Adaptive fault diagnosis method and system for high-low voltage switch cabinet based on ai multi-modal fusion

Abstract

The application relates to the technical field of switch cabinet fault diagnosis, in particular to a high-low voltage switch cabinet self-adaptive fault diagnosis method and system based on AI multi-modal fusion, which unifies various monitoring parameters to a common time reference through hierarchical timestamp coding, and solves the problem that different sampling rate data cannot be fused. A Granger causality check is used to eliminate statistical false correlation caused by environmental common causes, a causal correlation matrix is constructed to represent the real causal relationship between monitoring parameters under a fault mode, a bias term is injected into a Transformer self-attention mechanism to guide the model to preferentially focus on mode combinations with physical causal relationships, and false positives caused by a pure data-driven model due to working condition changes are avoided. In addition, a relaxed physical regularization loss term is introduced to constrain model training, so that the model output does not violate basic principles such as the Joule law and mechanical dynamics, and the reliability of the diagnosis result is enhanced.

Description

Technical Field

[0001] This application relates to the field of switchgear fault diagnosis technology, specifically to an adaptive fault diagnosis method and system for high and low voltage switchgear based on AI multimodal fusion. Background Technology

[0002] High- and low-voltage switchgear are the most numerous and widely distributed critical equipment in power systems, and their operating status directly affects the safety of the power grid. Switchgear has a compact internal structure and a complex electromagnetic environment. Common faults include contact overheating, partial discharge, jamming of mechanical operating mechanisms, and insulation deterioration. If any fault is not dealt with in time, it may lead to equipment damage or even a large-scale power outage.

[0003] Currently, switchgear condition monitoring mainly relies on data from single-type sensors, such as infrared thermal imagers for temperature monitoring, ultrasonic sensors for partial discharge detection, and ultra-high frequency sensors for capturing electromagnetic wave signals. However, single-mode information is insufficient to comprehensively reflect the true health status of the equipment. Different fault types exhibit varying intensity of characterization under different modes, and single sensors are highly susceptible to electromagnetic interference and environmental noise, leading to false alarms or missed alarms.

[0004] In recent years, multi-sensor fusion technology has attracted attention in power equipment diagnosis. Existing methods mostly employ purely data-driven deep learning models, directly concatenating multi-source data before inputting it into a neural network. However, this method suffers from two prominent problems:

[0005] First, there is insufficient utilization of prior physical knowledge. Purely data-driven models learn the correlation between modes from scratch, which not only requires a large number of labeled fault samples, but also easily misjudges statistical correlations caused by environmental factors (such as the synchronous increase of ambient temperature and load current in summer) as fault correlations, leading to an increased false alarm rate when operating conditions change. At the same time, the model output may violate basic electrical principles such as Joule's law, lacking physical consistency guarantees.

[0006] Second, the diagnostic results lack interpretability. Deep learning models are essentially "black boxes," only outputting fault labels or probabilities, and cannot provide decision-making basis or evidence chains. Power industry operation and maintenance decisions require rigorous basis, and operation and maintenance personnel are unlikely to trust and implement AI diagnostic suggestions with unclear reasons, which seriously restricts the practical promotion and application of the technology. Summary of the Invention

[0007] In view of this, the purpose of this application is to provide an adaptive fault diagnosis method and system for high and low voltage switchgear based on AI multimodal fusion, so as to solve the problems in the background art.

[0008] To achieve the above objectives, this application adopts the following technical solution:

[0009] The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion proposed in this application includes the following steps:

[0010] Acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points;

[0011] The time sequence of each monitoring parameter in the multimodal monitoring data is aligned based on hierarchical timestamp encoding to obtain aligned multimodal monitoring data;

[0012] Historical fault sample data is filtered from the aligned multimodal monitoring data, and a causal correlation matrix between various fault modes and monitoring parameters is constructed based on the historical fault sample data. The causal correlation matrix represents the causal relationship between monitoring parameters under various fault modes.

[0013] The causal correlation matrix is ​​injected as a bias term into the multi-head self-attention mechanism of the Transformer network, and the Transformer network fused with the causal correlation matrix is ​​trained based on the aligned multimodal monitoring data and the pre-constructed loss function with physical regularization loss term to obtain the fault identification model.

[0014] Based on the fault identification model, the multimodal monitoring data generated by the high and low voltage switchgear during the current operation are identified to obtain the fault identification result; and multi-head self-attention weights are extracted from the fault identification model, and the root cause is located based on the multi-head self-attention weights to obtain the root cause parameters leading to the fault.

[0015] This application also provides an adaptive fault diagnosis system for high and low voltage switchgear based on AI multimodal fusion, including:

[0016] The acquisition module is used to acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points;

[0017] The timing alignment module is used to perform timing alignment on each monitoring parameter in the multimodal monitoring data based on hierarchical timestamp encoding to obtain aligned multimodal monitoring data;

[0018] The causal association module is used to filter historical fault sample data from the aligned multimodal monitoring data, and construct a causal association matrix between various fault modes and monitoring parameters based on the historical fault sample data. The causal association matrix represents the causal relationship between monitoring parameters under various fault modes.

