Fault detection system for realizing high and low voltage power distribution cabinet based on intelligent AI

Abstract

The invention discloses a system for realizing fault detection of a high and low voltage power distribution cabinet based on intelligent AI, and the system comprises a sample construction module which is used for constructing N fault classification samples corresponding to N historical fault events, a model training module which is used for generating a fault classification model, and a necessary vector construction module which is used for constructing a necessary vector based on the fault classification samples of the N historical fault events. The system comprises a typical vector construction module for anchoring a typical necessary state vector, a typical template construction module for constructing a typical vector template, a real-time vector construction module for constructing a real-time typical necessary state vector, and a real-time classification detection module for performing early fault classification detection of a target fault category. The method not only reduces unnecessary data overhead, but also shields features irrelevant to the current fault, and prevents normal fluctuation from being misjudged as abnormity, thereby improving the accuracy and credibility of early warning in a real power distribution cabinet operation environment.

Description

Technical Field

[0001] This invention relates to the field of power distribution cabinet fault detection, specifically a fault detection system for high and low voltage power distribution cabinets based on intelligent AI. Background Technology

[0002] High- and low-voltage switchgear are critical power distribution and protection devices in power systems, and their operating status directly affects power supply safety and equipment lifespan. In actual operation and maintenance, common fault types such as poor contact, insulation degradation, and abnormal partial discharge often manifest as only slight abnormalities in individual electrical or thermal parameters in the early stages. If these are not identified in time, they can easily develop into serious accidents such as short circuits and fires.

[0003] Currently, fault detection in power distribution cabinets mainly relies on two types of methods: One method is single-parameter monitoring based on threshold alarms. This method is simple to implement, but it is difficult to capture early signs. Second, there is the intelligent diagnostic method based on multi-parameter fusion, which comprehensively judges the results by collecting multiple operating status parameters such as current, voltage, temperature, and partial discharge. While this type of method can improve detection sensitivity, it still faces significant challenges in practical applications: First, not all collected parameters are valuable for identifying specific faults. Using all parameters indiscriminately not only increases transmission costs but may also mask the true fault signal due to noise interference from irrelevant parameters. Secondly, there is a lack of explicit extraction and solidification mechanisms for "key feature combinations." Even if the model can classify correctly, operations and maintenance personnel still cannot determine "which parameters were abnormal and caused this alarm." Therefore, how to accurately select the features that are truly necessary for a specific fault category from multi-dimensional operating status parameters, and how to achieve interpretable early fault detection based on the selection results, has become an urgent problem to be solved in the current intelligent operation and maintenance of power distribution cabinets. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fault detection system for high and low voltage distribution cabinets based on intelligent AI. This system solves the technical problems mentioned in the background by introducing the labeling of necessary features and the construction of typical vector templates.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A fault detection system for high and low voltage switchgear based on intelligent AI includes the following application steps: The sample construction module is used to construct N fault classification samples corresponding to N historical fault events; The model training module is used to input N fault classification samples into a general multilayer perceptron for supervised training, and generate a fault classification model that can predict the target fault category and its complete state confidence. The necessary vector construction module is used to construct N necessary state vectors based on fault classification samples from N historical fault events. Typical vector construction module, used to anchor typical necessary state vectors among N necessary state vectors; The typical template building module is used to construct typical vector templates based on typical necessary state vectors. The real-time vector construction module is used to construct real-time typical necessary state vectors based on typical vector templates. The real-time classification and detection module is used to input the real-time typical necessary state vectors into the fault classification model to perform early fault classification and detection of the target fault category.

[0006] In some specific embodiments, the construction steps of the sample construction module include: S1-1, Determine the target fault category to be trained; S1-2. In the historical operation and maintenance database, anchor historical fault events for N target fault categories; S1-3. Construct fault classification samples for each of the N historical fault events.

[0007] In some specific embodiments, constructing fault classification samples for each of the N historical fault events includes: S1-3-1, Anchor the fault occurrence timestamps corresponding to N historical fault events; S1-3-2. Based on the fault occurrence time axis, trace back M time lengths and define an early state sampling timestamp that represents the early fault observation window; S1-3-3. Based on the early state sampling timestamp, obtain K operating state parameters of the high and low voltage switchgear. S1-3-4. Characterize the K operating state parameters to generate K operating state features; S1-3-5. Assign a unique feature number to each of the K running state features, and concatenate the K running state features based on the feature numbers to generate a running state vector. S1-3-6. Pair the operating state vector with the target fault category to form fault classification samples; S1-3-7. Traverse N historical fault events and repeat the pairing process to obtain N fault classification samples.

