Cross-device fault diagnosis method and system and electronic device

Abstract

The invention provides a cross-device fault diagnosis method and system, and an electronic device. The method comprises the steps of collecting a health state signal of operation of a target device; according to the health state signal and a preset hierarchical decoupling network HDN model, a local health prototype and an entity offset vector of the target device are calculated, and the entity offset vector is used for representing the offset between the local health prototype and a preset global health prototype; obtaining a to-be-detected signal of operation of the target equipment, and inputting the to-be-detected signal to the HDN model to obtain a to-be-detected feature vector; and determining a fault diagnosis result of the to-be-detected feature vector based on the entity offset vector. According to the method, cross-equipment fault diagnosis can be quickly and accurately realized.

Description

Technical Field

[0001] This invention belongs to the field of equipment fault diagnosis technology, and specifically relates to a method, system and electronic equipment for diagnosing cross-equipment faults. Background Technology

[0002] The safe and stable operation of power transformers is crucial to the power grid. Analyzing vibration signals from the surface of their enclosures to diagnose internal mechanical conditions (such as loose windings or core faults) is an effective non-invasive monitoring method. Among related technologies, deep learning-based methods have demonstrated high accuracy in processing vibration data from specific transformer models.

[0003] However, the relevant technologies face a fundamental generalization challenge: a model trained on transformer type A (source domain) will exhibit a sharp performance decline when directly applied to transformer type B (target domain) due to domain shift. The physical reason lies in the fact that the monitored vibration signals contain highly device-specific "baseline features" determined by factors such as equipment structure, material properties, and operating conditions. This generalization challenge also exists in industrial scenarios involving monitoring discharge signals, gas signals, and sound signals from other equipment. These device differences interfere with the extraction of fault features, resulting in the model's learned features lacking cross-transfer capabilities. Furthermore, re-collecting large amounts of labeled data covering all fault types for each new equipment model is costly and impractical in industrial practice, severely hindering the large-scale application of intelligent diagnostic technologies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing a method, system and electronic device for diagnosing cross-device faults, which can quickly and accurately diagnose cross-device faults.

[0005] In a first aspect, the present invention provides a method for diagnosing cross-device faults, comprising: acquiring health status signals of the target device; calculating the local health prototype and entity offset vector of the target device based on the health status signals and a preset hierarchical disentangled network (HDN) model, wherein the entity offset vector is used to characterize the offset between the local health prototype and a preset global health prototype; acquiring the test signal of the target device and inputting it into the HDN model to obtain the test feature vector; and determining the fault diagnosis result of the test feature vector based on the entity offset vector.

[0006] In some embodiments, determining the fault diagnosis result of the feature vector under test based on the entity offset vector specifically includes: calibrating the feature vector under test based on the entity offset vector to determine the fault diagnosis result of the feature vector under test, or calibrating a preset global fault prototype based on the entity offset vector to determine the fault diagnosis result of the feature vector under test.

[0007] In some embodiments, the test feature vector is calibrated based on the entity offset vector to determine the fault diagnosis result of the test feature vector. Specifically, this includes: calibrating the test feature vector based on the entity offset vector; calculating the probability distribution of the calibrated test feature vector belonging to the global health prototype and the global fault prototype, respectively; and determining the fault diagnosis result of the test feature vector based on the probability distribution.

[0008] In some embodiments, calibrating a preset global fault prototype based on the entity offset vector to determine the fault diagnosis result of the feature vector to be tested specifically includes: calibrating a preset global fault prototype based on the entity offset vector to generate a local virtual fault prototype of the target device; calculating the probability distribution of the feature vector to be tested belonging to the local healthy prototype and the local virtual fault prototype, respectively; and determining the fault diagnosis result of the feature vector to be tested based on the probability distribution.

[0009] In some embodiments, before calculating the local health prototype and entity offset vector of the target device based on the health status signal and the preset hierarchical decoupling network (HDN) model, the diagnostic method further includes: constructing a preset hierarchical decoupling network (HDN) model, wherein the HDN model includes a feature extraction backbone network, a fault branch, a device / operating condition branch, and a reconstruction decoder, the fault branch is used to extract general fault features, and the device / operating condition branch is used to extract device-specific baseline features.

[0010] In some embodiments, constructing a predefined hierarchical decoupling network (HDN) model specifically includes: acquiring operating signals of at least two different source devices under several states; employing a self-supervised learning method to pre-train a feature extraction backbone network and a reconstruction decoder based on the operating signals to learn a general feature representation; adding parallel fault branches and device / operating condition branches at the output of the pre-trained feature extraction backbone network and the input of the reconstruction decoder to construct the HDN model; training the HDN model using a meta-learning paradigm to decouple general fault features and device-specific baseline features from the general feature representation; and freezing the weights of the HDN model in response to HDN model convergence to obtain the predefined hierarchical decoupling network (HDN) model.

[0011] In some embodiments, a meta-learning paradigm is used to train the HDN model, specifically including: dividing the running signal into several meta-tasks, wherein each meta-task includes a support set and a query set; for each meta-task, training the HDN model based on the support set and calculating a composite loss function based on the query set to optimize the HDN model, wherein the composite loss function includes prototype network loss, orthogonal loss and reconstruction loss.

[0012] In some embodiments, the diagnostic method further includes: calculating a global health prototype and a global fault prototype based on a preset HDN model and the operating signals of at least two different source devices in several states; and storing the global health prototype and the global fault prototype in a global knowledge base.

[0013] In some embodiments, the fault diagnosis result includes an unknown anomaly. After storing the global health prototype and the global fault prototype in the global knowledge base, the diagnosis method further includes: if the fault diagnosis result is an unknown anomaly; determining the fault type label to which the unknown anomaly belongs; collecting the fault status signal corresponding to the fault type label; calculating a new global fault prototype based on the preset HDN model and the fault status signal; and adding the new global fault prototype to the global knowledge base.

[0014] Secondly, the present invention also provides a cross-device fault diagnosis system, comprising: a data acquisition module for acquiring health status signals of a target device; an adaptive calibration module connected to the data acquisition module for calculating a local health prototype and an entity offset vector of the target device based on the health status signals and a preset hierarchical decoupling network (HDN) model, wherein the entity offset vector is used to characterize the offset between the local health prototype and a preset global health prototype; and a diagnosis module connected to the adaptive calibration module for acquiring a test signal of the target device and inputting it into the HDN model to obtain a test feature vector, and a fault diagnosis result for determining the test feature vector based on the entity offset vector.

