Artificial intelligence-based electrical automation device fault diagnosis method

Abstract

The application discloses an electrical automation equipment fault diagnosis method based on artificial intelligence, and particularly relates to the technical field of equipment fault diagnosis, and the method comprises the following steps: collecting multi-source electrical signals, suppressing noise through adaptive morphological filtering of a particle swarm optimization, performing bispectrum analysis on the filtered signals to extract energy peak values, frequency pairs and bispectrum entropy as preliminary features, obtaining a feature vector F1 through Relief-F screening and PCA dimension reduction; aligning F1 with equipment operating conditions, control instructions and static semantic information in time and space, constructing a sequence and adding sinusoidal position coding, obtaining a joint feature tensor F2 through a Transformer encoder to fuse multi-modal data; and constructing a fault mechanism knowledge graph containing fault phenomenon, reason and location entities, wherein the application can extract early weak fault features from strong noise, fuse multi-source information to reduce false positives, and assist operation and maintenance personnel to quickly and accurately repair through an interpretable reasoning process.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and more specifically, to a method for diagnosing faults in electrical automation equipment based on artificial intelligence. Background Technology

[0002] Electrical automation equipment (such as frequency converters, servo drives, high-voltage switchgear, transformers, etc.) is a core component of modern industrial production, and its operating status directly affects production safety and efficiency. Electrical faults (including insulation degradation, partial discharge, power device aging, poor contact, parameter drift, etc.) are characterized by their suddenness, high concealment, and great harm. Early fault signals are often extremely weak and easily drowned out by strong power frequency signals and background noise. Traditional fault diagnosis methods mainly rely on the analysis of single electrical quantities, such as Fourier transform, wavelet analysis, or simple threshold judgment. These methods are effective for processing stationary signals, but they are difficult to effectively extract features for early, intermittent, non-stationary, and weak electrical fault signals, resulting in insufficient fault early warning capabilities. Often, the fault is only detected when it has developed to a severe stage, causing equipment downtime or even safety accidents.

[0003] With the development of artificial intelligence technology, fault diagnosis methods based on neural networks have gradually become a research hotspot. Existing methods typically collect electrical parameters such as current and voltage from equipment and use deep neural networks for feature learning and fault classification. However, these methods still have the following drawbacks: First, they rely solely on electrical data from a single mode, ignoring rich information such as the equipment's control logic, operating conditions, and historical maintenance records. When a minor electrical anomaly occurs, it is impossible to determine whether it is a normal operation or a precursor to a fault by combining it with the current control commands, resulting in a high false alarm rate. Second, deep neural networks, as "black box" models, can output fault types but cannot provide a reasoning process that conforms to physical laws. Maintenance personnel can only passively accept the diagnostic results, making it difficult to understand the mechanism and evolution path of the fault, which is not conducive to rapid and accurate on-site repair.

[0004] To address the aforementioned issues, how to effectively extract early, subtle electrical fault characteristics from a background of strong noise, how to integrate multi-dimensional information for comprehensive analysis to reduce false alarms, and how to make the diagnostic process interpretable have become pressing technical challenges in the field of electrical equipment fault diagnosis. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for diagnosing faults in electrical automation equipment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An artificial intelligence-based fault diagnosis method for electrical automation equipment includes the following steps:

[0008] Multi-source electrical signals are acquired and preprocessed. The feature vector F1 of early micro-weak electrical faults is extracted by adaptive morphological filtering and high-order spectral analysis.

[0009] The operating condition data, control commands, and static semantic information of the collected equipment are spatiotemporally aligned with the feature vector F1, and multimodal fusion is performed through a Transformer encoder to obtain the joint feature tensor F2.

[0010] A fault mechanism knowledge graph containing fault phenomena, fault causes, fault location entities and their causal relationships is constructed. The joint feature tensor F2 is mapped to the graph nodes, and causal reasoning is performed using a graph neural network to output interpretable diagnostic results.