[0019] The model building module is used to inject the causal correlation matrix as a bias term into the multi-head self-attention mechanism of the Transformer network, and to train the Transformer network that fuses the causal correlation matrix based on the aligned multimodal monitoring data and the pre-built loss function that introduces a physical regularization loss term, so as to obtain a fault identification model.

[0020] The fault and root cause identification module is used to identify the multimodal monitoring data generated by the high and low voltage switchgear during the current operation based on the fault identification model, and obtain the fault identification result; extract the multi-head self-attention weight from the fault identification model, and perform root cause localization on the fault identification result based on the multi-head self-attention weight to obtain the root cause parameters leading to the fault.

[0021] The beneficial effects of this application are as follows: The adaptive fault diagnosis method and system for high and low voltage switchgear based on AI multimodal fusion solves the problem of difficulty in fusing data with different sampling rates by unifying various monitoring parameters to a common time base through hierarchical timestamp encoding. Granger causality verification is used to eliminate statistical spurious correlations caused by environmental common factors, and a causal correlation matrix is ​​constructed to represent the true causal relationship between monitoring parameters under fault modes. Furthermore, a bias term is injected into the Transformer self-attention mechanism to guide the model to prioritize mode combinations with physical causal relationships, avoiding false alarms caused by changes in operating conditions in purely data-driven models. In addition, a relaxed physical regularization loss term is introduced to constrain model training, ensuring that the model output does not violate basic principles such as Joule's law and mechanical dynamics, thus enhancing the reliability of the diagnostic results. Attached Figure Description

[0022] The present application will be further described below with reference to the accompanying drawings and embodiments:

[0023] Figure 1 This is a flowchart illustrating an adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion in one embodiment of this application;

[0024] Figure 2 This is a schematic diagram of the model training phase in one embodiment of this application;

[0025] Figure 3 This is a flowchart illustrating the fault diagnosis stage in one embodiment of this application;

[0026] Figure 4 This is a logical structure diagram of an adaptive fault diagnosis system for high and low voltage switchgear based on AI multimodal fusion, as shown in one embodiment of this application. Detailed Implementation

[0027] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the layers related to this application and are not drawn according to the actual number, shape and size of the layers in the actual implementation. In the actual implementation, the form, number and proportion of each layer can be arbitrarily changed, and the layer layout may also be more complex.

[0029] Numerous details are explored in the following description to provide a more thorough explanation of embodiments of this application; however, it will be apparent to those skilled in the art that embodiments of this application may be practiced without these specific details.

[0030] Figure 1 This is a flowchart illustrating an adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion, as shown in one embodiment of this application. Figure 1 As shown, the adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion in this application mainly includes the following steps:

[0031] S110, multi-source heterogeneous data acquisition;

[0032] S120, heterogeneous timing alignment;

[0033] S130, Prior Knowledge Extraction;

[0034] S140, Knowledge Injection Training;

[0035] S150, Interpretable Diagnosis;

[0036] S160, Physical rule verification;

[0037] The specific execution method and principle of each step are explained below.

[0038] Figure 2 This is a schematic diagram of the model training phase in one embodiment of this application, as shown below. Figure 2 As shown, the model training phase includes S110-S140, which includes:

[0039] S110, multi-source heterogeneous data acquisition;

[0040] Acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points;

[0041] Multimodal monitoring data refers to monitoring data collected from switchgear using various types of sensors, reflecting different physical dimensions. These various monitoring parameters include, but are not limited to:

[0042] Electrical quantity monitoring parameters: RMS value of three-phase current, RMS value of three-phase voltage, zero-sequence current, and current harmonic distortion rate.

[0043] Thermal monitoring parameters: temperature of switchgear contact area, temperature of busbar connection point, and surface temperature of cabinet.

[0044] Vibration / acoustic monitoring parameters: cabinet surface vibration acceleration, ultrasonic signal amplitude, and vibration waveform when the operating mechanism is activated.

[0045] Partial discharge monitoring parameters: UHF partial discharge signal, transient ground voltage (TEV) signal, and partial discharge pulse phase.

[0046] Environmental monitoring parameters: ambient temperature, ambient humidity, and concentration of characteristic gases inside the switchgear.

[0047] Among them, the sampling rate of current / voltage signals is usually at the kHz level, vibration / ultrasonic signals at the MHz level, and temperature signals at the Hz level. The sampling frequencies and time scales of various parameters differ significantly, and they need to be time-aligned by subsequent layered timestamp encoding before they can be fused.