[0015] Thirdly, the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores one or more computer programs executable by the at least one processor, the one or more computer programs being executed by the at least one processor to enable the at least one processor to perform a cross-device fault diagnosis method as described in the first aspect.

[0016] The cross-device fault diagnosis method, system, and electronic equipment provided by this invention, due to the feature decoupling capability of the Hierarchical Decoupling Network (HDN) model, can be applied to entirely new devices that have never been seen before, fundamentally solving the cross-device generalization problem. Furthermore, when the target device is connected, only a small number of its operational health status signals need to be collected, eliminating the need to collect a large number of fault signals covering all fault types, significantly reducing the cost, time, and data dependency of deploying the new model. By combining entity offset vectors for calibration to determine the fault diagnosis result, the systematic offset caused by personalized factors such as device structure and materials on signal characteristics can be effectively offset, ensuring high accuracy and robustness of fault diagnosis on new devices, thereby achieving fast and accurate cross-device fault diagnosis.

[0017] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. The above and other features and advantages will become more apparent to those skilled in the art from the detailed description of exemplary embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 A flowchart illustrating a cross-device fault diagnosis method provided in an embodiment of the present invention;

[0020] Figure 2 A flowchart illustrating another method for diagnosing cross-device faults provided in an embodiment of the present invention;

[0021] Figure 3 A flowchart illustrating a method for constructing a hierarchical decoupling network (HDN) model according to an embodiment of the present invention;

[0022] Figure 4 A schematic diagram illustrating entity offset vector-based calibration as provided in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram illustrating the evolution of a global knowledge base, as provided in an embodiment of the present invention.

[0024] Figure 6 A block diagram of a cross-device fault diagnosis system provided in an embodiment of the present invention;

[0025] Figure 7 This is a block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] Where there is no conflict, the various embodiments of the present invention and the features thereof may be combined with each other.

[0028] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0030] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the meaning consistent with their meaning in the context of the relevant art and the invention, and will not be interpreted as having an idealized or overly formal meaning unless expressly so defined herein.

[0031] Firstly, this embodiment provides a cross-device fault diagnosis method. This method can be applied to monitor and diagnose the operation of physical equipment in industrial scenarios. The operating signals can be vibration signals, discharge signals, gas signals, sound signals, etc. It can also be applied to monitor and diagnose the performance of virtualized environments (such as virtual machines, containers, etc.). The operating signals can be the performance indicators of the virtual machine (such as CPU utilization). The diagnosis method is executed by electronic devices such as terminal devices or servers. The terminal device can be user equipment (UE), mobile devices, computing devices, vehicle-mounted devices, etc. The method can be implemented by a processor calling computer-readable program instructions stored in memory. Alternatively, the method can be executed by a server.

[0032] Reference Figure 1 This embodiment provides a method for diagnosing cross-device faults, including:

[0033] Step 101: Collect the health status signal of the target device.

[0034] Step 102: Based on the health status signal and the preset hierarchical decoupling network (HDN) model, calculate the local health prototype and entity offset vector of the target device. The entity offset vector is used to represent the offset between the local health prototype and the preset global health prototype.

[0035] Step 103: Obtain the signal to be tested from the target device and input it into the HDN model to obtain the feature vector to be tested.

[0036] Step 104: Determine the fault diagnosis result of the feature vector to be tested based on the entity offset vector.

[0037] In this context, "cross-device" refers to transfer learning, the process from a source domain device to a target domain device. A device can be a physical device or a virtual device. The target domain refers to a specific device or set of devices to be monitored for health status or diagnosed for faults. The source domain refers to a set of devices with known health / fault states, serving as a knowledge source. The target device refers to a device within the target domain, such as a new type of transformer being connected to the power grid for the first time. The source device refers to a device within the source domain. Upon initial connection of the target device, a small number of health status signals are collected, such as signals less than a preset threshold. A local health prototype of the target device is calculated based on the health status signals and a preset HDN model. Then, an entity offset vector is calculated based on the offset (e.g., difference) between the local health prototype and the preset global health prototype. The entity offset vector is used for adaptive calibration. Based on the entity offset vector, inherent individual differences between the target device and the source device can be eliminated, mapping the target device's test feature vector to a trained global feature space, or mapping global features to the target device's localized feature space, thereby significantly improving the accuracy and reliability of fault diagnosis for the target device (or new device). A global health prototype refers to a baseline health prototype constructed by extracting and aggregating the common features of signals from multiple source devices operating in a healthy state using an HDN model. The global health prototype is pre-set in a global knowledge base.

[0038] In some embodiments, calculating the local health prototype of the target device based on the health status signal and a preset HDN model includes: inputting the health status signal into the HDN model and calculating the characteristic mean of the health status signal as the local health prototype of the target device.

[0039] In this embodiment, the Hierarchical Decoupling Network (HDN) model possesses feature decoupling capabilities, enabling its application to entirely new, unseen devices and fundamentally solving the cross-device generalization problem. Other neural network models with feature decoupling capabilities can also be used with the HDN model. Furthermore, when the target device is connected, only a small amount of its operational health status signals need to be collected, eliminating the need to collect a large number of fault signals covering all fault types, significantly reducing the cost, time, and data dependency of deploying the new model. By combining the entity offset vector with calibration to determine the fault diagnosis result, the systematic offset caused by personalized factors such as device structure and materials on signal characteristics can be effectively offset, ensuring high accuracy and robustness of fault diagnosis on new devices, thereby achieving fast and accurate cross-device fault diagnosis.

[0040] In some embodiments, the fault diagnosis result based on the entity offset vector to determine the feature vector to be measured includes:

[0041] The test feature vector is calibrated based on the entity offset vector to determine the fault diagnosis result of the test feature vector, or a preset global fault prototype is calibrated based on the entity offset vector to determine the fault diagnosis result of the test feature vector.

[0042] There are two adaptive calibration methods: one is to calibrate the target feature vector based on the entity offset vector, i.e., the calibration data; the other is to calibrate the global fault prototype based on the entity offset vector, i.e., the calibration prototype. The global fault prototype refers to a baseline fault prototype constructed by extracting and aggregating the common features of signals from multiple source devices operating in fault states through the HDN model. The global fault prototype is pre-set in the global knowledge base. There can be one or more global fault prototypes.

[0043] In some embodiments, calibrating the feature vector to be tested based on the entity offset vector to determine the fault diagnosis result of the feature vector to be tested includes: calibrating the feature vector to be tested based on the entity offset vector; calculating the probability distribution of the calibrated feature vector to be tested belonging to the global health prototype and the global fault prototype, respectively; and determining the fault diagnosis result of the feature vector to be tested based on the probability distribution.

