attribution scores greater than a preset attribution threshold as target dimensions; adjusting the weights of each target dimension using preset weight coefficients to generate optimized modal feature vectors; and replacing the modal feature vectors to be optimized with optimized modal feature vectors to obtain the optimized modal feature vector set.

[0082] The purpose of this step is to proactively identify and enhance the feature signals most relevant to the root cause of the fault by mining the deep causal relationships behind the data before performing multimodal feature fusion, and to suppress the interference of spurious correlation features, thereby improving the accuracy and robustness of the diagnostic model.

[0083] Specifically, the platform first applies a causal discovery algorithm to analyze all variables in the multi-modal operational data acquired in step S201, constructing a causal relationship graph that represents the causal relationships between the variables. For example, the platform integrates variables from sensor time-series data and operational log text data. Thirty-two physical quantities from the sensor data (such as "phase A input voltage" and "IGBT module temperature") are directly used as variables. Simultaneously, key events (such as "mains voltage drop" and "switching to battery mode") extracted from the log text through rule matching are also used as event variables. The platform employs a constraint-based PC causal discovery algorithm. This algorithm first performs conditional independence checks on all variable pairs, constructing an undirected graph skeleton. For example, the check finds that "phase A input voltage" and "switching to battery mode" are independent given the "mains voltage drop" event. Subsequently, the algorithm applies rules such as V-structure recognition to determine the direction of the edges in the graph. Finally, a directed acyclic graph is generated as the causal relationship graph. For example, the diagram might contain a path like this: [Main grid fluctuation] → [Phase A input voltage drop] → [System switches to battery mode], where the arrows indicate the cause and effect.

[0084] Furthermore, in the constructed causal relationship graph, the platform identifies the variables in the operational data corresponding to nodes with an in-degree of zero (i.e., nodes that appear only as causes in the graph and are not affected by other variables) as target variables. Subsequently, in the target modal feature vector group generated in step S201, the platform locates the modal feature vector to be optimized generated from the operational data containing the target variable. For example, in the causal relationship graph generated above, the platform identifies the node "mains power grid fluctuation" as having an in-degree of zero. Therefore, the platform identifies the original data variable "A-phase input voltage" associated with "mains power grid fluctuation" as the target variable for this diagnostic event. Since "A-phase input voltage" is part of the sensor time series data, the platform locates the 128-dimensional sensor time series feature vector generated from the sensor time series data as the modal feature vector to be optimized in this case within the target modal feature vector group.

[0085] Furthermore, the platform uses a pre-defined attribution analysis method to calculate the attribution score of each component (each component represents a specific value in the corresponding dimension of the vector) of the modal feature vector to be optimized for the target variable, and identifies all components with attribution scores greater than a pre-defined attribution threshold as target components (i.e., target dimensions). For example, the platform uses the SHAP algorithm as the attribution analysis method. For the 128-dimensional sensor time-series feature vector located in the previous step, the SHAP algorithm calculates the contribution of each of its 128 components to the specific situation of voltage drop in the target variable "phase A input voltage". This contribution is the attribution score. For example, after calculation, it is found that the attribution scores of components 5, 12, and 38 are 0.85, 0.91, and 0.88, respectively, while the attribution scores of other components are all below 0.5. If the preset attribution threshold is 0.8 (this threshold is used to select the feature components most relevant to the target variable from the feature vector of the modality to be optimized. Its core purpose is to achieve a balance between "noise reduction" and "preservation of effective information". Too low a threshold will introduce interference from spurious correlations or noisy features; too high a threshold may miss minor but still diagnostically valuable features), then the platform will determine the three components with attribution scores greater than the threshold, namely the 5th, 12th and 38th, as the target components that need to be weighted in this operation. These target components mathematically represent the features used by the model to encode key information such as voltage drops and waveform distortion.

[0086] Furthermore, the platform enhances each identified target component using a preset weighting coefficient, generating an optimized modal feature vector. Finally, this optimized modal feature vector replaces the unoptimized modal feature vector in the target modal feature vector group, resulting in the final optimized modal feature vector group. For example, the platform sets a preset weighting coefficient of 1.5 (this coefficient enhances the signal strength of the target dimension, giving it more attention in subsequent cross-modal fusion. Its value should be significantly greater than 1, but not too large, to prevent numerical instability or overfitting of the model to a few features). Subsequently, the 5th, 12th, and 38th target components identified in the previous step are enhanced by multiplying their values ​​by 1.5, while keeping the values ​​of the remaining 125 components unchanged. After this operation, the original 128-dimensional sensor time-series feature vector is modified into a new "optimized modal feature vector" whose internally relevant signals are significantly amplified. Finally, the platform replaces the original sensor time-series feature vector with this newly generated optimized modal feature vector from the original "target modal feature vector set" containing three vectors. This results in an optimized modal feature vector set enhanced with causal knowledge.