[0048] S120, heterogeneous timing alignment;

[0049] The sampling rates of various sensors differ significantly (from kHz to MHz), and the time bases are not unified. Directly splicing these sensors can lead to misjudgments of the temporal relationships between cross-modal events in the model, making it impossible to accurately construct a causal correlation matrix. Therefore, time sequence alignment is necessary to lay the foundation for subsequent fault diagnosis.

[0050] Based on hierarchical timestamp encoding, each monitoring parameter in the multimodal monitoring data is time-series aligned to obtain aligned multimodal monitoring data, specifically including:

[0051] S120-1, Extract the data point sequence for each monitoring parameter from the multimodal monitoring data, wherein the data point sequence includes multiple data points, and each data point includes a value and a timestamp;

[0052] From the multimodal monitoring data, data point sequences are extracted according to the monitoring parameter categories. Each data point is a binary tuple. ,in This is the absolute timestamp recorded by the hardware clock when the data point was acquired. For the value to be taken.

[0053] S120-2 divides various monitoring parameters into high-frequency monitoring parameters and low-frequency monitoring parameters;

[0054] Based on the sampling frequency of each sensor, various monitoring parameters are divided into high-frequency monitoring parameters and low-frequency monitoring parameters, as shown in Table 1:

[0055] Table 1. Examples of high and low frequency monitoring parameter division

[0056]

[0057] S120-3, based on a pre-built sliding window, extracts time-frequency domain statistical features of multiple time windows from the data point sequence of high-frequency monitoring parameters to obtain token sequences corresponding to multiple time windows;

[0058] For high-frequency monitoring parameters, a sliding window strategy is adopted: a fixed-length time window slides across the data point sequence. The original sampling points within each window are not directly retained; instead, their time-frequency domain statistical features are extracted, and each window is compressed into a feature vector (i.e., a token). Let the window length be... The sliding step size is The original N sampling points were compressed into Tokens.

[0059] Fault information in high-frequency signals is often reflected in shifts in the statistical distribution (such as increased vibration amplitude) or changes in the spectral structure (such as increased energy in a specific frequency band), rather than the absolute values ​​of a single sampling point. Therefore, the extracted time-frequency domain statistical features include, but are not limited to:

[0060] Time-domain characteristics: mean, variance, peak value, RMS value, peak-to-peak value.

[0061] Frequency domain characteristics: spectral centroid, frequency band energy proportion (after wavelet packet decomposition or FFT).

[0062] Envelope characteristics: peak value of the envelope spectrum.

[0063] By compressing high-frequency raw sampling points into low-dimensional feature tokens, the length of the Transformer's input sequence is significantly reduced, enabling the model to handle heterogeneous data ranging from Hz to MHz.

[0064] S120-4, take the value of each time point in the data point sequence of low frequency monitoring parameters as a token to obtain a token sequence corresponding to multiple time points;

[0065] For low-frequency monitoring parameters, no sliding window processing is required. The sampled values ​​at each time point are normalized (e.g., Min-Max normalization or Z-score normalization) and directly used as a feature vector (token) to obtain the token sequence corresponding to multiple time points.

[0066] S120-5, convert the token sequence into a hierarchical timestamp code based on an independent linear projection layer, wherein the mathematical expression of the hierarchical timestamp code is:

[0067]

[0068] In the formula, Indicates layered time coding. To collect timestamps, For parameter index;

[0069] This is the global absolute time, usually converted to a Unix timestamp (the number of seconds since January 1, 1970, such as 1785613937.123).

[0070] For parameters A specific timescale location indicates that the token is in a modal context. Internal time location;

[0071] For sinusoidal position coding functions, scalar time values ​​are mapped to high-dimensional vectors;

[0072] The global time projection matrix, For parameters A dedicated time projection matrix; and Two independent, learnable projection matrices are used to encode "global absolute time" and "modal intra-temporal temporality," respectively. These have distinct semantics: the former answers "When did this happen?" (e.g., August 18, 2024, 14:32), while the latter answers "On this sensor's own timeline, which point / window is this?" (e.g., the 123456th sampling point). By using these independent projection matrices, the model learns the difference between these two temporal semantics, rather than conflating them.

[0073] In the above calculation formula, By enabling all modalities to share the same time scale, the model can determine the absolute chronological order of events, which is the mathematical basis for Granger causality tests to determine that the cause precedes the effect. as well as It preserves the unique temporal granularity of each modality, enabling the model to understand the essential physical and semantic differences between the microsecond-level rhythm of high-frequency signals and the second-level rhythm of low-frequency signals.

[0074] S120-6 uses hierarchical timestamp encoding based on multiple monitoring parameters to perform time-series alignment, resulting in aligned multimodal monitoring data.

[0075] All modal (monitoring parameter) tokens carry After encoding, the data is arranged and stored in a unified vector space to form aligned multimodal monitoring data.