[0008] In some specific embodiments, the construction steps of the necessary vector construction module include: S3-1. Traverse the N fault classification samples and select the fault classification sample to be processed one by one; S3-2. Input the fault classification samples to be processed into the pre-trained fault classification model, and output the corresponding target fault category and complete state confidence. S3-3. Extract K operating status features arranged by feature number from the fault classification samples to be processed; S3-4. Along the ordered direction of the K feature numbers, mask the running state features one by one to generate a mask state vector; S3-5. Input the mask state vector into the pre-trained fault classification model and output the corresponding target fault category and mask state confidence. S3-6. Based on state confidence and mask state confidence, determine the necessary state vectors for N fault classification samples.

[0009] In some specific embodiments, the necessary state vectors for N fault classification samples are determined based on state confidence and mask state confidence, including: S3-6-1. Calculate the confidence difference between the complete state confidence and the masked state confidence. S3-6-2. If the confidence difference is greater than the preset confidence sensitivity threshold, the masked operating status feature will be marked as a necessary feature. S3-6-3. Traverse the K operating state features of the fault classification sample to be processed, and repeatedly perform the marking of necessary features until the J necessary features and KJ non-necessary features corresponding to the fault classification sample to be processed are marked. S3-6-4. Map the J necessary features back to a unified K-dimensional feature space to generate a K-dimensional necessary state vector; where the positions of non-necessary features are 0, and the positions of necessary features retain their original values; S3-6-5. Traverse N fault classification samples and repeatedly execute the mapping to generate N necessary state vectors.

[0010] In some specific embodiments, the construction steps of the typical vector construction module include: S4-1. Calculate the G similarities between each pair of the N necessary state vectors; where G = N*(N-1) / 2; S4-2. Arrange the G similarities in descending order to form a similarity sequence, and extract the H similarities from the first third of the sequence; where H = (1 / 3) * G S4-3. Extract 2H necessary state vectors corresponding to H similarities and count the frequency of occurrence of necessary state vectors; S4-4. Mark the necessary state vector that appears most frequently as the typical necessary state vector.

[0011] In some specific embodiments, the construction steps of the typical template building module include: S5-1, Anchor the feature numbers of each necessary and non-necessary feature in the typical necessary state vector; S5-2. Structure the feature indices and mask markers of necessary and unnecessary features into typical vector templates.

[0012] In some specific embodiments, the construction steps of the real-time vector construction module include: S6-1. Analyze the feature labels of the J necessary features in a typical vector template; S6-2. Based on the feature labels of J necessary features, obtain the corresponding J real-time running status features; S6-3. Fill the J real-time running state features into the feature numbers corresponding to the J dimensions in the typical vector template to generate a real-time typical necessary state vector.

[0013] This invention provides a fault detection system for high and low voltage switchgear based on intelligent AI, which has the following advantages: This invention identifies operational state features that would significantly decrease confidence after removal by masking each historical fault sample and calculating the difference in confidence of the model for the target fault category before and after masking. These features are then marked as necessary features. Subsequently, these necessary features are mapped back to K-dimensional space according to their original feature indices, retaining the original values ​​and setting the positions of non-necessary features to 0, thus forming the necessary state vector corresponding to the sample.

[0014] Based on this, the similarity between all pairs of necessary state vectors is calculated, the high similarity part is extracted and the frequency of occurrence of each vector is counted, and the necessary state vector with the highest frequency of occurrence is taken as the typical necessary state vector; the feature index corresponding to the non-zero element in this vector is the set of key features that have a common discriminative effect on the current fault category.

[0015] Furthermore, these non-zero positions are solidified into typical vector templates to form a binary mask structure, clearly identifying which feature indices belong to the necessary features.

[0016] During real-time detection, the system only collects the real-time operating status parameter values ​​corresponding to the necessary feature numbers specified by the template, fills them into the corresponding positions of the K-dimensional vector, and sets the remaining positions to 0, generating a real-time typical necessary state vector, which is then input into the original fault classification model for inference. Thus, early fault detection no longer relies on the input of all K operating status parameters, but is driven by a few necessary features that have been validated by a large number of historical fault samples. On the one hand, this reduces the overhead of data collection, transmission, and processing of non-essential features; on the other hand, it shields features irrelevant to the current fault (such as ignoring irrelevant parameters like cabinet ambient temperature or circuit breaker operation count when judging partial discharge anomalies using bus temperature and partial discharge quantity), preventing reasonable fluctuations of these features under normal operating conditions from being misjudged as abnormal by the model. This improves the accuracy of early alarms and the credibility of diagnostic results for maintenance personnel in the real distribution cabinet operating environment. Attached Figure Description