[0044] In some embodiments, calibrating the feature vector to be tested based on the entity offset vector includes: translating and calibrating the feature vector to be tested according to the entity offset vector to obtain the calibrated feature vector to be tested.

[0045] In some embodiments, calculating the probability distribution of the calibrated test feature vector belonging to the global health prototype and the global fault prototype respectively includes: calculating the Euclidean distance between the calibrated test feature vector and the global health prototype and the global fault prototype respectively to obtain the corresponding score vector; inputting each score vector into a probability distribution transformation function (such as the Softmax function) to obtain the probability distribution corresponding to the calibrated test feature vector belonging to the global health prototype and the global fault prototype.

[0046] In some embodiments, determining the fault diagnosis result of the tested feature vector based on the probability distribution includes: sorting the probability distributions corresponding to the global health prototype and the global fault prototype in descending order; determining the magnitude of the maximum value of the probability distribution relative to a confidence threshold; and, in response to the maximum value of the probability distribution being greater than the confidence threshold, determining the fault diagnosis result of the tested feature vector as the global prototype corresponding to the maximum value of the probability distribution, where the global prototype includes both the global health prototype and the global fault prototype. It should be noted that the fault diagnosis result also includes a confidence level. In response to the maximum value of the probability distribution being less than or equal to the confidence threshold, determining the fault diagnosis result of the tested feature vector as an unknown anomaly.

[0047] In some embodiments, calibrating a preset global fault prototype based on an entity offset vector to determine the fault diagnosis result of the feature vector to be tested includes: calibrating a preset global fault prototype based on an entity offset vector to generate a local virtual fault prototype of the target device; calculating the probability distribution of the feature vector to be tested belonging to the local health prototype and the local virtual fault prototype, respectively; and determining the fault diagnosis result of the feature vector to be tested based on the probability distribution.

[0048] The calibration of the preset global fault prototype based on the entity offset vector includes: calibrating the preset global fault prototype by translation based on the entity offset vector to generate a set of local virtual fault prototypes for the target device. The Euclidean distance between the feature vector to be tested and the local healthy prototype and the local virtual fault prototype is calculated to obtain the corresponding score vectors. Each score vector is input into a probability distribution transformation function (such as the Softmax function) to obtain the probability distributions corresponding to the global healthy prototype and the global fault prototype for the feature vector to be tested. The probability distributions corresponding to the global healthy prototype and the global fault prototype are sorted in descending order. The magnitude of the maximum value of the probability distribution and the confidence threshold are compared. If the maximum value of the probability distribution is greater than the confidence threshold, the fault diagnosis result of the feature vector to be tested is determined to be the global prototype corresponding to the maximum value of the probability distribution. The global prototype includes both the global healthy prototype and the global fault prototype. It should be noted that the fault diagnosis result also includes the confidence level. If the maximum value of the probability distribution is less than or equal to the confidence threshold, the fault diagnosis result of the feature vector to be tested is determined to be an unknown anomaly.

[0049] In this embodiment, by introducing a decision-making mechanism based on confidence threshold, the robustness and security of the diagnostic method are greatly improved. It can effectively avoid misjudgment of low-confidence samples or unknown fault modes and prevent false alarms and missed alarms. In addition, since the fault diagnosis results provide confidence, they provide transparent and quantifiable decision-making basis, which enhances the interpretability and accuracy of the results.

[0050] In some embodiments, before calculating the local health prototype and entity offset vector of the target device based on the health status signal and the preset hierarchical decoupling network (HDN) model, the diagnostic method further includes: constructing a preset hierarchical decoupling network (HDN) model, wherein the HDN model includes a feature extraction backbone network, a fault branch, a device / operating condition branch, and a reconstruction decoder, the fault branch is used to extract general fault features, and the device / operating condition branch is used to extract device-specific baseline features.

[0051] In this embodiment, self-supervised learning methods (such as contrastive learning, generative methods, and non-contrastive learning) can be used to pre-train the feature extraction backbone network, enabling it to learn robust, surface-variation-insensitive low-level feature representations from unlabeled signals. Then, based on the pre-trained feature extraction backbone network, an HDN model is constructed, comprising the feature extraction backbone network, a fault branch, a device / condition branch, and a reconstruction decoder. The HDN model is trained using a meta-learning paradigm, enabling it to decouple general fault features from signals. Because the fault branch and the device / condition branch decouple general fault features from device-specific baseline features, it can be applied to entirely new, unseen devices, fundamentally solving the generalization problem of fault diagnosis across devices.

[0052] In some embodiments, constructing a predefined hierarchical decoupling network (HDN) model includes:

[0053] Step 201: Obtain the operating signals of at least two different source devices in several states.

[0054] Step 202: A self-supervised learning method is adopted to pre-train the feature extraction backbone network and the reconstruction decoder based on the running signal in order to learn a general feature representation.

[0055] Step 203: Add parallel fault branches and equipment / condition branches to the output of the pre-trained feature extraction backbone network and the input of the reconstruction decoder to construct the HDN model.

[0056] Step 204: The HDN model is trained using a meta-learning paradigm to decouple general fault features and device-specific baseline features from the general feature representation.

[0057] Step 205: In response to the convergence of the HDN model, freeze the weights of the HDN model to obtain the preset hierarchical decoupling network HDN model.

[0058] Examples of the states include a healthy state, a first fault type state, a second fault type state, and a third fault type state. Operating signals from at least two different source devices under these states can be acquired via sensors and data acquisition cards. The process of building a pre-defined HDN model can be performed offline.

[0059] In this embodiment, a hierarchical decoupling network (HDN) capable of decoupling general fault features and personalized equipment features (i.e., equipment-specific baseline features) is trained by using self-supervised pre-training and meta-learning paradigm. The entity offset vector is calculated by combining a small number of health status signals (which can be understood as a small number of health samples) collected in the online stage to adaptively calibrate the global prototype, thereby achieving rapid and high-precision fault diagnosis of the target equipment.

[0060] In some embodiments, after step 201 and before step 202, the diagnostic method further includes data preprocessing of operating signals from at least two different source devices in several states. Data preprocessing includes time alignment, signal slicing, and normalization.

[0061] In some embodiments, training the HDN model using a meta-learning paradigm includes:

[0062] Step 2041: Divide the running signal into several meta-tasks, where each meta-task includes a support set and a query set.

[0063] Step 2042: For each meta-task, train the HDN model based on the support set and calculate the composite loss function based on the query set to optimize the HDN model. The composite loss function includes the prototype network loss, orthogonal loss, and reconstruction loss.