[0076] S130, Prior Knowledge Extraction;

[0077] The purpose of extracting the causal correlation matrix between monitoring parameters from historical fault data is to provide prior knowledge for the Transformer, enabling it to prioritize mode combinations with physical causal relationships during self-attention computation. This avoids purely data-driven models misclassifying environmental common causes as fault correlations and provides interpretability for subsequent diagnostic results. The core idea is to use the fault occurrence time as an anchor point to extract the fault evolution window, quantify the correlation strength between parameters using the Pearson correlation coefficient, then use the Granger causality test to screen for truly causal pairs, and finally fuse them using the Hadamard product to obtain the causal correlation matrix, which serves as prior knowledge injected into the Transformer.

[0078] Historical fault sample data is filtered from the aligned multimodal monitoring data, and a causal correlation matrix between various fault modes and monitoring parameters is constructed based on the historical fault sample data. The causal correlation matrix represents the causal relationship between monitoring parameters under various fault modes.

[0079] The process of constructing a causal correlation matrix includes:

[0080] S130-1, Obtain multiple fault occurrence times;

[0081] Obtain the precise time point of each failure from historical failure records. The sources of the time of failure include:

[0082] Protection action records (such as circuit breaker tripping time);

[0083] Online monitoring system alarm time (such as infrared temperature over-limit alarm);

[0084] The moment when the amplitude of the partial discharge pulse suddenly increases;

[0085] The time when the fault was discovered, as recorded in the maintenance and repair log.

[0086] S130-2, Construct a data extraction window based on the time point of the fault occurrence, and filter historical fault sample data from the aligned multimodal monitoring data based on the data extraction window;

[0087] At each time point of failure Use anchor points to construct data extraction windows .in, This refers to the length of time before a fault occurs (capturing precursor features). This refers to the time elapsed after a fault (capturing fault response characteristics). The window length is configured differently based on the fault type, as shown in Table 2.

[0088] Table 2. Examples of Window Length Differentiation Configurations

[0089]

[0090] For each type of fault All samples belonging to this type of fault use the same window length configuration.

[0091] S130-3, Extract the fault sample value sequence of each monitoring parameter from the historical fault sample data;

[0092] Specifically, within the data window captured by S130-2, for each monitoring parameter Extract its values ​​at all time points within the window to form a fault sample value sequence for that parameter. There are a total of Various monitoring parameters were obtained. A sequence of values.

[0093] S130-4, use any two monitoring parameters as a parameter group;

[0094] from Choose any two from the monitoring parameters. and The parameter group is composed of... There are 10 parameter sets, each corresponding to a set of two sequences to be analyzed. .

[0095] S130-5, Calculate the Pearson correlation between the fault sample value sequences of any two monitoring parameters in each parameter group. And based on the Pearson correlation of all parameter groups Construct a correlation coefficient matrix and normalize the correlation coefficient matrix to obtain the degree of correlation matrix. ,in, and This indicates two different monitoring parameters. Fault mode type;

[0096] Pearson correlation The mathematical expression is:

[0097]

[0098] In the formula, For monitoring parameters In the fault The mean of all samples below, For monitoring parameters In the fault Next The values ​​of each sample; For monitoring parameters In the fault The mean of all samples below, For monitoring parameters In the fault Next The values ​​of each sample.

[0099] Organize the correlation coefficients of all parameter groups as follows The matrix is ​​then normalized (so that all elements are in the interval [0,1]) to obtain the correlation matrix, whose elements are... Quantified the fault Monitoring parameters and The strength of the association.

[0100] S130-6, Granger causality check is performed based on the fault sample value sequence of any two monitoring parameters in each parameter group to obtain the causality check flag bit for each parameter group; and a causality check mask matrix is ​​constructed based on the causality check flag bits of all parameter groups. Wherein, a causal verification flag of 1 indicates the existence of a causal relationship, and a causal verification flag of 0 indicates the absence of a causal relationship;

[0101] Specifically, for each parameter group Granger causality test was used to determine the monitoring parameters. Does historical information affect monitoring parameters? The predictions are statistically significant. This test is based on the following regression model:

[0102]

[0103] In the formula, For monitoring parameters exist The value at time, and All are lagging indexes. The lag order is... and All are regression coefficients. For monitoring parameters exist The value at time, For monitoring parameters exist The value at time, This is the residual term.

[0104] Establish the F-statistic to test the null hypothesis. (Right now Historical information (No predictive ability). The lag order is determined using the Bayesian information criterion. .

[0105] At the same time, considering the When performing pairwise tests on different modalities, multiple comparisons can occur. To address this, Bonferroni correction is introduced, with the significance level set at:

[0106]

[0107] In the formula, This represents the significance probability value of the Granger causality test.

[0108] Determine only if the correction condition is met. right Granger causality exists. Each parameter group... The verification results are organized as follows Causality check mask matrix Its elements This indicates the existence of a causal relationship. This indicates that it does not exist.

[0109] S130-7, Based on the aforementioned correlation matrix and the causal check mask matrix Calculate the causal relationship matrix , In the formula, It is the product of Hadama.