[0017] Figure 1 This is a structural block diagram of a fault detection system for high and low voltage distribution cabinets based on intelligent AI, according to the present invention. Figure 2 This is a schematic diagram of the processing flow of a fault detection system for high and low voltage distribution cabinets based on intelligent AI according to the present invention. Figure 3 This is a schematic diagram illustrating the construction of the necessary state vectors described in this invention; Figure 4 This is a schematic diagram illustrating the construction process of the typical necessary state vector described in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figures 1 to 2 This invention provides a fault detection system for high and low voltage distribution cabinets based on intelligent AI, including the following application modules: The sample construction module is used for S1 to construct N fault classification samples corresponding to N historical fault events; The model training module is used in S2 to input N fault classification samples into a general multilayer perceptron for supervised training, and generate a fault classification model that can predict the target fault category and its complete state confidence. The necessary vector construction module is used by S3 to construct N necessary state vectors based on fault classification samples from N historical fault events. Typical vector construction module, used in S4 to anchor typical necessary state vectors among N necessary state vectors; Typical template building module, used in S5 to build typical vector templates based on typical necessary state vectors; The real-time vector construction module is used in S6 to construct real-time typical necessary state vectors based on typical vector templates. The real-time classification and detection module is used by S7 to input the real-time typical necessary state vectors into the fault classification model to perform early fault classification and detection of the target fault category.

[0020] In this embodiment, by converting historical fault samples into necessary state vectors and extracting typical necessary state vectors and their templates, the fault classification model can rely only on necessary feature combinations for reasoning during the real-time detection stage. This not only preserves the discriminative ability but also enables focused identification of early fault symptoms, avoiding missed detections caused by redundant feature interference.

[0021] Example 2: See Figures 3 to 4 The technical solution of this embodiment 2 differs from that of embodiment 1 in that it discloses the specific application steps of each application module in embodiment 1.

[0022] Specifically, in this embodiment, the construction steps of the sample construction module include: S1-1, Determine the target fault category to be trained; Specifically, the target fault category refers to typical fault types that are prone to causing safety accidents during the operation of the distribution cabinet; the category includes: poor contact, insulation degradation, abnormal partial discharge, overload operation, grounding fault, and circuit breaker contact erosion; this fault category uses the fault types recorded in the historical operation and maintenance logs as tags, and is classified and labeled through expert systems or manual review. S1-2. In the historical operation and maintenance database, anchor historical fault events for N target fault categories; The historical operation and maintenance database refers to a structured dataset that stores power distribution cabinet operating status parameters, fault occurrence time, and maintenance records; historical fault events are identified as fault events with unique numbers, each event containing its occurrence timestamp, the number of the equipment to which it belongs, the fault type label, and the associated raw sensor data.

[0023] S1-3. Construct fault classification samples for each of the N historical fault events.

[0024] In this embodiment, by using real fault events in historical operation and maintenance logs as a basis and combining expert annotations to clarify the target fault category, the consistency between training data and the operation scenario of the power distribution cabinet is ensured.

[0025] Steps S1-3 further include: S1-3-1, Anchor the fault occurrence timestamps corresponding to N historical fault events; S1-3-2. Based on the fault occurrence time axis, trace back M time lengths and define an early state sampling timestamp that represents the early fault observation window; S1-3-3. Based on the early state sampling timestamp, obtain K operating state parameters of the high and low voltage switchgear. The operating status parameters include quantitative parameters that represent the operating status of the distribution cabinet, such as three-phase current, voltage, bus temperature, ambient temperature inside the cabinet, power factor, number of circuit breaker operations, and insulation resistance.

[0026] S1-3-4. Characterize the K operating state parameters to generate K operating state features; In this embodiment, characterization refers to processing the running state parameters into numerical features suitable for model input.

[0027] S1-3-5. Assign a unique feature number to each of the K running state features, and concatenate the K running state features based on the feature numbers to generate a running state vector. In this embodiment, the feature number is used to uniquely identify and locate each running state feature; of course, concatenating them in an orderly manner can also structure the feature dimensions and ensure the consistency of the model input format. S1-3-6. Pair the operating state vector with the target fault category to form fault classification samples; S1-3-7. Traverse N historical fault events and repeat the pairing process to obtain N fault classification samples.