[0011] Specifically, the step of extracting the early micro-current fault feature vector F1 includes:

[0012] The original three-phase voltage and current signals are synchronously acquired using high-frequency sensors;

[0013] The particle swarm optimization algorithm is used to adaptively select the optimal structuring element type and length, and morphological opening-closing and closing-opening combination operations are performed on the original signal to suppress noise.

[0014] The filtered signal is subjected to segmented bispectral analysis to extract bispectral energy peaks, frequency pairs, and diagonal slice entropy as preliminary features.

[0015] The preliminary features are screened using Relief-F importance assessment, and then dimensionality is reduced by principal component analysis to obtain a low-dimensional, highly discriminative feature vector F1.

[0016] Specifically, the particle swarm optimization algorithm includes:

[0017] With the goal of maximizing the signal-to-noise ratio of the filtered signal, each particle is encoded as a two-dimensional vector of structuring element type and length. The particle velocity and position are updated iteratively until the maximum number of iterations or the fitness value converges, and the globally optimal combination of structuring elements is output.

[0018] Specifically, in the Relief-F importance assessment step:

[0019] Calculate the weight of each feature in distinguishing different fault categories, and retain the features with the highest weight ranking; in the principal component analysis step, select the principal components whose cumulative variance contribution rate reaches a preset threshold, and project the original features onto the principal component direction to obtain the dimensionality-reduced feature vector F1.

[0020] Specifically, the multimodal fusion includes:

[0021] Align the feature vector F1, the real-time read control commands and operating parameters, and the static semantic information obtained from the equipment ledger on the time axis;

[0022] The aligned data is constructed into a sequence with fixed time windows, where electrical features and operating parameters are arranged by time step, and static semantic features are copied as global tokens and concatenated to each time step.

[0023] A sinusoidal positional code is added to the sequence, and then input into a Transformer encoder. Cross-modal associations are learned through a self-attention mechanism, and a joint feature tensor F2 is output.

[0024] Specifically, the sinusoidal position code is generated in the following manner:

[0025] The sine and cosine function values ​​are calculated based on the time step position and dimension index to form a position encoding matrix, which is then added element by element to the input token sequence.

[0026] Specifically, the steps for constructing the fault mechanism knowledge graph include:

[0027] Define the entity types of electrical faults, fault phenomena, fault causes, fault locations, influencing parameters, and maintenance measures;

[0028] Knowledge is extracted from electrical equipment design manuals, fault repair case libraries, and expert experience, and the causal and influence relationships between entities are stored in the form of triples.

[0029] Specifically, the steps of mapping the joint feature tensor F2 to graph nodes include:

[0030] Predefine a semantic prototype vector for each fault phenomenon node;

[0031] The F2 and prototype vectors are mapped to the same semantic space using a trainable projection matrix, and the attention score is calculated as the initial confidence of each phenomenon node.

[0032] Set the initial confidence level of other types of nodes in the graph to zero.

[0033] Specifically, the steps for causal reasoning performed by the graph neural network include:

[0034] Using the initial confidence level of each node as input, information is propagated and aggregated along the causal edges of the knowledge graph to update the confidence levels of the fault cause and fault location nodes;

[0035] Output one or more fault causes with the highest confidence level, and trace back the message passing path of the graph neural network to generate an interpretable diagnostic report containing the inference logic chain and recommended maintenance solutions.

[0036] The technical effects and advantages of this invention are as follows:

[0037] This invention effectively suppresses strong background noise and extracts nonlinear, non-Gaussian features of early, minor electrical faults by combining adaptive morphological filtering with high-order spectral analysis. Relief-F filtering and PCA dimensionality reduction are then applied to obtain low-dimensional, highly discriminative feature vectors, significantly improving the sensitivity and accuracy of fault detection. Simultaneously, electrical features are spatiotemporally aligned with equipment operating conditions, control commands, and static semantic information. Multimodal data is fused using a Transformer encoder, and a self-attention mechanism is employed to learn cross-modal correlations. Position coding is introduced to enhance temporal perception, enabling accurate judgment of abnormal signals and significantly reducing false alarms caused by changes in operating conditions.