[0110] Pearson correlation identifies "parameters that change together," while Granger causality identifies "parameters with time-series predictive relationships." Only association pairs that simultaneously satisfy both statistical correlation and causality are considered to have physically meaningful fault causal relationships. The Hadamard product is a mathematical tool for performing this dual screening.

[0111] The above steps integrate the dual screening of correlation strength and causal validity into a unified causal correlation matrix, providing high-quality and highly reliable prior knowledge for the subsequent step S140 of injecting the Transformer.

[0112] S140, Knowledge Injection Training;

[0113] The core role of knowledge-injected training is to integrate the causal prior knowledge extracted from historical fault data into the model, guide the Transformer's multi-head self-attention mechanism to prioritize modal combinations with physical causal relationships, avoid the pure data-driven model from misjudging environmental common causes as fault correlations, and at the same time constrain the model output to not violate basic physical laws through physical regularization loss, thereby improving diagnostic accuracy and reliability.

[0114] The core idea is to inject causal relationships into the self-attention score in the form of bias terms, so that the model prioritizes modal combinations with causal relationships when calculating attention; at the same time, the training process is constrained by physical regularization loss to ensure that the model output does not violate basic physical laws, and finally a fault recognition model with both data-driven learning ability and physical knowledge constraints is trained.

[0115] Specifically, the causal correlation matrix is ​​injected as a bias term into the multi-head self-attention mechanism of the Transformer network, and the Transformer network fused with the causal correlation matrix is ​​trained based on the aligned multimodal monitoring data and a pre-constructed loss function that incorporates a physical regularization loss term to obtain a fault identification model, including:

[0116] S140-1, Configure bias terms based on the causal correlation matrix. , wherein the bias term The mathematical expression is:

[0117]

[0118] In the formula, Monitoring parameters in the causal correlation matrix With monitoring parameters The correlation matrix, The set anti-crash value;

[0119] Specifically, The function will The correlation weight mapping of the interval is as follows The bias value of the interval, and the attention score Terms of the same numerical order of magnitude facilitate addition. Positions without causal relationship. Take directly In the subsequent softmax iteration, the probability at this position approaches 0, which is equivalent to masking. Using... Cut off, avoid This can cause numerical overflow.

[0120] S140-2, the bias term By introducing the multi-head self-attention score from the Transformer network, we obtain the updated multi-head self-attention score. Among them, the updated multi-head self-attention score The mathematical expression is:

[0121]

[0122]

[0123] In the formula, For monitoring parameters The query vector of the token. For monitoring parameters The key vector of the token, To query the dimensions of a vector / key vector, This is the prior confidence coefficient. As a time distance factor, For monitoring parameters The original collection timestamp, For monitoring parameters The original collection timestamp, This represents the maximum time delay of physical signal propagation.

[0124] The self-attention score mentioned above is composed of three superimposed components, each serving a different function:

[0125] The first item is a data-driven item. It is a data-driven similarity score, which is learned autonomously by the model from the current input data. It is used to capture the similarity between modes in the current input data and has flexibility.

[0126] The second item is a knowledge-guided item. Knowledge-driven causal bias The learnable confidence coefficient (initial value 0.5) controls the guiding strength of prior knowledge. By injecting causal priors mined from historical data, the model obtains the correct association direction in the early stages of training, avoiding learning from scratch and reducing dependence on a large number of labeled samples.

[0127] The third item is a timing penalty item. As a penalty for the reasonableness of the timing, Calculations based on the original acquisition timestamps of the two tokens, when the time difference exceeds the maximum physical propagation delay. It generates a negative penalty. It is used to block attention calculations between tokens that are too far apart in time, and to eliminate meaningless cross-temporal correlations from a physical mechanism perspective (such as making a cross-temporal correlation between the current current and the temperature 5 seconds ago).

[0128] The combination of these three factors allows the attention score to be driven by data similarity, causal priors, and physical time sequence simultaneously, resulting in stronger physical consistency and sample efficiency compared to purely data-driven models.

[0129] S140-3, based on the updated multi-head self-attention score Update the multi-head self-attention mechanism.

[0130] Specifically, the updated attention score calculated by S140-2 is... Normalized to attention weight distribution using the softmax function and weighted aggregate value vector :

[0131]

[0132]

[0133] In the formula, For multi-head self-attention score The matrix, This indicates multi-head self-attention output.

[0134] Meanwhile, a causal mask matrix constructed using S130 Generate sparse attention masks and skip during training and inference. The position of the causally significant modal pairs will restrict the computation to the position of the causally significant modal pairs.

[0135] Softmax will score Mapped to probability distribution Modal pairs with higher scores receive higher attention weights. Through engineering optimization using causal masks, computations on non-causally related modal pairs are skipped directly, reducing unnecessary computational overhead. The standard Transformer's self-attention architecture is retained, with only the score calculation method modified, resulting in a simple implementation and strong compatibility.