[0064] In some embodiments, the diagnostic method further includes: calculating a global health prototype and a global fault prototype based on a preset HDN model and the operating signals of at least two different source devices in several states; and storing the global health prototype and the global fault prototype in a global knowledge base.

[0065] In this embodiment, the operating signals under several states are grouped according to their real labels. For example, the operating signals are grouped according to healthy state, first fault type state, second fault type state, and third fault type state. The grouped operating signals are then used as samples corresponding to the real labels and input into a preset HDN model. After extracting fault features, the mean is calculated to obtain a preset global prototype. Each global healthy prototype and each global fault prototype corresponds one-to-one with a real label.

[0066] In some embodiments, the fault diagnosis result includes unknown anomalies. After storing the global health prototype and global fault prototype in a global knowledge base, the diagnostic method further includes:

[0067] Step 301: If the fault diagnosis result is an unknown anomaly, determine the fault type label to which the unknown anomaly belongs.

[0068] Step 302: Collect the fault status signal corresponding to the fault type label.

[0069] Step 303: Calculate a new global fault prototype based on the preset HDN model and fault state signals.

[0070] Step 304: Add the new global fault prototype to the global knowledge base.

[0071] The process involves determining the fault type label to which an unknown anomaly belongs based on human-machine collaboration or an expert system. Step 303 includes: inputting the fault state signal corresponding to the fault type label into a preset HDN model, calculating the fault feature vector and averaging it to obtain a new global fault prototype. Step 304 includes: associating the new global fault prototype with the corresponding fault type label and adding it to the global knowledge base. The purpose of step 304 is to supplement the global knowledge base data, thereby realizing the evolution of the global knowledge base and expanding diagnostic capabilities.

[0072] In this embodiment, by expanding the global knowledge base data—that is, once a new global prototype is added to the global knowledge base—the capabilities of the entire diagnostic method are instantly expanded. From the next moment on, when any adapted device experiences the same type of fault, it can be correctly identified through the entity offset vector calibration mechanism, realizing closed-loop learning and capability evolution from the unknown to the known. In summary, due to the human-machine collaborative knowledge evolution closed loop that includes handling unknown anomalies, this invention has excellent generalization ability and can be effectively applied to unseen devices; it has complete scalability and can continuously evolve through lifelong learning, thus fundamentally solving the technical problems of poor generalization ability and inability to handle unknown faults in traditional intelligent diagnostic models, and has important theoretical significance and application value.

[0073] In some embodiments, since a robust diagnostic method must be able to distinguish between “known normal / fault states” and “true unknown events” to avoid misclassifying or ignoring newly emerging fault types, and needs to consider the actual needs of knowledge growth, the ways to determine the fault diagnosis result of the feature vector to be tested as an unknown anomaly include one or a combination of the following.

[0074] The first method, determining the fault diagnosis result of the feature vector to be tested as an unknown anomaly, includes: when the maximum value of the probability distribution is less than or equal to the confidence threshold, determining the fault diagnosis result of the feature vector to be tested as an unknown anomaly. Here, the maximum value of the probability distribution being less than or equal to the confidence threshold indicates that there is insufficient information to classify the sample into any known type.

[0075] The second method, determining the fault diagnosis result of the feature vector to be tested as an unknown anomaly, includes: calculating the minimum distance between the feature vector to be tested and the global healthy prototype and the global faulty prototype, respectively; determining the magnitude of the minimum distance compared to the spatial distance threshold; and determining the fault diagnosis result of the feature vector to be tested as an unknown anomaly if the minimum distance is greater than the spatial distance threshold. Specifically, if the minimum distance between the feature vector to be tested and the global healthy prototype and the global faulty prototype is greater than the spatial distance threshold, it indicates that the feature vector to be tested is in a "no-man's land" in the feature space, far from all known prototype clusters.

[0076] In other words, when the maximum value of the probability distribution is less than or equal to the confidence threshold, and / or the minimum distance is greater than the spatial distance threshold, the fault diagnosis result of the feature vector to be tested is determined to be an unknown anomaly.

[0077] In some embodiments, after determining that the fault diagnosis result of the feature vector to be measured is an unknown anomaly, the diagnostic method further includes generating an alarm signal. Generating an alarm signal facilitates the initiation of human-machine collaboration or provides an alarm for abnormal conditions during equipment operation monitoring.

[0078] The cross-device fault diagnosis method provided in this embodiment, due to the feature decoupling capability of the Hierarchical Decoupled Network (HDN) model, can be applied to entirely new devices that have never been seen before, fundamentally solving the cross-device generalization problem. Other neural network models with feature decoupling capabilities can also be selected using the HDN model. Furthermore, when the target device is connected, only a small number of its operational health status signals need to be collected, eliminating the need to collect a large number of fault signals covering all fault types, significantly reducing the cost, time, and data dependency of deploying new models. The fault diagnosis result is determined after calibration using entity offset vectors, effectively offsetting the systematic offset caused by personalized factors such as device structure and materials on signal characteristics, ensuring high accuracy and robustness of fault diagnosis on new devices, thereby achieving fast and accurate cross-device fault diagnosis. Furthermore, by introducing a decision mechanism based on confidence thresholds, the robustness and security of the diagnostic method are greatly improved, effectively avoiding misjudgments of low-confidence samples or unknown fault modes, preventing false negatives and false negatives; in addition, since the fault diagnosis result provides a confidence level, it provides a transparent and quantifiable decision basis, thereby enhancing the interpretability and accuracy of the diagnostic result. By employing self-supervised pre-training and a meta-learning paradigm to train a hierarchical decoupling network (HDN) capable of decoupling general fault features from personalized equipment features (i.e., equipment-specific baseline features), and combining this with a small amount of health status signals collected during the online phase (which can be understood as a small number of healthy samples), entity offset vectors are calculated to adaptively calibrate the global prototype, thereby achieving rapid and high-precision fault diagnosis of the target equipment. Furthermore, by supplementing the global knowledge base with additional data, the global knowledge base evolves to expand diagnostic capabilities.

[0079] Example 1:

[0080] For example, as a specific form of this embodiment, 110kV and 220kV transformers are used as source devices, and a brand-new, untrained 500kV transformer is used as the target device. Vibration signals are used as operating signals. The entire implementation process of the cross-device fault diagnosis method of this embodiment is described in detail. The cross-device fault diagnosis method is applied to a cross-device fault diagnosis system. Figure 2 As shown, a method for diagnosing cross-device faults includes:

[0081] Offline model training phase:

[0082] Step S1: Acquire operating signals from at least two different source devices under several states and perform data preprocessing, i.e., acquire multi-source vibration signals and preprocess them.