[0038] Furthermore, this invention constructs a fault mechanism knowledge graph containing fault phenomena, causes, and location entities. Through semantic attention, fused features are mapped to initial confidence levels for phenomenon nodes. A graph neural network is then used to propagate along causal edges to update the confidence levels, ultimately outputting the most probable fault cause and backtracking the complete reasoning path. This method not only provides high-confidence diagnostic conclusions but also generates interpretable logical chains and recommended maintenance solutions that conform to physical laws. This enables maintenance personnel to understand the fault mechanism and quickly locate and repair the fault, significantly improving the intelligence level of electrical equipment fault diagnosis and on-site maintenance efficiency. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] 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.

[0041] like Figure 1 As shown, the steps of the artificial intelligence-based fault diagnosis method for electrical automation equipment are as follows:

[0042] Step 1: Feature extraction of weak electrical faults. First, electrical signals are synchronously acquired using a high-frequency sensor. Adaptive morphological filtering with particle swarm optimization is used to suppress noise while preserving signal edges. Then, bispectral analysis is performed on the filtered signal to extract features such as energy peaks, frequency pairs, and bispectral entropy. Finally, Relief-F filtering and PCA dimensionality reduction are applied to obtain a low-dimensional, high-discrimination feature vector F1. The process is as follows:

[0043] Multi-source electrical signal synchronous acquisition: High-frequency voltage and current sensors are deployed at key nodes (such as input / output terminals, IGBT modules, and DC buses) of the monitored electrical automation equipment (taking frequency converters as an example) to synchronously acquire the original three-phase voltage (Ua, Ub, Uc) and three-phase current (Ia, Ib, Ic) signals at a sampling rate of not less than 1MHz.

[0044] Adaptive morphological filtering preprocessing: First, the acquired raw electrical signal is normalized. Second, a structure element pool consisting of various structural elements such as lines, triangles, and semicircles is constructed. Then, using a particle swarm optimization algorithm, with the goal of maximizing the signal-to-noise ratio of the filtered signal, the optimal structure element type and length are adaptively selected online, and morphological opening-closing and closing-opening combination operations are performed on the signal. This effectively suppresses random white noise and impulse interference while protecting the signal's edge and impulse characteristics, which is crucial for preserving the steep rising edge of the early discharge signal.

[0045] Specifically, the particle swarm optimization algorithm is configured as follows: Each particle is encoded as a two-dimensional vector (Type, L), where Type takes the value 1, 2, or 3, corresponding to linear, triangular, and semi-circular structuring elements, respectively; L is the length of the structuring element, with the search range set to an integer from 3 to the sampling rate / 100 to ensure that the structuring element length covers the signal feature scale without excessive smoothing. The population size is set to 30, and the maximum number of iterations is 50. The fitness function is defined as the signal-to-noise ratio of the filtered signal. ,in For signal-to-noise ratio, To extract the expected frequency band signal power from the original signal using a bandpass filter. This represents the noise power in the filtered residual signal. It is a logarithmic function to the base 10. The algorithm iteratively updates the velocity and position of particles until it reaches the maximum number of iterations or the fitness value no longer increases within 10 consecutive iterations. Finally, it outputs the globally optimal (Type, L) combination, which is used for morphological open-close and closed-open combination operations of actual signals.

[0046] Fault feature extraction based on higher-order spectrum: The filtered signal is segmented, with each segment serving as an analysis window; the bispectrum (two-dimensional Fourier transform of the third-order cumulant) is calculated for the signal within each window. Bispectral analysis can automatically suppress Gaussian noise and detect secondary phase coupling phenomena in the signal, which are typical characteristics of nonlinear faults (such as partial discharge and loose contact).

[0047] From the obtained bispectral matrix, the energy peak of the non-Gaussian signal, the corresponding frequency pair (f1, f2), and the bispectral diagonal slice entropy are extracted as sensitive feature vectors F1 for early weak electrical faults. These features can accurately quantify the nonlinearity and non-Gaussianity of the fault.