[0087] S203: Input the optimized modal feature vector group into the preset cross-modal attention network to obtain the fused feature vector and attention weight matrix.

[0088] For example, the platform generates corresponding query, key, and value vectors for each vector in the optimized modal feature vector group through multiple independent linear projection layers; it calculates the importance of association between different modalities through a cross-modal attention network (such as Scaled Dot-ProductAttention) to obtain attention weights; it weights and fuses the value vectors of each modality based on the attention weights to generate an update vector and concatenates them into a fused feature vector; and it integrates all attention weights to construct an attention weight matrix representing the modal association strength.

[0089] In one possible implementation, the optimized modality feature vector set is input into a preset cross-modal attention network to obtain a fused feature vector and an attention weight matrix. Specifically, this includes: performing matrix multiplication on the target feature vector in the optimized modality feature vector set through preset query linear projection layers, key linear projection layers, and value linear projection layers to generate corresponding query vectors, key vectors, and value vectors for the target feature vector, where the target feature vector is any one of the feature vectors in the optimized modality feature vector set; performing information interaction between the query vector and key vector through the cross-modal attention network to calculate the correlation importance between any two different modalities, where any two different modalities represent any two feature vectors in the optimized modality feature vector set; weighting and fusing the value vectors of each optimized modality feature vector based on the correlation importance to generate a fused feature vector; and constructing an attention weight matrix based on the correlation importance.

[0090] The purpose of this step is to perform deep and dynamic fusion of multiple modal feature vectors enhanced with causal knowledge. By simulating human attention mechanisms, the network can automatically learn and quantify the importance of interrelationships between different data sources, thereby generating a single fused feature that is more informative and better reflects the nature of the fault, providing the optimal input for the final fault classification.

[0091] Specifically, the platform first processes each feature vector in the optimized modal feature vector group through multiple preset, independent linear projection layers, generating a corresponding query vector, key vector, and value vector for each feature vector. For example, the input optimized modal feature vector group contains three vectors: an optimized 128-dimensional sensor time-series feature vector, an unoptimized 768-dimensional semantic feature vector, and an unoptimized 256-dimensional fault map feature vector. The platform internally presets three sets of independent linear projection layers, each containing a query layer, a key layer, and a value layer, with the output dimension of all projection layers set to 64. First, the 128-dimensional sensor time-series feature vector is simultaneously input into the three linear projection layers of the first group, generating three 64-dimensional vectors: Q_sensor, K_sensor, and V_sensor through multiplication with their respective weight matrices. Next, the 768-dimensional semantic feature vector is processed in the same way, generating three 64-dimensional vectors: Q_log, K_log, and V_log, through a second set of independent linear projection layers. Finally, the 256-dimensional fault map feature vector is also processed through a third set of independent linear projection layers, generating three 64-dimensional vectors: Q_graph, K_graph, and V_graph. Thus, the platform has generated its own Q, K, and V vectors, each representing a different role, for each of the three input modal feature vectors.

[0092] Furthermore, the platform uses a cross-modal attention network to interact with the query vector and key vector generated in the previous step, calculating the importance of the association between any two different modalities. The platform employs a standard scaled dot product attention mechanism. Taking the sensor modality as an example, to calculate the attention it should invest in other modalities, the platform performs dot product operations between its query vector Q_sensor and the key vectors (K_sensor, K_log, K_graph) of all modalities, and then divides the result by a scaling factor (usually the square root of the dimension of the K vector, i.e., ...). This process yields three initial attention scores. These scores are then normalized using the Softmax function to obtain a set of attention weights that sum to 1. For example, the calculated weights might be: [Weight(sensor->sensor) = 0.2, Weight(sensor->log) = 0.7, Weight(sensor->graph) = 0.1], where Weight(sensor->sensor) represents the sensor modality's attention weight to its own information, Weight(sensor->log) represents the sensor modality's attention weight to log modality information, and Weight(sensor->graph) represents the sensor modality's attention weight to historical fault graph modality information. This result mathematically quantifies "association importance," indicating that in the current diagnostic scenario, the sensor modality considers log modality information to be the most important (weight 0.7), its own information to be second most important (weight 0.2), and the historical graph information to have the lowest correlation (weight 0.1). The platform repeats this process for the query vector (Q_log, Q_graph) of each modality, and finally calculates the pairwise importance of all modalities.