[0083] Step S1.1: Based on the sensing system and data acquisition card, acquire the operating signals of at least two different source devices under several states. At key locations on the tank walls of two 110kV and one 220kV source transformers, install uniaxial ICP (Integrated Circuit Piezoelectric) piezoelectric accelerometers (e.g., model PCB356B21). The measurement points cover the tank walls corresponding to the high, medium, and low voltage windings and the tank bottom corresponding to the iron core, for a total of 8 measurement points and 8 channels. The vibration signals are continuously acquired at a sampling frequency of 51.2kHz using a data acquisition card (e.g., NI cRIO-9068 and NI 9234 modules). The acquired vibration signals cover the following states: (1) Healthy state: Continuous signals of the transformer operating normally for more than 1 month under different load rates (30%-90%). (2) Winding loosening fault: Vibration signals of transformers confirmed to have slight winding loosening after short-circuit impact test. (3) Core failure: Loosening the core fastening bolts simulates the vibration signal of core failure.

[0084] Step S1.2: Data preprocessing. (1) Time alignment: All data are time-stamped with GPS (Global Positioning System) clock synchronization to ensure a unified time reference. (2) Signal slicing: Using a sliding window with a length of 2048 sampling points (corresponding to a duration of 0.4 seconds) and a step size of 512 sampling points, the continuous signal is divided into a large number of overlapping samples. (3) Normalization: For each slice sample, Z-score normalization is performed independently to eliminate the influence of vibration amplitude dimensions under different channels and different operating conditions.

[0085] Step S2: A self-supervised learning method is used to pre-train the feature extraction backbone network and the reconstruction decoder based on vibration signals to learn a general feature representation. This is the self-supervised pre-training of the feature extraction backbone network.

[0086] Step S2.1: The feature extraction backbone network f(·) adopts an architecture combining neural networks (such as one-dimensional convolutional neural networks 1D-CNN, temporal convolutional networks, multilayer perceptrons, and recurrent neural networks) with an attention mechanism. Taking 1D-CNN as an example, its network structure is as follows:

[0087] ① Input layer: Receives 2048-dimensional raw vibration signals;

[0088] ② Convolutional block 1: 32 1×5 convolutional kernels, stride 1, ReLU activation, batch normalization, max pooling (pooling window 2).

[0089] ③ Convolutional block 2: 64 1×3 convolutional kernels, stride 1, ReLU activation, batch normalization, max pooling (pooling window 2).

[0090] ④ Convolutional block 3: 128 1×3 convolutional kernels, stride 1, ReLU activation, batch normalization;

[0091] ⑤ Spatiotemporal Attention Module: Calculates the temporal and spatial attention weights of the feature map;

[0092] ⑥ Global average pooling: Outputs a 256-dimensional shared feature vector h;

[0093] Step S2.2: Set a pre-training objective so that the feature extraction backbone network f(·) can learn the inherent, condition- and noise-insensitive structured patterns in the vibration signal without relying on any fault type labels.

[0094] Step S2.3: Self-supervised pre-training employs a contrastive learning method, using either the SimCLR framework or the MoCo framework. Taking the SimCLR framework as an example, for each input vibration signal sample x, a data augmentation module generates two different augmented views x. i and x j These form positive sample pairs. The data augmentation strategy library is as follows:

[0095] ① Random time-based cropping and scaling: Randomly crop out a subsequence of 70%-100% of the original length, and then scale it back to the original length.

[0096] ② Gaussian white noise: Add Gaussian white noise with a signal-to-noise ratio of 20dB to 30dB.

[0097] ③ Random amplitude jitter: Multiply the overall amplitude of the sample by a random coefficient between 0.8 and 1.2.

[0098] ④ Random Time Reversal: Reverses the signal on the time axis with a 50% probability.

[0099] Step S2.4: Pre-training process setup. Vibration data from all source devices (regardless of health or malfunction) are used as an unlabeled data pool. A temporary projection head g(·) is a two-layer MLP (Multi-layer Perceptron), projecting the 256-dimensional features from the feature extraction backbone network onto a 64-dimensional network. The NT-Xent loss function is used, with the temperature hyperparameter τ set to 0.1. The Adam optimizer is employed, with a learning rate of 5e-4, a batch size of 256, and training for 100 epochs. After pre-training, the weights of the feature extraction backbone network f(·) are saved, and the projection head g(·) is discarded.

[0100] Step S3: Add parallel fault branches and equipment / condition branches to the output of the pre-trained feature extraction backbone network and the input of the reconstruction decoder to construct the HDN model. Then, train the HDN model using a meta-learning paradigm to decouple general fault features and equipment-specific baseline features from the general feature representation. This involves constructing a hierarchical decoupled network HDN model and meta-learning.

[0101] Step S3.1: Construct the HDN model. (Refer to...) Figure 3 Its feature extraction backbone network f(·) loads the pre-trained weights from step S2. The fault branch (F-Branch) g_f(·) and the equipment / condition branch (C-Branch) g_c(·) are both three-layer MLPs, mapping the 256-dimensional shared feature vector h to the 64-dimensional fault feature z. f and equipment / operating condition characteristics z c The decoder network (Decoder) d(·) employs a transposed convolutional structure that is symmetrical to the feature extraction backbone network.

[0102] Step S3.2: Construct Meta-Tasks. Specifically, the vibration data (i.e., operating signals) is divided into several meta-tasks, each including a support set and a query set. For example, a 3-way 5-shot task setup is used (i.e., three types: {health, winding looseness, core fault}, with 5 samples per class in the support set). The query set contains 15 samples per class. In each training iteration, cross-task tasks are constructed with a 70% probability (e.g., the support set is sampled from 110kV data, and the query set is sampled from 220kV data), and intra-domain tasks are constructed with a 30% probability (both the support set and query set come from the same device).

[0103] Step S3.3: Composite Loss and Optimization. Specifically, for each meta-task, the HDN model is trained based on the support set, and a composite loss function is calculated based on the query set to optimize the HDN model. The composite loss function includes the prototype network loss, orthogonality loss, and reconstruction loss. For example, the composite loss function for meta-learning is as follows:

[0104]

[0105] The weighting coefficients were optimized through experiments. For prototype network loss, For orthogonal loss, To reconstruct the loss, an orthogonal loss weight coefficient λ is set. ortho = 1.0, reconstruction loss weight coefficient λ recon = 0.2, L ortho By minimizing z f and z cThe Frobenius norm of the correlation coefficient matrix of the intra-batch feature matrix is ​​implemented. The Adam optimizer is used, with a learning rate of 1e-3 for F-Branch, C-Branch, and Decoder, and a learning rate of 1e-5 for the feature extraction backbone network for fine-tuning. A total of 20,000 meta-tasks were trained iteratively.