[0136] The training process for the Transformer network based on the aforementioned multi-head self-attention mechanism includes:

[0137] S140-4, Based on multiple historical fault time points, extract fault sample segments from the aligned multimodal monitoring data, and extract normal sample segments from the aligned multimodal monitoring data;

[0138] Fault diagnosis is essentially a supervised classification task, requiring both positive (fault) and negative (normal) samples for the model to learn the discrimination boundary. Differentiated window configuration ensures that the key evolutionary stages of different fault types are fully captured. Temporal partitioning avoids the same degradation process of the same device appearing simultaneously in the training and test sets, validating the model's generalization ability to unseen faults.

[0139] Specifically, multiple historical fault time points obtained using S130-1 As the anchor point, the data extraction window is configured according to S130-2. Extract fault sample segments. At the same time, randomly extract time segments of the same length as the fault samples from the normal operation period of the equipment (without any alarm records) as normal samples.

[0140] The ratio of positive to negative samples should be controlled between 1:3 and 1:5 to prevent the model from being biased towards predicting fault-free scenarios due to an excessive number of normal samples (class imbalance). Data should be divided according to device number and fault occurrence time (e.g., training with data from the first 3 years and testing with data from the last year) to avoid inflated generalization ability caused by leakage of data from the same source.

[0141] S140-5, Tokenize and align the faulty sample fragment and the normal sample fragment to obtain a data sample;

[0142] Specifically, following the tokenization strategies of S120-3 and S120-4, the fault sample segments and normal sample segments extracted in S140-4 are processed: high-frequency signals are extracted using a sliding window to extract time-frequency domain statistical features and compressed into token sequences; low-frequency signals are directly normalized as tokens. Subsequently, each token is generated according to the hierarchical timestamp encoding method of S120-5. Encode and add.

[0143] Sequence length of all samples The number of tokens is standardized to a fixed value (e.g., 512 tokens). Samples that are too short are zero-paddinged, and the padding mask is recorded to ensure that the Transformer does not include the padding positions in its attention calculation. This generates standardized data samples with uniform format and length, meeting the requirements for batch training of the Transformer.

[0144] S140-6, The data samples are labeled to obtain training samples with label vectors, wherein the label vectors include flag bits for the determination results of various fault types;

[0145] Data normalization has already been implemented during tokenization. However, this normalization involves globally standardizing the token feature vectors of all modalities using Z-score, and employing statistics (mean). and standard deviation The features of each modality are scaled uniformly. Then the data is labeled to generate label vectors. , This represents the total number of fault types. If a sample belongs to a fault... ,but Multiple labels can be set to 1 simultaneously (multi-label annotation).

[0146] For fault samples: based on their fault type, the corresponding Set all values ​​to 1, and the rest to 0. For normal samples: all All are 0.

[0147] S140-7, Input the training samples into the Transformer network that fuses the causal correlation matrix to obtain the prediction results;

[0148] The training samples (input vectors) generated in S140-6 The input is fed into a Transformer network that incorporates a causal correlation matrix, and passes through the following layers in sequence:

[0149] Embedding layer (input vector + layered timestamp encoding) → Multi-head self-attention layer (using S140-2 updated) (Calculation method) → Feedforward neural network → Layer normalization → Repeated multi-layer encoding → Fully connected classification layer → Sigmoid activation function, finally obtaining the prediction result. ,in The independent predicted probabilities for each fault type.

[0150] S140-8, Based on the pre-constructed loss function that introduces a physical regularization loss term, calculate the physical constraint loss of the data sample, and calculate the prediction loss of the prediction result and the label vector to obtain the total loss;

[0151] The mathematical expression for the loss function that introduces the physical regularization loss term is:

[0152]

[0153]

[0154] In the formula, For the total loss, For binary cross-entropy loss, For the first The loss value of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, Loss constrained by Joule's law, For mechanical vibration constraint loss, This is the partial discharge constraint loss.

[0155] (Binary cross-entropy loss): measures the probability of prediction With real labels The difference between them is the standard classification loss. Its gradient-driven model improves the accuracy of fault identification.

[0156] (Physical Regularization Constraint Loss): Calculated based on sensor measurements in the current batch, this loss incurs a high penalty when the sample itself violates physical laws. Its gradient backpropagation forces the model to weaken "anti-physical" noise inputs during the feature extraction stage, indirectly guiding the model's internal representation to optimize in a direction that conforms to physical laws.

[0157] The mathematical expressions for the losses due to the three physical constraints are as follows:

[0158]

[0159]

[0160]

[0161] In the formula, This refers to the batch sample size. The change in temperature For the time change, This is the effective value of the current. Joule heat coefficient, This is a heat dissipation correction item. This is the inherent error tolerance band of Joule's law. To measure the vibration amplitude, For operating force or driving force, Indicates a given operating force or driving force The maximum vibration amplitude at that time, The inherent error tolerance band of vibration constraint, To measure the apparent discharge quantity, To apply voltage, To apply a given voltage Theoretical discharge amount at that time This is the inherent error tolerance band for partial discharge constraint.