[0093] Furthermore, based on the calculated correlation importance, the platform performs weighted fusion of the value vectors of each modality to generate the final fused feature vector. Taking the sensor modality as an example, the platform uses the weights [0.2, 0.7, 0.1] calculated in the previous steps to perform a weighted summation of the value vectors (V_sensor, V_log, V_graph) of all modalities. The calculation formula is: Updated_V_sensor = 0.2 × V_sensor + 0.7 × V_log + 0.1 × V_graph. The calculated result, Updated_V_sensor, is a new 64-dimensional vector that fuses the information from all modalities and is an updated version of the sensor modality. The platform performs the same operation on the log and graph modalities to obtain their respective updated vectors, Updated_V_log and Updated_V_graph. Finally, the platform concatenates these three 64-dimensional updated vectors to form a single 192-dimensional (64+64+64) fused feature vector. This vector is the final representation of all modal information after dynamic weighting and deep interaction.

[0094] Furthermore, the platform constructs an attention weight matrix based on all calculated association importances. For example, the platform integrates all calculated attention weights into an N×N matrix (N represents the number of modalities), which is a 3×3 matrix in this example. The element in the i-th row and j-th column of the matrix represents the attention weight of the i-th modality on the j-th modality. For example, the matrix is ​​shown below:

[0095]

[0096] The first row and first column represent the sensor mode, the second row and second column represent the log mode, and the third row and third column represent the graph mode. This attention weight matrix intuitively shows the correlation strength between features from different data sources at the current diagnostic time (for example, the first row shows that the correlation between sensor data and log data is the strongest), providing key and quantifiable evidence for subsequent fault path reasoning.

[0097] In one possible implementation, matrix multiplication is performed on the target feature vector in the optimized modality feature vector group through preset query linear projection layers, key linear projection layers, and value linear projection layers to generate corresponding query vectors, key vectors, and value vectors for the target feature vectors. Specifically, this includes: performing matrix multiplication on the target feature vector with the query weight matrix of the query linear projection layer to generate a query vector; performing matrix multiplication on the target feature vector with the key weight matrix of the key linear projection layer to generate a key vector; and performing matrix multiplication on the target feature vector with the value weight matrix of the value linear projection layer to generate a value vector.

[0098] Specifically, the optimized 128-dimensional sensor temporal feature vector from the optimized modal feature vector set is used as the target feature vector for current processing. The platform's preset target output dimension (i.e., the dimensions of the Q, K, and V vectors) is 64. The platform internally pre-defines three linear projection layers with different weight matrices: a query linear projection layer containing a 128×64 query weight matrix (Wq); a key linear projection layer containing a 128×64 key weight matrix (Wk); and a value linear projection layer containing a 128×64 value weight matrix (Wv). The platform performs matrix multiplication operations on this 128-dimensional target feature vector (which can be considered a 1×128 matrix) with the three weight matrices mentioned above. First, it multiplies it with the 128×64 query weight matrix Wq to generate a 1×64 matrix, i.e., a 64-dimensional query vector (Q_sensor). Next, it is multiplied with the key weight matrix Wk in the same way to generate a 64-dimensional key vector (K_sensor). Finally, it is multiplied with the value weight matrix Wv to generate a 64-dimensional value vector (V_sensor). Through this series of parallel matrix multiplication operations, the platform successfully generates its own unique, dimension-uniform Q, K, and V vectors for the input sensor time-series feature vectors. For other feature vectors in the optimized modal feature vector group (such as semantic feature vectors and fault map feature vectors), the platform performs the same processing flow through their respective linear projection layers of the corresponding size.

[0099] In one possible implementation, a cross-modal attention network is used to interact with the query vector and the key vector to calculate the association importance between any two different modalities. Specifically, this includes: for the optimized modal feature vectors of any two different modalities, the query vector of the first modality and the key vector of the second modality are calculated using a preset similarity function through the cross-attention mechanism of the cross-modal attention network to obtain an initial attention score, wherein any two different modalities include the first modality and the second modality; the initial attention score is normalized to obtain the attention weight, and the attention weight is used as the association importance.