[0106] Step S3.4: In response to the convergence of the HDN model, freeze the weights of the HDN model to obtain the preset hierarchical decoupling network HDN model.

[0107] Step S4: Establish a global knowledge base. Specifically, based on the preset HDN model and the operating signals of at least two different source devices in several states, calculate the global health prototype and the global fault prototype respectively, and store the global health prototype and the global fault prototype in the global knowledge base.

[0108] Step S4.1: Calculate the global prototype, which includes a global health prototype and a global fault prototype. Input all 110kV and 220kV training data (separated by category) into the frozen HDN model, and only take its F-Branch output z. f For all z in each category f The mean of the vectors is used to obtain the global prototypes: global healthy centroid (P_healthy_global), global winding looseness centroid (P_fault_global_winding_looseness), and global core looseness centroid (P_fault_global_core_issue).

[0109] Step S4.2: Storage. Serialize and store the frozen HDN model weight file and the knowledge base file (such as JSON or pickle format) containing the above three global prototypes. The detailed process is shown in Algorithm 1 in Table 1.

[0110] Table 1 Offline Model Training

[0111]

[0112]

[0113]

[0114] Online Diagnosis and Adaptation Phase:

[0115] Step S5: Collect the health status signal of the target device, and calculate the local health prototype and entity offset vector of the target device based on the health status signal and the preset hierarchical decoupling network (HDN) model. This involves collecting a small number of health samples and adaptive calibration.

[0116] Step S5.1: Data Acquisition and Mapping. A brand-new, untrained 500kV transformer is connected to the power grid. After grid connection, its healthy operation is confirmed, and 50 vibration signal samples (each sample length 2048) are continuously acquired. These 50 healthy samples are input into the deployed, frozen HDN model to obtain 50 64-dimensional health feature vectors.

[0117] Step S5.2: Calibration Vector Calculation. Calculate the mean of these 50 vectors to obtain the local health prototype P_healthy_local_500kV for the 500kV transformer. Read the preset global health prototype P_healthy_global from the knowledge base and calculate the entity offset vector of the device (the entity offset vector is used to represent the offset between the local health prototype and the preset global health prototype): Offset_Vector_500kV = P_healthy_local_500kV - P_healthy_global.

[0118] Step S5.3: Global Knowledge Base Update. Associate the two vectors P_healthy_local_500kV and Offset_Vector_500kV with the unique asset ID of the 500kV transformer and store them in the dynamic global knowledge base.

[0119] Step S6: Real-time diagnosis of the vibration signal under test. Specifically, the signal under test of the target equipment is acquired and input into the HDN model to obtain the feature vector under test, and the fault diagnosis result of the feature vector under test is determined based on the entity offset vector.

[0120] Step S6.1: Signal acquisition and feature extraction. A real-time vibration signal X to be measured is acquired from the 500kV transformer every minute. new Inputting it into a pre-defined HDN model yields its feature vector z to be tested. new .

[0121] Step S6.2: Localized Prototype Calibration. Specifically, a preset global fault prototype is calibrated based on the entity offset vector to generate a local virtual fault prototype for the target device. See example below. Figure 4 The Offset_Vector_500kV of the 500kV transformer is read from the global knowledge base, and then all global fault prototypes in the global knowledge base are translated and calibrated in real time to generate a set of local virtual fault prototypes.

[0122] P_fault_local_winding_looseness=P_fault_global_winding_looseness+Offset_Vector_500kV;

[0123] P_fault_local_core_issue=P_fault_global_core_issue+Offset_Vector_500kV.

[0124] Step S6.3: Calculate the probability distributions of the tested feature vector belonging to the local healthy prototype and the local virtual fault prototype, respectively, and determine the fault diagnosis result of the tested feature vector based on the probability distributions. The probabilistic decision-making mainly includes three parts:

[0125] ① Construct a set {P_candidates} containing all candidate prototypes, including the device’s local health prototype P_healthy_local_500kV and all local virtual fault prototypes (such as P_fault_local_winding_looseness, P_fault_local_core_issue).

[0126] ② Calculate the eigenvector z to be measured new With each prototype P in the candidate prototype set k The squared Euclidean distance between them. For the k-th prototype, its score is score_k = -||z new - P k ||^2.

[0127] ③ Construct a vector [score_1, score_2, ..., score_K] from all scores, where K is the total number of candidate prototypes.

[0128] ④ Input the score vector into the Softmax function to convert it into a probability distribution vector [prob_1, prob_2, ..., prob_K]. The formula for calculating the Softmax function is:

[0129]

[0130] Where T is a temperature coefficient used to adjust the smoothness of the probability distribution, and in this embodiment, it can be set to 1. The value of j ranges from 1 to K. Each prob_k represents the feature vector z to be measured. new The confidence probability of belonging to the k-th prototype label.

[0131] ⑤ Find the term prob_max with the highest probability value in the probability distribution vector. The prototype label corresponding to this highest probability (such as "healthy", "loose winding", or "core fault") is the final fault diagnosis result. Simultaneously output the diagnosis label and its corresponding confidence probability prob_max. If all probability values ​​are below a preset confidence threshold (e.g., 0.5), the fault diagnosis result is "unknown anomaly, low confidence," and an alarm signal is generated and output.

[0132] Step S7: Unknown anomaly detection and global knowledge base evolution, such as... Figure 5 As shown.

[0133] Step S7.1: Real-time detection and judgment of unknown anomalies.

[0134] A robust diagnostic system must be able to distinguish between "known normal / fault states" and "truly unknown events" to avoid misclassifying or ignoring newly emerging fault types and to provide input for the knowledge growth of the diagnostic system. In the decision-making process of step S6.3, the diagnostic system calculates the target feature vector z... new After assigning the probability distribution [prob_1, prob_2, ..., prob_K] to all candidate prototypes, the following double-determination is performed:

[0135] ① Confidence threshold determination: Calculate the maximum probability value prob_max in the probability distribution. If prob_max is less than a preset confidence threshold τ_prob (e.g., τ_prob = 0.5), it means that the model is not confident enough in classifying the sample into any known category.

[0136] ② Distance threshold determination (for enhancement): Calculate the feature vector z to be tested. new The minimum distance d_min between the sample and all candidate prototypes. If d_min is greater than a preset spatial distance threshold τ_dist, it means that the sample is in a "no man's land" in the feature space, far away from all known pattern clusters.