[0162] The inputs to the three physical constraint losses mentioned above are the raw sensor measurements of each sample in the current batch, not the model predictions. These measured values ​​are directly extracted from the input data of the training samples and substituted into the corresponding physical formulas to calculate the bias. The physical meaning is: if there is a sample in the current batch that "violates the laws of physics" (e.g., extremely low current but extremely high temperature rise), then the bias of that sample... This results in a high penalty. The gradient propagates back through this loss, forcing the model to learn to weaken these "anti-physical" noise inputs during the feature extraction stage, thereby indirectly guiding the model's internal representation to optimize in a direction that conforms to physical laws.

[0163] The three physical constraint losses correspond to the physical meanings shown in Table 3, and are applicable to the scenarios shown in Table 3:

[0164] Table 3. Applicable Scenarios and Physical Meanings of the Three Physical Constraint Losses

[0165]

[0166] Each constraint uses a Huber-style relaxation loss method. ,in The inherent error tolerance band is determined statistically through the fluctuation range of historical normal operating data. When the physical deviation is within... Within a certain range, fluctuations caused by measurement noise and thermal inertia are considered normal and no penalty is imposed; only when the deviation exceeds a certain range... Only when the limit is exceeded will a linear penalty be applied.

[0167] S140-9, Propagate forward based on the total loss to adjust the internal parameters of the Transformer network that fuses the causal correlation matrix;

[0168] Specifically, calculation The gradients of all trainable parameters in the model are propagated layer by layer using the backpropagation algorithm, and the optimizer updates the parameters to reduce the loss value. The AdamW optimizer (more suitable for Transformers) is used, with an initial learning rate of 1×10⁻⁶. −4 A cosine annealing learning rate scheduling strategy is employed. Gradient clipping is also implemented, with a maximum gradient norm of 1.0, to prevent the physical loss from generating large gradients in the early stages of training, which could lead to training instability.

[0169] Trainable parameters include: the parameters of each layer of the Transformer , , Projection matrix; weights of fully connected classification layers; layer normalization parameters; learnable prior confidence coefficients The causal correlation matrix remains fixed during training, does not participate in gradient updates, and serves as invariant prior knowledge.

[0170] S140-10, return to inputting the training samples into the Transformer network that fuses the causal correlation matrix to obtain the prediction results, until training is complete and the fault identification model is obtained.

[0171] Repeat steps S140-7 to S140-9 to complete one epoch of training, then evaluate the model performance on the validation set. Monitor the performance on the validation set. and If the validation loss does not decrease for 15 consecutive epochs, early stopping is triggered, terminating training and rolling back to the model weights of the epoch with the lowest validation loss. After training, the model weights on the validation set are selected. The model with the smallest weight is used as the final fault identification model.

[0172] Figure 3 This is a flowchart illustrating the fault diagnosis stage in one embodiment of this application, as shown below. Figure 3 As shown, the fault diagnosis phase includes S150-S160, which includes:

[0173] S150, Interpretable Diagnosis;

[0174] Based on the fault identification model, multimodal monitoring data generated by high and low voltage switchgear during current operation are identified to obtain fault identification results. Multi-head self-attention weights are extracted from the fault identification model, and root cause localization is performed on the fault identification results based on these multi-head self-attention weights to obtain root cause parameters leading to the fault. Specifically, this includes:

[0175] S150-1, Extract the current data segment from the multimodal monitoring data generated during the current operation;

[0176] Specifically, from the real-time multimodal monitoring data stream generated during the current operation, taking the current moment... Using the right endpoint as the starting point, a fixed-length data segment is truncated forward. The truncated length remains consistent with that of the training phase (S140-4) to match the model's input dimension. The system continuously maintains a dynamic data cache, with new data continuously flowing in and expired data being evicted, achieving streaming data processing.

[0177] S150-2, Tokenize and align the current data segment to obtain aligned data;

[0178] Following the same tokenization strategy as training phase S140-5, the data segment extracted in S150-1 is processed as follows:

[0179] High-frequency monitoring parameters (current, voltage, vibration, etc.): Time-frequency domain statistical features (mean, variance, spectral centroid, wavelet packet band energy, etc.) are extracted using a sliding window and compressed into token sequences;

[0180] Low-frequency monitoring parameters (infrared temperature, ambient humidity, etc.): directly take the value as the token.

[0181] Subsequently, each token is generated according to the hierarchical timestamp encoding method of S120-5. Encode the sequence and add it element-wise to the content feature vector of the token. If the sequence length is insufficient, zero-padding is performed, and the padding mask is recorded.

[0182] S150-3, Normalize the aligned data to obtain the input vector;

[0183] The tokenized data obtained in S150-2 is Z-score standardized using the normalization parameters (mean and standard deviation of each feature) saved in the training phase S140-6 to uniformly eliminate the dimensional differences between the monitoring parameters.