[0100] Specifically, the platform uses a cross-attention mechanism within a cross-modal attention network to calculate an initial attention score by comparing the query vector of one modality (as the first modality) with the key vectors of all modalities (as the second modality) using a preset similarity function. For example, the sensor modality is used as the current "first modality," and its 64-dimensional query vector (Q_sensor) generated in step S203 is used. The preset similarity function is the scaled dot product. The platform interacts with the Q_sensor with the key vectors of all modalities, namely K_sensor, K_log, and K_graph. First, the dot product of Q_sensor and K_sensor is calculated, and the result is divided by a scaling factor of 8 (the square root of the K vector dimension 64) to obtain an initial attention score representing the "sensor modality on itself," for example, a score of 5.6. Next, the dot product of Q_sensor and K_log is calculated and scaled in the same way to obtain an initial attention score representing the "sensor mode versus log mode," for example, a result of 12.5. Finally, the dot product of Q_sensor and K_graph is calculated and scaled to obtain an initial attention score representing the "sensor mode versus graph mode," for example, a result of 3.2. Thus, the platform obtains a set of initial attention scores: [5.6, 12.5, 3.2].

[0101] Furthermore, the platform normalizes the obtained initial attention scores to obtain the final attention weights, which are then used as the quantified correlation importance. For example, the Softmax function is used to normalize the initial attention score vector [5.6, 12.5, 3.2] obtained in the previous step. The Softmax function can convert a set of arbitrary real numbers into a probability distribution that sums to 1, where the larger the input value, the higher the probability of the output. After calculation by the Softmax function, the original score vector is converted into a new weight vector, for example: [0.006, 0.991, 0.003]. Each value in this vector is the final attention weight of the sensor modality to other modalities, i.e., the correlation importance in this embodiment. This result clearly shows that, in the current diagnostic scenario, the sensor modality considers the information of the log modality to be of absolute importance (weight 0.991), while its own information and the information of the historical graph modality are almost negligible. This quantification result provides precise guidance for subsequent weighted fusion.

[0102] S204: Using a pre-defined fault knowledge graph, perform path reasoning on target features with weight values ​​greater than a pre-defined threshold within the attention weight matrix to generate a diagnostic path pointing to the cause of the fault.

[0103] For example, the platform first traverses the attention weight matrix and identifies the interaction modality features with weights exceeding a preset threshold as target features; then, it maps them to the starting node in the fault knowledge graph through a preset feature mapping rule; subsequently, based on the graph, it uses a graph search algorithm to search for a reasoning path that connects the starting node and points to the root cause node; if a path is found, it is determined as the final diagnostic path.

[0104] In one possible implementation, a path reasoning is performed on target features with weight values ​​greater than a preset threshold within an attention weight matrix using a preset fault knowledge graph to generate a diagnostic path pointing to the cause of the fault. Specifically, this includes: mapping each target feature to a starting node in the fault knowledge graph using a preset feature mapping rule; and using a graph search algorithm to search for a reasoning path that connects the starting node to a root node representing the cause of the fault, based on the fault knowledge graph, and using the reasoning path as the diagnostic path.

[0105] The purpose of this step is to transform the abstract numerical results (attention weights) output by the deep learning model in the previous steps into a clear and understandable diagnostic reasoning chain that conforms to the logic of human experts. This transforms the entire diagnostic process from an opaque black box into a transparent and traceable white box, greatly enhancing the interpretability and credibility of the diagnostic results.

[0106] Specifically, the platform first traverses the attention weight matrix, identifying the pair of modal features corresponding to those interacting with each other, whose weight values ​​are greater than a preset threshold, as target features. For example, consider the 3×3 attention weight matrix generated in step S203. The platform's preset weight threshold is 0.6 (this threshold is used to filter the most reliable cross-modal associations from the attention weight matrix to construct an interpretable diagnostic path. Its selection focuses more on the logic and interpretability of the diagnostic path than simply pursuing numerical indicators). By traversing this matrix, the platform finds that the weight value in the first row and second column is 0.7, which is greater than 0.6. This weight value represents the attention paid by the "sensor modality" to the "log modality." Therefore, the platform identifies this pair of interacting modal features, i.e., (sensor feature, log feature), as the "target features" for this diagnosis. Mathematically, this target feature indicates that the model believes there is a strong correlation between anomalies in the sensor data and events recorded in the log text, which is the most crucial clue for this fault diagnosis.