[0137] When any one or two of the above criteria are met, the diagnostic system will not forcibly output the label with the highest probability, but will instead classify the event as an "unknown anomaly." The diagnostic system will immediately generate a dedicated alarm and automatically record and store the following key information:

[0138] ① Alarm timestamp.

[0139] ② The original vibration signal segment X that triggered the alarm new .

[0140] ③ The feature vector z to be measured newAnd its coordinates in F-Space.

[0141] ④ z new A list of distances to all localized prototypes.

[0142] ⑤ z new The complete probability distribution belonging to all categories.

[0143] ⑥ Relevant operating conditions at the time of the event (such as load, temperature, etc.).

[0144] Step S7.2: Expert intervention and confirmation of fault mechanism.

[0145] "Unknown anomaly" alarms are pushed to professional equipment maintenance or diagnostic experts via the monitoring system. These experts utilize the detailed information recorded by the system, combined with their own professional knowledge, to conduct an in-depth analysis of the event. The experts are then required to investigate X. new Perform spectral analysis, time-frequency analysis, or cross-compare with other monitoring methods (such as oil chromatography) to determine the physical root cause of the anomaly.

[0146] Step S7.3: Create a new global fault prototype and evolve the global knowledge base.

[0147] After expert analysis, it was confirmed that the "unknown anomaly" was a new type of fault with clear physical meaning and potential for recurrence, such as "partial discharge flashover of the bushing." Experts provided a standardized new label for this fault type and began the following steps:

[0148] ① New Sample Collection: Maintenance experts will continuously monitor the equipment. During subsequent operation, if the same type of anomaly reappears and is confirmed by experts, vibration signal samples will be collected and stored. Once a certain number (e.g., 5-10) of new fault samples of the same type are collected, {X} new_fault After that, the global knowledge base update process is initiated.

[0149] ② New global fault prototype calculation: The collected {X} new_fault The sample set is input into the deployed, frozen HDN model (i.e., the pre-defined HDN model), and the corresponding fault feature vectors {z} are calculated. f_new_fault Then, the mean of these new fault feature vectors is calculated to create a completely new global fault prototype: P_fault_global_new_type = mean({z f_new_fault}).

[0150] ③ Global knowledge base addition: Associate this new global fault prototype P_fault_global_new_type with the new label provided by the expert (such as "bucket flashover") and add it as a new entry to the global knowledge base.

[0151] Step S7.4: Immediate expansion of diagnostic system capabilities.

[0152] Once a new global fault prototype is added to the global knowledge base, the capabilities of the entire diagnostic system are instantly expanded. From that moment on, when any adapted equipment (whether 110kV, 220kV, or 500kV) experiences a similar "buffer flashover" fault, the diagnostic system can correctly identify it through the entity offset vector calibration mechanism, achieving closed-loop learning and capability evolution from the unknown to the known. The specific algorithm process is shown in Table 2:

[0153] Table 2 Online Diagnosis and Adaptive Update

[0154]

[0155]

[0156]

[0157] It is understood that the various method embodiments mentioned above in this invention can be combined with each other to form combined embodiments without violating the underlying principles and logic. Due to space limitations, these will not be elaborated upon further. Those skilled in the art will understand that the specific execution order of each step in the above methods of specific implementation should be determined by its function and possible internal logic.

[0158] In addition, this embodiment also provides a cross-device fault diagnosis system and electronic device. The above can be used to implement any cross-device fault diagnosis method provided by the present invention. The corresponding technical solutions and descriptions are described in the corresponding descriptions in the method section, and will not be repeated here.

[0159] Secondly, referring to Figure 6 This invention provides a cross-device fault diagnosis system, including a data acquisition module 61, an adaptive calibration module 62, and a diagnosis module 63.

[0160] The acquisition module 61 is used to acquire health status signals of the target device.

[0161] The adaptive calibration module 62, connected to the acquisition module 61, is used to calculate the local health prototype and entity offset vector of the target device based on the health status signal and the preset hierarchical decoupling network (HDN) model. The entity offset vector is used to characterize the offset between the local health prototype and the preset global health prototype.

[0162] The diagnostic module 63, connected to the adaptive calibration module 62, is used to acquire the test signal of the target device and input it into the HDN model to obtain the test feature vector, as well as the fault diagnosis result used to determine the test feature vector based on the entity offset vector.

[0163] In some embodiments, the diagnostic module is further configured to calibrate the feature vector to be tested based on the entity offset vector to determine the fault diagnosis result of the feature vector to be tested. Alternatively, the diagnostic module is further configured to calibrate a preset global fault prototype based on the entity offset vector to determine the fault diagnosis result of the feature vector to be tested.

[0164] In some embodiments, the diagnostic module is used to calibrate the feature vector to be tested based on the entity offset vector, calculate the probability distribution of the calibrated feature vector to be tested belonging to the global health prototype and the global fault prototype, respectively, and determine the fault diagnosis result of the feature vector to be tested based on the probability distribution.

[0165] In some embodiments, the diagnostic module is used to calibrate a preset global fault prototype based on the entity offset vector to generate a local virtual fault prototype of the target device, and to calculate the probability distribution of the test feature vector belonging to the local healthy prototype and the local virtual fault prototype, respectively, and to determine the fault diagnosis result of the test feature vector based on the probability distribution.

[0166] In some embodiments, the cross-device fault diagnosis system further includes a construction module. The construction module is used to construct a pre-defined hierarchical decoupled network (HDN) model, wherein the HDN model includes a feature extraction backbone network, a fault branch, a device / condition branch, and a reconstruction decoder. The fault branch is used to extract general fault features, and the device / condition branch is used to extract device-specific baseline features.

[0167] In some embodiments, the building module is used to acquire the operating signals of at least two different source devices in several states, and to pre-train the feature extraction backbone network and the reconstruction decoder based on the operating signals using a self-supervised learning method to learn a general feature representation. It is also used to add parallel fault branches and device / condition branches at the output of the pre-trained feature extraction backbone network and the input of the reconstruction decoder to build an HDN model. Furthermore, it is used to train the HDN model using a meta-learning paradigm to decouple general fault features and device-specific baseline features from the general feature representation. In response to the convergence of the HDN model, it freezes the weights of the HDN model to obtain a preset hierarchical decoupled network HDN model.

[0168] In some embodiments, the building module includes a training unit, which is used to divide the running signal into several meta-tasks, wherein each meta-task includes a support set and a query set. The training unit is also used to train an HDN model based on the support set for each meta-task and to calculate a composite loss function based on the query set to optimize the HDN model, wherein the composite loss function includes prototype network loss, orthogonal loss and reconstruction loss.