[0048] To improve the discriminative power of features and reduce redundancy, feature selection and dimensionality reduction mechanisms are added to make the final feature vector F1 input to the subsequent fusion model more representative and compact. The process is as follows:

[0049] The initial feature set extracted from the bispectral matrix (including energy peaks, corresponding frequency pairs, bispectral diagonal slice entropy, etc.) is usually high-dimensional and contains redundant information. To remove irrelevant features and reduce computational complexity, the following steps are used for feature selection and dimensionality reduction:

[0050] Feature importance assessment based on Relief-F: For each sample, its initial feature vector is denoted as... ,in The initial total number of features is given (e.g., M=50). The Relief-F algorithm is used to calculate the value of each feature. weight This reflects the feature's ability to distinguish between different fault categories. The weight update formula is: ;in, For the first The weights of each feature are initially set to 0 and are updated iteratively. For feature index, For the sample Category to which it belongs For different Category The feature vector of the current sample. To and The first of its kind The nearest neighbor samples, To and Different classes The The nearest neighbor samples, This represents the number of nearest neighbors (K=10). For the number of iterations, For category The prior probability, Calculate samples and In the Normalized differences on each feature.

[0051] Calculate the weights of all features After that, only the top-ranked items by weight are retained. Features (e.g.) =30), forming the filtered feature vector. .

[0052] Dimensionality reduction based on principal component analysis (PCA): Dimensionality reduction of the selected feature vectors Perform PCA dimensionality reduction. First, calculate the covariance matrix of the features in the training set. ,in The covariance matrix of the features of the training set. For the first Feature vectors of samples after Relief-F filtering ( dimension), For matrix transpose, For the number of training samples, The eigenvalues ​​are the eigenvalues. Then, the eigenvalues ​​of the covariance matrix are solved. and the corresponding feature vectors Select the first d principal components such that the cumulative variance contribution rate is... Original features Projecting onto the selected principal component directions yields the final dimensionality-reduced feature vector:

[0053] ;

[0054] in The final feature vector after dimensionality reduction (dimension d=128), which is the output of step one, It is a matrix composed of the eigenvectors of the first d principal components. , This is the feature vector of the current sample after Relief-F filtering. In this embodiment, d=128, consistent with the input dimension d1 of the subsequent Transformer encoder.

[0055] Through this mechanism, the original high-dimensional features are compressed into a low-dimensional, highly discriminative feature vector F1, which effectively reduces the computational burden of the subsequent multimodal fusion model and improves the accuracy of fault diagnosis.

[0056] Step 2: Multimodal data fusion. The electrical features F1 are spatiotemporally aligned with the operating conditions (instructions, parameters) read from the controller and the static semantic information obtained from the ledger. Multimodal data is fused using a Transformer encoder, and the associations are learned using a self-attention mechanism. A joint feature tensor F2 is output, and position encoding is added to enhance temporal awareness. The process is as follows:

[0057] Heterogeneous data acquisition and preprocessing:

[0058] Electrical data stream: Receives the continuous feature vector F1 extracted in step one;

[0059] Control and status data stream: Read control commands (such as start, stop, accelerate, decelerate), operating parameters (such as switching frequency, modulation ratio, given speed), and fault alarm codes from the device controller (PLC or embedded controller) in real time;

[0060] Semantic data: Retrieves static information such as the equipment model, rated parameters, commissioning date, and historical maintenance records from the equipment ledger management system;

[0061] Spatiotemporal alignment:

[0062] Time alignment: Unify all data streams onto the same timeline. Downsample high-speed electrical characteristic data (F1) or timestamp match it with lower-speed control command data to ensure that electrical characteristics within the same time window correspond to the operating conditions of the equipment;

[0063] Spatial alignment: Establishing a mapping relationship between electrical signal measurement points and control logic units. For example, associating the collector voltage signal of the IGBT module with the control module that generates the IGBT drive pulse;

[0064] Multimodal feature fusion: Construct a fusion model based on a Transformer encoder; use the time-aligned electrical feature vector F1 as a numerical input, encode control commands and operating parameters into continuous numerical vectors, and embed textual information such as equipment model and historical maintenance records into semantic feature vectors through a pre-trained BERT model.