[0107] Furthermore, the platform uses pre-defined feature mapping rules to map each identified target feature to one or more starting nodes in a pre-defined fault knowledge graph. For example, the platform internally pre-defines a fault knowledge graph containing causal relationship nodes between various UPS fault phenomena, intermediate events, and root causes. Simultaneously, the platform also pre-defines mapping rules. For the target features (sensor features, log features) identified in the previous step, the rules further analyze the specific alarm information most closely associated with that feature. For example, the analysis reveals that the strongest association with the sensor feature is "A-phase input voltage drop," and the strongest association with the log feature is "system switches to battery power mode." According to the mapping rules, the platform maps these two specific events to two existing nodes in the knowledge graph: "Node A: Phenomenon - Abnormal mains input voltage" and "Node B: Action - System switches to bypass / battery mode." These two nodes will serve as the starting nodes for this path reasoning.

[0108] Furthermore, based on the aforementioned fault knowledge graph, the platform uses a graph search algorithm to search for reasoning paths that connect these starting nodes and ultimately point to a root node representing the root cause, starting from the initial nodes determined in the previous step. For example, in the platform's fault knowledge graph, "Node A: Phenomenon - Abnormal Mains Input Voltage" has an edge pointing to it from "Node C: Event - Upstream Power Grid Fluctuation"; simultaneously, "Node C" also has an edge pointing to "Node B: Action - System Switches to Bypass / Battery Mode," because power grid fluctuation is a reasonable explanation for the UPS switching mode. The platform employs a bidirectional graph search algorithm, simultaneously starting from both node A and node B and searching upstream (i.e., in opposite directions of the arrows). The search starting from node A quickly finds its parent node C; at the same time, the search starting from node B also finds its parent node C. The search is complete when the searches in both directions intersect at the node "Node C: Event - Upstream Power Grid Fluctuation." Since node C is defined as a root node with zero in-degree in the knowledge graph (i.e., it is an external root cause), the platform has successfully found a reasoning path.

[0109] Furthermore, if a reasoning path is found, it is determined as the final diagnostic path. For example, based on the search results from the previous step, the platform organizes the found paths and ultimately generates a clear diagnostic path pointing to the cause of the fault: "Upstream power grid fluctuations (root cause) → leading to abnormal mains input voltage (phenomenon) → triggering the system to switch to bypass / battery mode (action)." This diagnostic path not only explains why the attention network in step S203 pays close attention to the correlation between sensors and logs, but also clearly reveals the ins and outs of this UPS alarm event in a logical manner, providing valuable decision support for maintenance personnel.

[0110] S205: Integrate and fuse the fault classification results and diagnostic paths corresponding to the feature vectors to generate a fault diagnosis report.

[0111] The purpose of this step is to finally converge and structure the two information streams obtained in parallel from the previous steps: fault type (classification results) and fault cause (diagnostic path). By generating a comprehensive, logically clear, and human-computer co-readable diagnostic report, it provides maintenance personnel with complete decision support from fault symptoms to root causes, thus achieving a closed loop in the entire diagnostic process.

[0112] Specifically, the platform first inputs the fused feature vector into a pre-defined fault classifier to obtain a quantitative classification result for the current fault. For example, the platform inputs the 192-dimensional fused feature vector generated in step S203 into a pre-defined fault classifier consisting of two fully connected layers and a Softmax activation function. The output layer of this classifier has 50 neurons, each corresponding to one of the 50 predefined common UPS fault types in the database. After calculation, the classifier outputs a vector containing 50 probability values, with the 17th dimension (corresponding to fault code "F017") having the highest probability value of 0.985. Therefore, the platform determines this result as the fault classification result for this instance, namely, the fault type is "F017: External power grid fluctuation or interference," with a confidence score of 98.5%. This result accurately identifies the fault category.