[0169] In some embodiments, the diagnostic module further includes a global knowledge base, and is further configured to calculate a global health prototype and a global fault prototype based on a preset HDN model and the operating signals of at least two different source devices in several states, and to store the global health prototype and the global fault prototype in the global knowledge base.

[0170] In some embodiments, the diagnostic module is further configured to determine the fault type label to which the unknown abnormality belongs if the fault diagnosis result is an unknown abnormality, and to collect the fault status signal corresponding to the fault type label, and to calculate a new global fault prototype based on the preset HDN model and the fault status signal, and to add the new global fault prototype to the global knowledge base.

[0171] Thirdly, referring to Figure 7 This invention provides an electronic device, comprising:

[0172] At least one processor 701; a memory 702 communicatively connected to the at least one processor; and one or more I / O interfaces 703 connected between the processor 701 and the memory 702; wherein the memory 702 stores one or more computer programs executable by the at least one processor 701, the one or more computer programs being executed by the at least one processor 701 to enable the at least one processor 701 to perform a cross-device fault diagnosis method of the first aspect.

[0173] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0174] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method of diagnosing a cross-device fault, characterized by, The method comprises the following steps: Collecting a health state signal of a target device; According to the health state signal and a preset hierarchical decoupling network HDN model, a local health prototype and an entity offset vector of the target device are calculated, wherein the entity offset vector is used to represent the offset between the local health prototype and a preset global health prototype; Obtaining a to-be-tested signal of the target device in operation and inputting it into the HDN model to obtain a to-be-tested feature vector; Based on the entity offset vector, a fault diagnosis result of the to-be-tested feature vector is determined.

2. The diagnostic method of cross-device malfunctions according to claim 1, characterized in that, Based on the entity offset vector, a fault diagnosis result of the to-be-tested feature vector is determined, specifically including: Based on the entity offset vector, the to-be-tested feature vector is calibrated to determine the fault diagnosis result of the to-be-tested feature vector, Or, based on the entity offset vector, a preset global fault prototype is calibrated to determine the fault diagnosis result of the to-be-tested feature vector.

3. The diagnostic method of cross-device malfunctions according to claim 2, characterized in that, Based on the entity offset vector, the to-be-tested feature vector is calibrated to determine the fault diagnosis result of the to-be-tested feature vector, specifically including: Based on the entity offset vector, the to-be-tested feature vector is calibrated; The probability distribution of the calibrated to-be-tested feature vector belonging to the global health prototype and the global fault prototype respectively is calculated; According to the probability distribution, the fault diagnosis result of the to-be-tested feature vector is determined.

4. The diagnostic method of cross-device malfunctions according to claim 2, characterized in that, Based on the entity offset vector, the preset global fault prototype is calibrated to determine the fault diagnosis result of the to-be-tested feature vector, specifically including: Based on the entity offset vector, a local virtual fault prototype of the target device is generated by calibrating the preset global fault prototype; The probability distribution of the to-be-tested feature vector belonging to the local health prototype and the local virtual fault prototype respectively is calculated; According to the probability distribution, the fault diagnosis result of the to-be-tested feature vector is determined.

5. The method of diagnosing a cross-device fault according to claim 1, wherein, Before calculating the local health prototype and the entity offset vector of the target device according to the health state signal and the preset hierarchical decoupling network HDN model, the diagnosis method further comprises: Building a preset hierarchical decoupling network HDN model, wherein the HDN model comprises a feature extraction backbone network, a fault branch, a device / working condition branch, and a reconstruction decoder, the fault branch is used to extract general fault features, and the device / working condition branch is used to extract device-specific baseline features.

6. The diagnostic method of cross-device malfunctions according to claim 5, characterized in that, The building of the preset hierarchical decoupling network HDN model specifically includes: Obtaining running signals of at least two different source devices in several states; Using a self-supervised learning method, the feature extraction backbone network and the reconstruction decoder are pre-trained based on the running signals to learn general feature representations; Parallel fault branches and device / working condition branches are added at the output end of the pre-trained feature extraction backbone network and the input end of the reconstruction decoder to build the HDN model; Using a meta-learning paradigm, the HDN model is trained to decouple the general fault features and the device-specific baseline features from the general feature representations; In response to the convergence of the HDN model, the weights of the HDN model are frozen to obtain a preset hierarchical decoupling network HDN model.

7. The diagnostic method of cross-device malfunctions according to claim 6, characterized in that, The HDN model is trained using a meta-learning paradigm, specifically including: The operating signal is divided into a plurality of meta-tasks, wherein each meta-task includes a support set and a query set; For each meta-task, the HDN model is trained based on the support set, and a composite loss function is calculated based on the query set to optimize the HDN model, wherein the composite loss function includes a prototype network loss, an orthogonal loss, and a reconstruction loss.

8. The diagnostic method of cross-device malfunctions according to claim 2, characterized in that, The diagnostic method further includes: According to the preset HDN model and the operating signals of at least two different source devices in several states, the global health prototype and the global fault prototype are calculated respectively; The global health prototype and the global fault prototype are stored in a global knowledge base.

9. The diagnostic method of cross-device malfunctions according to claim 8, characterized in that, The fault diagnosis result includes an unknown abnormality, After storing the global health prototype and the global fault prototype in the global knowledge base, the diagnostic method further includes: If the fault diagnosis result is an unknown abnormality; Determine the fault type label to which the unknown abnormality belongs; Collect the fault state signal corresponding to the fault type label; According to the preset HDN model and the fault state signal, a new global fault prototype is calculated; The new global fault prototype is supplemented to the global knowledge base.

10. A diagnostic system for cross-device failures, the system comprising: Including: The acquisition module is configured to acquire a health state signal of a target device; The adaptive calibration module is connected to the acquisition module and is configured to calculate a local health prototype and an entity offset vector of the target device based on the health state signal and a preset hierarchical decoupling network HDN model, wherein the entity offset vector is used to represent the offset between the local health prototype and a preset global health prototype; The diagnostic module is connected to the adaptive calibration module and is configured to obtain a to-be-tested signal of the target device and input the to-be-tested signal into the HDN model to obtain a to-be-tested feature vector, and determine a fault diagnosis result of the to-be-tested feature vector based on the entity offset vector.

11. An electronic device, comprising: Including: At least one processor; And The memory is in communication connection with the at least one processor; wherein The memory stores one or more computer programs executable by the at least one processor, and the one or more computer programs are executed by the at least one processor to enable the at least one processor to execute the diagnostic method of the cross-device fault as claimed in any one of claims 1-9.

Images