[0028] In this embodiment, by defining an early observation window M time units before the fault occurs and extracting K operating state parameters as features, the generated fault classification samples reflect the state in the early stage of the fault rather than the deterioration stage.

[0029] Specifically, in this embodiment, the construction steps of the necessary vector construction module include: S3-1. Traverse the N fault classification samples and select the fault classification sample to be processed one by one; S3-2. Input the fault classification samples to be processed into the pre-trained fault classification model, and output the corresponding target fault category and complete state confidence. In this embodiment, the fault classification model adopts a multilayer perceptron (MLP) that supports probabilistic output. Its network structure includes several fully connected hidden layers and a softmax activated output layer. The softmax layer converts the forward propagation feature vector into a normalized state confidence (i.e., a probability distribution), where each dimension corresponds to a predefined fault category. The actual output of the model is the confidence and its corresponding fault category, rather than a single hard label.

[0030] For example, Softmax activation can be represented as: ; in: This represents the unnormalized score of the i-th fault category, which is the unnormalized score output by the penultimate layer of the neural network, reflecting the model's original prediction strength for that category. This represents the total number of predefined target fault categories. For example, if there are 6 categories such as poor contact and insulation degradation, then C=6. This indicates that the exponential function acts on the logit to amplify high scores and enhance classification discrimination. This represents the exponential sum of the logit values ​​for all categories, used as the normalized denominator to ensure that the total output sum is 1. This represents the final confidence level (probability) for the i-th category, with a value ranging from (0,1).

[0031] S3-3. Extract K operating status features arranged by feature number from the fault classification samples to be processed; S3-4. Along the ordered direction of the K feature numbers, mask the running state features one by one to generate a mask state vector; S3-5. Input the mask state vector into the pre-trained fault classification model and output the corresponding target fault category and mask state confidence. It should be noted that the masked state vector is a K-dimensional vector formed by replacing the feature value corresponding to a specific index with a preset mask value (such as 0 or the historical mean of the feature) on the basis of the original running state vector, while keeping the other features unchanged. It simulates the model input state when the feature is unavailable by fixing the input dimension and only masking a single feature, thereby evaluating the sensitivity of the feature to fault detection.

[0032] S3-6. Based on state confidence and mask state confidence, determine the necessary state vectors for N fault classification samples.

[0033] In this embodiment, by masking features one by one and comparing changes in confidence, the contribution of each running state feature to the output is quantified, thereby identifying the necessary features that play a key role in the current sample discrimination.

[0034] Step S3-6 further includes: S3-6-1. Calculate the confidence difference between the complete state confidence and the masked state confidence. The confidence difference indicates the degree to which the model’s confidence in the target fault category decreases after a certain operational feature is removed; in other words, the larger the difference, the more critical the contribution of the removed feature to maintaining the current fault judgment.

[0035] S3-6-2. If the confidence difference is greater than the preset confidence sensitivity threshold, the masked operating status feature will be marked as a necessary feature. Specifically, the essential feature means that for the current fault classification sample to be processed, this feature is the key discrimination criterion that the model relies on when identifying its target fault category. Its absence will lead to a significant reduction in the model's confidence in that category.

[0036] S3-6-3. Traverse the K operating state features of the fault classification sample to be processed, and repeatedly perform the marking of necessary features until the J necessary features and KJ non-necessary features corresponding to the fault classification sample to be processed are marked. S3-6-4. Map the J necessary features back to a unified K-dimensional feature space to generate a K-dimensional necessary state vector; where the positions of non-necessary features are 0, and the positions of necessary features retain their original values; S3-6-5. Traverse N fault classification samples and repeatedly execute the mapping to generate N necessary state vectors.

[0037] In this embodiment, necessary features are mapped back to a unified K-dimensional space and non-necessary features are set to zero to generate necessary state vectors with consistent dimensions. This not only preserves the physical location information of the original features, but also provides a comparable data foundation for cross-sample similarity calculation.