[0065] The Transformer model automatically learns the intrinsic correlations between different modal data through its self-attention mechanism. For example, when a weak spike pulse appears in the electrical characteristics, the model simultaneously "pays attention" to whether the control command at that time is a normal PWM switching action. If both match, the importance of that feature is reduced (suppressing false alarms); if the control command is smooth but the electrical characteristics are abnormal, the importance of that feature is increased. The model outputs a joint feature tensor F2 that integrates electrical, operating condition, and semantic information.

[0066] The specific construction method of the fusion model based on the Transformer encoder is as follows:

[0067] Sequence Construction: Data within a time window T=1 second is treated as a sample. First, an electrical feature vector F1 (with dimensions of 1) is extracted every 0.1 seconds within this window. =128) are arranged in chronological order to form an electrical characteristic subsequence of length 10; simultaneously, control commands (such as speed setpoint) and operating parameters (such as output frequency) at the same moment within this window are normalized and also arranged in chronological order to form a working condition parameter subsequence of length 10 (dimension: =16). The two numerical sequences mentioned above will be used as numerical tokens for the Transformer. Furthermore, the static semantic feature vector extracted from the BERT model (dimension: 16) will be used. =768) is used as a special global token, copied 10 times, and concatenated into the features of each time step. Finally, the sequence length of the input Transformer is 10, and the dimension of each token is d1+d2+d3=912.

[0068] The process of adding sinusoidal positional encoding to the input sequence is as follows:

[0069] After sequence construction is complete, the input token at each time step contains a concatenated vector of electrical features, operating parameters, and static semantic information, with dimensions of [dimensional value missing]. The sequence length is L=10. To enable the Transformer model to be aware of the temporal order, a sinusoidal position code is added to the tokens at each time step before inputting them to the encoder.

[0070] Position encoding matrix The The elements in the row and the i-th column are calculated using the following formula:

[0071] ;

[0072] in, This is a position encoding matrix with dimensions L×D. Indicates the time step position index. This represents the dimension index, where D is the dimension of the input token at each time step, and L is the sequence length (L=10). , These are sine and cosine functions, respectively. This encoding method enables unique and distinguishable positional representations between different time steps, and the relative positions of any two time steps can be expressed through a linear transformation.

[0073] Compare the position encoding matrix with the input token sequence: By adding elements one by one, we obtain the final input containing temporal information: ,in The final input sequence after adding position encoding has dimensions L×D. The original input token sequence is composed of electrical features, operating parameters, and static semantic features, with dimensions L×D. This is the position encoding matrix;

[0074] Then, The data is then fed into a Transformer encoder for subsequent multimodal fusion. By introducing position encoding, the model can effectively utilize temporal sequence information, which is of great significance for the accurate localization of the gradual process of early faults or transient events.

[0075] Model Structure: The encoder part uses a 4-layer Transformer block, with each layer containing 8 attention heads. The feedforward network has a dimension of 2048. Through a self-attention mechanism, the model learns the interdependencies between different time steps and between electrical quantities and operating conditions.

[0076] Training Method: Supervised learning was employed for training. Historical operational data was used, with known fault types (including "no fault," "DC bus capacitor aging," and "IGBT drive anomaly") as labels. Cross-entropy loss was used as the loss function, AdamW was selected as the optimizer, and the initial learning rate was set to 1e-4, with a batch size of 32. The entire fusion network was trained by minimizing the error between predicted and actual faults, ultimately outputting a discriminative joint feature tensor F2 (256 dimensions).

[0077] Step 3: Graph Neural Network-Based Interpretable Diagnosis. A knowledge graph containing entities such as fault phenomena, causes, and locations is constructed. Semantic attention maps F2 to the initial confidence scores of the phenomenon nodes. The confidence scores are updated by propagating along causal edges using the graph neural network, inferring the most probable cause of the fault, and backtracking to generate an interpretable diagnostic report containing the reasoning path. The process is as follows:

[0078] Constructing a knowledge graph of electrical fault mechanisms: Ontology layer definition: Define entity types such as "electrical fault", "fault phenomenon", "fault cause", "fault location", "influencing parameters", and "maintenance measures".