[0113] Furthermore, the platform structurally combines the fault classification results obtained in the previous step with the diagnostic path generated in step S204 and related key evidence to generate the final fault diagnosis report. For example, the platform creates a data object to carry all the conclusive information of this diagnosis. Subsequently, the platform fills all the key information into the preset fields of this data object. First, the classification results from the previous step are filled in: fault_code: F017; fault_description: external power grid fluctuations or interference; confidence: 0.985. Next, the key evidence extracted in step S204 is filled in, namely the most important association in the attention weight matrix: key_evidence: [{source: sensor features, target: log features, attention_weight: 0.7}]. Finally, the complete reasoning path generated in step S204 is filled in: reasoning_path: [upstream power grid fluctuations (root cause), leading to abnormal mains input voltage (phenomenon), triggering the system to switch to bypass / battery mode (action).] Once all information is filled in, the platform serializes this complete data object (e.g., converts it to JSON format) to generate a final, structured fault diagnosis report. This report can be directly pushed to the monitoring terminal of the operations and maintenance personnel and displayed in a clear UI (User Interface), for example: "Diagnosis Conclusion: External power grid fluctuations or interference (confidence 98.5%). Diagnosis Path: Upstream power grid fluctuations were detected, causing abnormal mains input voltage, which in turn triggered the system to switch to battery mode. Key Evidence: The abnormal sensor voltage is highly correlated with the log records of the system switching to battery mode, with a correlation weight of 0.7."

[0114] Figure 3This is a schematic diagram of a multi-source information uninterruptible power supply fault diagnosis system according to an embodiment of this application. This system can be implemented through software, hardware, or a combination of both, becoming all or part of the overall system. For example... Figure 3 As shown, the system includes a data acquisition module 301, a data processing module 302, and a fault diagnosis module 303, wherein:

[0115] The data acquisition module 301 is configured to acquire the operating data of multiple modes of the uninterruptible power supply, and convert the operating data of multiple modes into a target mode feature vector group through multiple preset feature extraction models.

[0116] The data processing module 302 is configured to analyze the running data through a causal discovery algorithm, construct a causal relationship diagram between the variables of the running data, and adjust the target modal feature vector group according to the causal relationship diagram to obtain an optimized modal feature vector group.

[0117] The data processing module 302 is also configured to input the optimized modal feature vector group into a preset cross-modal attention network to obtain the fused feature vector and the attention weight matrix;

[0118] The fault diagnosis module 303 is configured to perform path reasoning on target features with weight values ​​greater than a preset threshold in the attention weight matrix through a preset fault knowledge graph, and generate a diagnostic path pointing to the cause of the fault.

[0119] The fault diagnosis module 303 is also configured to integrate the fault classification results and diagnosis paths corresponding to the fused feature vectors to generate a fault diagnosis report.

[0120] Optionally, the data acquisition module 301 is also configured to:

[0121] Acquire sensor time-series data collected by a multi-type sensor group pre-deployed on an uninterruptible power supply, and process the sensor time-series data through a pre-set convolutional neural network model to generate sensor time-series feature vectors;

[0122] Acquire the operation log text data of the uninterruptible power supply output recording device status and events, and encode the operation log text data through a preset natural language processing model to generate semantic feature vectors;

[0123] Acquire historical fault record data of uninterruptible power supply, construct a fault association graph based on the historical fault record data, and process the fault association graph through a preset graph neural network model to generate fault map feature vectors;

[0124] The sensor time-series feature vector, semantic feature vector, and fault map feature vector are time-aligned to form the target modal feature vector group.

[0125] Optionally, the data processing module 302 is also configured to:

[0126] In a causal relationship graph, the variables in the running data corresponding to nodes with an in-degree of zero are identified as target variables;

[0127] In the target modal feature vector set, locate the modal feature vector to be optimized generated from the running data containing the target variable;

[0128] Using a pre-defined attribution analysis method, the attribution score of each dimension of the modality feature vector to be optimized to the target variable is calculated.

[0129] All dimensions with attribution scores greater than a preset attribution threshold are used as target dimensions;

[0130] For each target dimension, the weights are adjusted using preset weight coefficients to generate an optimized modal feature vector;

[0131] Replace the feature vector of the modality to be optimized with the feature vector of the optimized modality to obtain the set of optimized modal feature vectors.

[0132] Optionally, the data processing module 302 is also configured to:

[0133] Matrix multiplication is performed on the target feature vector in the optimized modal feature vector group through preset query linear projection layer, key linear projection layer and value linear projection layer respectively, to generate corresponding query vector, key vector and value vector for the target feature vector. The target feature vector is any feature vector in the optimized modal feature vector group.

[0134] By using a cross-modal attention network to exchange information between the query vector and the key vector, the correlation importance between any two different modalities can be calculated. Any two different modalities represent any two feature vectors in the optimized modal feature vector group.

[0135] Based on the importance of association, the value vectors of each optimized modality feature vector are weighted and fused to generate a fused feature vector;

[0136] An attention weight matrix is ​​constructed based on the importance of association.