[0038] Specifically, in this embodiment, the construction steps of the typical vector construction module include: S4-1. Calculate the G similarities between each pair of the N necessary state vectors; where G = N*(N-1) / 2; The similarity can be cosine similarity, which can measure the directional consistency of two necessary state vectors without being affected by the magnitude of their respective vectors (i.e. the absolute value of the feature), and is suitable for feature alignment of early faults in different devices. The similarity calculation formula is: ; Where u and v represent any two K-dimensional necessary state vectors, and u⋅v is the dot product of the two vectors; , These are the Euclidean norms (i.e., vector lengths) of the two vectors, respectively. S4-2. Arrange the G similarities in descending order to form a similarity sequence, and extract the H similarities from the first third of the sequence; where H = (1 / 3) * G S4-3. Extract 2H necessary state vectors corresponding to H similarities and count the frequency of occurrence of necessary state vectors; It should be noted that each similarity is calculated from two necessary state vectors, and the H similarities are the first third of the high similarity results extracted after all G similarities are sorted in descending order; therefore, there must be a large number of duplicate terms in the corresponding 2H necessary state vectors.

[0039] For example, if a certain high similarity is calculated from necessary state vectors A and B, and another high similarity is calculated from vectors X and B, then among the four vectors involved (A, B, X, B), vector B appears twice, and its frequency of occurrence is 50%.

[0040] S4-4. Mark the necessary state vector that appears most frequently as the typical necessary state vector. The highest frequency of occurrence means that the necessary state vector is repeatedly associated in highly similar sample pairs, representing a set of fault instances with highly consistent combinations of necessary features; therefore, it can be regarded as a typical early symptom pattern under the current fault category and has strong representativeness.

[0041] Specifically, in this embodiment, the construction steps of the typical template construction module include: S5-1, Anchor the feature numbers of each necessary and non-necessary feature in the typical necessary state vector; S5-2. Structure the feature indices and mask markers of necessary and unnecessary features into typical vector templates.

[0042] Specifically, a typical vector template represents a binary mask structure, in which the necessary feature indices that are discriminative for the target fault category and their corresponding physical meanings are marked, while the positions of unnecessary features are marked as negligible items; For example, if the typical necessary state vector has only the 2nd position (bus temperature) and the 5th position (partial discharge quantity) non-zero in the K=6 dimensional space, then its corresponding template can be represented as [0,1,0,0,1,0], where 1 represents a necessary feature and 0 represents a non-necessary feature.

[0043] Specifically, in this embodiment, the construction steps of the real-time vector construction module include: S6-1. Analyze the feature labels of the J necessary features in a typical vector template; S6-2. Based on the feature labels of J necessary features, obtain the corresponding J real-time running status features; S6-3. Fill the J real-time running state features into the feature numbers corresponding to the J dimensions in the typical vector template to generate a real-time typical necessary state vector.

[0044] In this embodiment, by extracting real-time observations of only J necessary features based on a typical vector template, and constructing a K-dimensional sparse real-time typical necessary state vector, the fault classification model relies only on a small number of key features that have been verified to be discriminative for the target fault category during the inference stage. It does not need to process all K operating state features, which reduces real-time computational overhead and avoids interference introduced by unnecessary features. It achieves highly focused inference for early faults while ensuring detection accuracy.

[0045] In summary, this invention achieves effective detection of early faults in high and low voltage distribution cabinets through the following process: First, K operational state parameters are extracted from the early observation window before the occurrence of historical failure events to construct failure classification samples reflecting the initial state of the failure. Second, using a pre-trained multilayer perceptron model, necessary features that play a key role in the current failure category are identified by masking operational state features one by one and comparing the difference between the confidence of the complete state and the confidence of the masked state. These necessary features are then stored in K-dimensional space with their original values ​​preserved and non-necessary features set to 0, generating N necessary state vectors. Next, the cosine similarity between each pair of the N necessary state vectors is calculated, the high similarity portions are extracted, and the frequency of vector occurrence is counted. The necessary state vector with the highest frequency is taken as the typical necessary state vector. Subsequently, the feature indices corresponding to the non-zero positions are solidified into a typical vector template. This template explicitly identifies the necessary feature positions that are discriminative for the target fault category in the form of a binary mask. Finally, in the real-time detection stage, only the real-time operating status parameter values ​​of the J necessary features specified by the template are read, the corresponding dimensions are filled in, and the remaining positions are set to 0 to construct a real-time typical necessary state vector, which is then input into the original fault classification model for inference. The output fault classification result is the target fault category and its corresponding high confidence probability value, which can characterize the degree of matching between the current equipment operating status and the historical typical early fault modes. This enables maintenance personnel to trigger targeted early warning and handling measures in a timely manner before the fault deteriorates.

[0046] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means.