[0079] Knowledge Extraction and Completion: Combining electrical equipment design manuals, fault repair case libraries, and expert experience, specific knowledge is stored in the graph database in the form of triples (entity-relationship-entity). For example, (IGBT overvoltage breakdown leads to drive circuit damage); (DC bus capacitor bulging manifests as increased voltage ripple); (insulation aging, the cause is over-temperature operation), etc. This knowledge graph contains the physical mechanisms and causal chains of electrical fault occurrence and evolution.

[0080] GNN-based fault state reasoning: The joint feature tensor F2 of the current device obtained in step 2 is mapped to the nodes of the knowledge graph as the initial state of the corresponding fault phenomenon node.

[0081] The mapping process from the joint feature tensor F2 to knowledge graph nodes is specifically implemented through a "semantic attention allocation layer":

[0082] First, predefine the set of all "fault phenomena" nodes in the knowledge graph. ,in This represents the total number of phenomenon nodes. For each phenomenon node... It generates corresponding semantic prototype vectors by combining keywords defined by experts and using a pre-trained word vector model. For example, for the node "voltage ripple increases", its prototype vector is obtained by weighted averaging of the word vectors of the three words "voltage", "ripple", and "increase".

[0083] Then, the joint feature tensor F2 (a 256-dimensional vector) output from step two is input into a single-layer neural network, and its attention score with the prototype vector of each phenomenon node is calculated:

[0084] ;

[0085] in Let be the initial confidence score (attention score) for the i-th fault phenomenon node. , This is an index for the faulty node. This represents the total number of nodes exhibiting fault symptoms. It is an exponential function. For the first The matching score of each node. , and The projection matrix is ​​a trainable matrix (both with dimensions 128×256). Through training, the F2 and prototype vectors are projected into the same semantic space and similarity matching is performed.

[0086] Finally, each phenomenon node Initial confidence in graph neural networks That is, the calculated attention score. The value ranges from [0,1] and the sum of the scores of all phenomenon nodes is 1. For other types of nodes in the graph, such as "fault cause" and "fault location", their initial confidence is uniformly set to 0, waiting for subsequent GNN message passing to update. In this way, the continuous feature vector F2 is successfully decoupled and discretized and assigned to the corresponding symbolic nodes in the graph, thereby starting the subsequent graph inference process.

[0087] Graph neural networks (such as GCN or GAT) are used to propagate and aggregate information on the graph. GNNs can update the confidence of each relevant node (fault cause, fault location) along the edges of "causal" and "influence" relationships.

[0088] For example, when the confidence of the two phenomenon nodes "increased voltage ripple" and "IGBT case temperature too high" is activated by F2, GNN will significantly increase the confidence of their common parent nodes "DC bus capacitor aging" and "IGBT drive signal abnormality" through graph propagation, and calculate the correlation probability between the two.

[0089] Generate interpretable diagnostic reports: The graph neural network ultimately outputs one or more of the most likely causes of failure, along with their confidence levels;

[0090] Interpretable Generation: Not only does it output "Fault Cause: DC bus capacitor aging (confidence 95%)", but more importantly, it can trace back the GNN's inference path to generate a clear and interpretable logical chain. For example: "Diagnostic Reasoning: ① At time t1, the equipment is in a constant-speed operating condition (from the control flow), but the electrical characteristics (step one) detect an abnormal increase in bispectral energy at 150Hz -> ② The fusion model determines that the abnormality is not caused by the operating condition and confirms it as the phenomenon of 'increased voltage ripple' (step two) -> ③ According to the knowledge graph, 'increased voltage ripple' is usually caused by 'capacitor aging' or 'load imbalance', and the 'IGBT case temperature' feature shows an abnormal temperature rise rate, further pointing to capacitor aging -> ④ After comprehensive calculation by the graph neural network, the root cause is inferred to be 'increased equivalent series resistance of the DC bus capacitor' (aging)."

[0091] Finally, the diagnostic conclusions, confidence levels, complete reasoning paths, and recommended repair solutions (such as "recommend replacing the DC bus capacitor bank") are displayed to the maintenance personnel.