[0137] Optionally, the data processing module 302 is also configured to:

[0138] The target feature vector is multiplied by the query weight matrix of the query linear projection layer to generate the query vector.

[0139] The target feature vector is multiplied by the key weight matrix of the key linear projection layer to generate the key vector.

[0140] The target feature vector is multiplied by the value weight matrix of the linear projection layer to generate a value vector.

[0141] Optionally, the data processing module 302 is also configured to:

[0142] For any two optimized modal feature vectors of different modalities, the query vector of the first modality and the key vector of the second modality are calculated by a preset similarity function through the cross-attention mechanism of the cross-modal attention network to obtain the initial attention score. Here, any two different modalities include the first modality and the second modality.

[0143] The initial attention scores are normalized to obtain attention weights, which are then used as the correlation importance.

[0144] Optionally, the fault diagnosis module 303 is also configured to:

[0145] By using preset feature mapping rules, each target feature is mapped to a starting node in the fault knowledge graph;

[0146] Based on the fault knowledge graph, a graph search algorithm is used to search for a reasoning path that connects the starting node and points to the root node representing the cause of the fault, starting from the starting node, and then uses the reasoning path as the diagnostic path.

[0147] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0148] This embodiment also discloses an electronic device, as shown in the reference. Figure 4 The electronic device may include: at least one processor 401, at least one communication bus 402, user interface 403, network interface 404, and at least one memory 405.

[0149] The communication bus 402 is used to enable communication between these components.

[0150] The user interface 403 may include a display screen and a camera. Optionally, the user interface 403 may also include a standard wired interface and a wireless interface.

[0151] The network interface 404 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0152] The processor 401 may include one or more processing cores. The processor 401 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 405, and by calling data stored in memory 405. Optionally, the processor 401 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 401 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 401.

[0153] The memory 405 may include random access memory (RAM) or read-only memory. Optionally, the memory 405 may include a non-transitory computer-readable storage medium. The memory 405 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 405 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 405 may also be at least one storage device located remotely from the aforementioned processor 401. Figure 4 As shown, the memory 405, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a multi-source information uninterruptible power supply fault diagnosis method.

[0154] exist Figure 4In the electronic device shown, the user interface 403 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 401 can be used to call the application program that stores a multi-source information uninterruptible power supply fault diagnosis method in the memory 405. When executed by one or more processors 401, the electronic device performs one or more methods as described in the above embodiments.

[0155] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0156] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0157] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0158] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0160] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 405 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory 405 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0161] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the disclosure in this specification. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope of this application is defined by the claims.

Claims

1. A method for diagnosing uninterruptible power supply faults using multi-source information, characterized in that, The method, applied to an uninterruptible power supply (UPS) fault diagnosis platform, includes: The system acquires operational data of multiple modes of an uninterruptible power supply (UPS), and converts the operational data of the multiple modes into a target mode feature vector group through multiple preset feature extraction models. The operational data is analyzed using a causal discovery algorithm to construct a causal relationship diagram among the variables of the operational data. Based on the causal relationship diagram, the target modal feature vector group is adjusted to obtain an optimized modal feature vector group. The optimized modal feature vector set is input into a preset cross-modal attention network to obtain a fused feature vector and an attention weight matrix; By using a pre-defined fault knowledge graph, path reasoning is performed on target features with weight values ​​greater than a pre-defined threshold within the attention weight matrix to generate a diagnostic path pointing to the cause of the fault. By integrating the fault classification results corresponding to the fused feature vectors and the diagnostic path, a fault diagnosis report is generated.

2. The method according to claim 1, characterized in that, The process of acquiring operational data of multiple modes of the uninterruptible power supply and converting the operational data of multiple modes into a target mode feature vector group through multiple preset feature extraction models specifically includes: Acquire sensor time-series data collected by a multi-type sensor group pre-deployed in the uninterruptible power supply, and process the sensor time-series data through a preset convolutional neural network model to generate sensor time-series feature vectors; The operation log text data of the uninterruptible power supply output recording device status and events is obtained, and the operation log text data is encoded through a preset natural language processing model to generate a semantic feature vector. The historical fault record data of the uninterruptible power supply is obtained, a fault association graph is constructed based on the historical fault record data, and the fault association graph is processed by a preset graph neural network model to generate a fault map feature vector. The sensor time-series feature vector, the semantic feature vector, and the fault map feature vector are time-aligned to form the target modal feature vector group.