[0047] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0048] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A fault detection system for high and low voltage switchgear based on intelligent AI, characterized in that, include: The sample construction module is used to construct N fault classification samples corresponding to N historical fault events; The model training module is used to input N fault classification samples into a general multilayer perceptron for supervised training, and generate a fault classification model that can predict the target fault category and its complete state confidence. The necessary vector construction module is used to construct N necessary state vectors based on fault classification samples from N historical fault events. Typical vector construction module, used to anchor typical necessary state vectors among N necessary state vectors; The typical template building module is used to construct typical vector templates based on typical necessary state vectors. The real-time vector construction module is used to construct real-time typical necessary state vectors based on typical vector templates. The real-time classification and detection module is used to input the real-time typical necessary state vectors into the fault classification model to perform early fault classification and detection of the target fault category.

2. The fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 1, characterized in that, The construction steps of the sample construction module include: S1-1, Determine the target fault category to be trained; S1-2. In the historical operation and maintenance database, anchor historical fault events for N target fault categories; S1-3. Construct fault classification samples for each of the N historical fault events.

3. The fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 2, characterized in that, The construction of fault classification samples for each of the N historical fault events includes: S1-3-1, Anchor the fault occurrence timestamps corresponding to N historical fault events; S1-3-2. Based on the fault occurrence time axis, trace back M time lengths and define an early state sampling timestamp that represents the early fault observation window; S1-3-3. Based on the early state sampling timestamp, obtain K operating state parameters of the high and low voltage switchgear. S1-3-4. Characterize the K operating state parameters to generate K operating state features; S1-3-5. Assign a unique feature number to each of the K running state features, and concatenate the K running state features based on the feature numbers to generate a running state vector. S1-3-6. Pair the operating state vector with the target fault category to form fault classification samples; S1-3-7. Traverse N historical fault events and repeat the pairing process to obtain N fault classification samples.

4. The fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 1, characterized in that, The construction steps of the necessary vector construction module include: S3-1. Traverse the N fault classification samples and select the fault classification sample to be processed one by one; S3-2. Input the fault classification samples to be processed into the pre-trained fault classification model, and output the corresponding target fault category and complete state confidence. S3-3. Extract K operating status features arranged by feature number from the fault classification samples to be processed; S3-4. Along the ordered direction of the K feature numbers, mask the running state features one by one to generate a mask state vector; S3-5. Input the mask state vector into the pre-trained fault classification model and output the corresponding target fault category and mask state confidence. S3-6. Based on state confidence and mask state confidence, determine the necessary state vectors for N fault classification samples.

5. A fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 4, characterized in that, Based on state confidence and masked state confidence, the necessary state vectors for N fault classification samples are determined, including: S3-6-1. Calculate the confidence difference between the complete state confidence and the masked state confidence. S3-6-2. If the confidence difference is greater than the preset confidence sensitivity threshold, the masked operating status feature will be marked as a necessary feature. S3-6-3. Traverse the K operating state features of the fault classification sample to be processed, and repeatedly perform the marking of necessary features until the J necessary features and KJ non-necessary features corresponding to the fault classification sample to be processed are marked. S3-6-4. Map the J necessary features back to a unified K-dimensional feature space to generate a K-dimensional necessary state vector; where the positions of non-necessary features are 0, and the positions of necessary features retain their original values; S3-6-5. Traverse N fault classification samples and repeatedly execute the mapping to generate N necessary state vectors.

6. The fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 5, characterized in that, The construction steps of the typical vector construction module include: S4-1. Calculate the G similarities between each pair of the N necessary state vectors; where G = N * (N-1) / 2; S4-2. Arrange the G similarities in descending order to form a similarity sequence, and extract the H similarities from the first third of the sequence; where H = (1 / 3) * G S4-3. Extract 2H necessary state vectors corresponding to H similarities and count the frequency of occurrence of necessary state vectors; S4-4. Mark the necessary state vector that appears most frequently as the typical necessary state vector.

7. A fault detection system for high and low voltage distribution cabinets based on intelligent AI as described in claim 6, characterized in that, The construction steps of the typical template building module include: S5-1, Anchor the feature numbers of each necessary and non-necessary feature in the typical necessary state vector; S5-2. Structure the feature indices and mask markers of necessary and unnecessary features into typical vector templates.

8. A fault detection system for high and low voltage distribution cabinets based on intelligent AI according to claim 7, characterized in that, The construction steps of the real-time vector construction module include: S6-1. Analyze the feature labels of the J necessary features in a typical vector template; S6-2. Based on the feature labels of J necessary features, obtain the corresponding J real-time running status features; S6-3. Fill the J real-time running state features into the feature numbers corresponding to the J dimensions in the typical vector template to generate a real-time typical necessary state vector.

Images