[0092] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0093] 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 programs 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 or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0094] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0097] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0098] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations 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. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A fault diagnosis method for electrical automation equipment based on artificial intelligence, characterized in that, Includes the following steps: Multi-source electrical signals are acquired and preprocessed. The feature vector F1 of early micro-weak electrical faults is extracted by adaptive morphological filtering and high-order spectral analysis. The operating condition data, control commands, and static semantic information of the collected equipment are spatiotemporally aligned with the feature vector F1, and multimodal fusion is performed through a Transformer encoder to obtain the joint feature tensor F2. A fault mechanism knowledge graph containing fault phenomena, fault causes, fault location entities and their causal relationships is constructed. The joint feature tensor F2 is mapped to the graph nodes, and causal reasoning is performed using a graph neural network to output interpretable diagnostic results.

2. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 1, characterized in that, The specific steps for extracting the early minor electrical fault feature vector F1 include: The original three-phase voltage and current signals are synchronously acquired using high-frequency sensors; The particle swarm optimization algorithm is used to adaptively select the optimal structuring element type and length, and morphological opening-closing and closing-opening combination operations are performed on the original signal to suppress noise. The filtered signal is subjected to segmented bispectral analysis to extract bispectral energy peaks, frequency pairs, and diagonal slice entropy as preliminary features. The preliminary features are screened using Relief-F importance assessment, and then dimensionality is reduced by principal component analysis to obtain a low-dimensional, highly discriminative feature vector F1.

3. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 2, characterized in that, The particle swarm optimization algorithm specifically includes: With the goal of maximizing the signal-to-noise ratio of the filtered signal, each particle is encoded as a two-dimensional vector of structuring element type and length. The particle velocity and position are updated iteratively until the maximum number of iterations or the fitness value converges, and the globally optimal combination of structuring elements is output.

4. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 2, characterized in that, In the Relief-F importance assessment steps: Calculate the weight of each feature in distinguishing different fault categories, and retain the features with the highest weight ranking; in the principal component analysis step, select the principal components whose cumulative variance contribution rate reaches a preset threshold, and project the original features onto the principal component direction to obtain the dimensionality-reduced feature vector F1.

5. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 1, characterized in that, The multimodal fusion specifically includes: Align the feature vector F1, the real-time read control commands and operating parameters, and the static semantic information obtained from the equipment ledger on the time axis; The aligned data is constructed into a sequence with fixed time windows, where electrical features and operating parameters are arranged by time step, and static semantic features are copied as global tokens and concatenated to each time step. A sinusoidal positional code is added to the sequence, and then input into a Transformer encoder. Cross-modal associations are learned through a self-attention mechanism, and a joint feature tensor F2 is output.

6. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 5, characterized in that, The sinusoidal position code is generated in the following manner: The sine and cosine function values ​​are calculated based on the time step position and dimension index to form a position encoding matrix, which is then added element by element to the input token sequence.

7. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 1, characterized in that, The steps for constructing the fault mechanism knowledge graph include: Define the entity types of electrical faults, fault phenomena, fault causes, fault locations, influencing parameters, and maintenance measures; Knowledge is extracted from electrical equipment design manuals, fault repair case libraries, and expert experience, and the causal and influence relationships between entities are stored in the form of triples.

8. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 1, characterized in that, The steps of mapping the joint feature tensor F2 to graph nodes include: Predefine a semantic prototype vector for each fault phenomenon node; The F2 and prototype vectors are mapped to the same semantic space using a trainable projection matrix, and the attention score is calculated as the initial confidence of each phenomenon node. Set the initial confidence level of other types of nodes in the graph to zero.

9. The method for fault diagnosis of electrical automation equipment based on artificial intelligence according to claim 1, characterized in that, The steps for causal reasoning performed by the graph neural network include: Using the initial confidence level of each node as input, information is propagated and aggregated along the causal edges of the knowledge graph to update the confidence levels of the fault cause and fault location nodes; Output one or more fault causes with the highest confidence level, and trace back the message passing path of the graph neural network to generate an interpretable diagnostic report containing the inference logic chain and recommended maintenance solutions.

Images