3. The method according to claim 2, characterized in that, The step involves analyzing the operational data using a causal discovery algorithm to construct a causal relationship graph among the variables of the operational data, and adjusting the target modal feature vector set based on the causal relationship graph to obtain an optimized modal feature vector set. Specifically, this includes: In the causal relationship graph, the variables in the running data corresponding to the nodes with an in-degree of zero are identified as target variables; In the target modal feature vector group, locate the modal feature vector to be optimized generated from the running data containing the target variable; Using a pre-defined attribution analysis method, the attribution score of each dimension of the modality feature vector to be optimized to the target variable is calculated. All dimensions whose attribution scores are greater than a preset attribution threshold are taken as target dimensions; For each of the target dimensions, the weights are adjusted using preset weight coefficients to generate an optimized modal feature vector; The feature vector of the modality to be optimized is replaced with the feature vector of the optimized modality to obtain the set of optimized modality feature vectors.

4. The method according to claim 3, characterized in that, The step of inputting the optimized modal feature vector group into a preset cross-modal attention network to obtain a fused feature vector and an attention weight matrix specifically includes: Matrix multiplication is performed on the target feature vector in the optimized modality feature vector group through preset query linear projection layer, key linear projection layer and value linear projection layer respectively, to generate corresponding query vector, key vector and value vector for the target feature vector, wherein the target feature vector is any feature vector in the optimized modality feature vector group; The cross-modal attention network interacts with the query vector and the key vector to calculate the correlation importance between any two different modalities, where any two different modalities represent any two feature vectors in the optimized modal feature vector group. Based on the aforementioned correlation importance, the value vectors of each optimized modality feature vector are weighted and fused to generate the fused feature vector; Based on the aforementioned importance of association, the attention weight matrix is ​​constructed.

5. The method according to claim 4, characterized in that, The step of performing matrix multiplication on the target feature vectors in the optimized modality feature vector group through preset query linear projection layers, key linear projection layers, and value linear projection layers to generate corresponding query vectors, key vectors, and value vectors for the target feature vectors specifically includes: The target feature vector is multiplied by the query weight matrix of the query linear projection layer to generate the query vector. The target feature vector is multiplied by the key weight matrix of the key linear projection layer to generate the key vector. The target feature vector is multiplied by the value weight matrix of the linear projection layer to generate the value vector.

6. The method according to claim 4, characterized in that, The step of using the cross-modal attention network to interact with the query vector and the key vector to calculate the association importance between any two different modalities specifically includes: For any two different modalities of the optimized modal feature vector, the query vector of the first modality and the key vector of the second modality are calculated using a preset similarity function through the cross-attention mechanism of the cross-modal attention network to obtain an initial attention score, wherein the two different modalities include the first modality and the second modality; The initial attention score is normalized to obtain the attention weight, and the attention weight is used as the association importance.

7. The method according to claim 1, characterized in that, The step involves using a preset fault knowledge graph to perform path reasoning on target features with weight values ​​greater than a preset threshold within the attention weight matrix, generating a diagnostic path pointing to the cause of the fault. Specifically, this includes: By using preset feature mapping rules, each target feature is mapped to a starting node in the fault knowledge graph; Based on the fault knowledge graph, a graph search algorithm is used to search for a reasoning path that connects the starting node and points to the root node representing the cause of the fault, starting from the starting node, and the reasoning path is used as the diagnostic path.

8. A multi-source information uninterruptible power supply fault diagnosis system, characterized in that, The system includes a data acquisition module, a data processing module, and a fault diagnosis module, among which: The data acquisition module is configured to acquire operating data of multiple modes of uninterruptible power supply, and convert the operating data of multiple modes into a target mode feature vector group through multiple preset feature extraction models. The data processing module is configured to analyze the running data using a causal discovery algorithm, construct a causal relationship diagram between the variables of the running data, and adjust the target modal feature vector group based on the causal relationship diagram to obtain an optimized modal feature vector group. The data processing module is also configured to input the optimized modal feature vector group into a preset cross-modal attention network to obtain a fused feature vector and an attention weight matrix; The fault diagnosis module is configured to perform path reasoning on target features with weight values ​​greater than a preset threshold in the attention weight matrix using a preset fault knowledge graph, and generate a diagnostic path pointing to the cause of the fault. The fault diagnosis module is also configured to integrate the fault classification results corresponding to the fused feature vector and the diagnosis path to generate a fault diagnosis report.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images