[0184] S150-4, Input the input vector into the fault identification model to obtain a fault identification vector, wherein the fault identification vector includes the probabilities of multiple fault types;

[0185] The input vector generated by S150-3 The input is fed into the fault identification model, and a forward propagation computation is performed. During this process, all network parameters are frozen (gradient calculation and backpropagation are not performed). After passing through multiple Transformer layers, the fault identification vector is output by a fully connected layer and a Sigmoid activation function. .

[0186] S150-5, the fault type with a probability greater than the preset adaptive threshold is taken as the fault identification result.

[0187] The various fault probabilities output by S150-4 Adaptive thresholds for each category determined in training phase S140-10 Compare them one by one. When If the condition is met, the fault type is determined to exist; otherwise, it is determined not to exist. Finally, all fault types that meet the conditions are output as the fault identification result.

[0188] Thresholds for each category The optimal value of F1 is determined on the validation set using the PR curve.

[0189] S150-6, Extract multi-head self-attention weights from the fault identification model;

[0190] During the forward inference process of the S150-4 model, the attention weight matrix of all attention heads is extracted from the multi-head self-attention mechanism of the last coding layer of the Transformer. The specific value is... The output probability distribution matrix, where Indicates the first The first one in the attention. Attention weights are assigned to each token.

[0191] Multiple attention heads compute attention from different representation subspaces, and the output of each head is retained for subsequent contribution calculation.

[0192] S150-7, Calculate the contribution of each monitoring parameter based on the multi-head self-attention weights. The contribution degree The mathematical expression is:

[0193]

[0194] In the formula, and All are indexes of monitoring parameters. To monitor the total number of parameters, For attention head index, This represents the total number of heads of multi-head attention. For token index, For monitoring parameters The set of tokens, For monitoring parameters The set of tokens, For the first The first one in the attention. Attention weights for each token;

[0195] The molecular part Indicates: for monitoring parameters All tokens corresponding to it All attention head Attention weights All values ​​are summed to obtain the total attention value obtained for this modality.

[0196] denominator Indicates: for all modes The attention weights of all tokens and all attention heads are summed to obtain the total attention value for all modalities (this value is always equal to...). ).

[0197] Dividing the numerator by the denominator gives the normalization, which represents the monitored parameter. The contribution percentage of all modalities in this diagnostic decision-making process. The sum is 1.

[0198] S150-8, based on the contribution of each monitoring parameter The various monitoring parameters are sorted, and the top P monitoring parameters are extracted as the initial root cause parameters.

[0199] The contribution of each monitoring parameter calculated by S150-7 Sort the parameters from largest to smallest and select the top P (TOP-P) monitoring parameters with the highest contribution as the initial root cause parameters (P is generally 3-5). In the diagnostic report, each parameter and its contribution are presented in a list format, with supplementary information such as the measured value of each parameter at the current time and whether it exceeds the limit, for reference by operation and maintenance personnel.

[0200] S160, Physical rule verification;

[0201] This step is an external verification of the S150 fault identification results. The core idea is to use the modal contribution calculated by S150. The AI ​​performs physical rule matching verification on the current real-time monitoring data. It determines that the more important a modality is, the greater the weight of its out-of-limit evidence in the rule verification. The AI ​​probability and rule matching score are then integrated into a comprehensive confidence score. The final diagnosis result is only output when the confidence level reaches a threshold. Through the dual-path collaboration of AI probability path and physical rule path, false alarms from purely data-driven models are effectively avoided, improving the reliability of diagnostic results. Specifically, this includes:

[0202] S160-1, Contribution based on all initial root cause parameters Calculate the rule matching score of the fault identification results. The rule matching score The mathematical expression is:

[0203]

[0204] In the formula, For indicator functions, Fault Corresponding monitoring parameters The statistical early warning threshold, To monitor the value of the parameter at the current moment;

[0205] This reflects whether those modalities deemed important by AI in the current real-time data have actually exceeded physical limits. A higher value indicates that the modality that AI relies heavily on has indeed experienced an anomaly, and the AI's judgment has a physical basis; if A value as low as 0 indicates that the modalities that AI relies on have not exceeded the limits, and that the AI's judgment lacks physical support.

[0206] The indicator function converts the Boolean judgment of whether the limit is exceeded into a numerical value of 0 or 1, which is related to the contribution level. The rule matching score is obtained by weighted summation. The larger the value, the more important the modality is in AI decision-making, and the greater the impact of whether it exceeds the limit on the rule validation results. This allows rule validation to focus on the modalities that AI relies on most, rather than treating all modalities equally.

[0207] S160-2, Matching score based on the rules Calculate the overall confidence level The overall confidence level The mathematical expression is:

[0208]

[0209] In the formula, These are the weighting coefficients for the AI's predicted path. For faults in the fault prediction results The probability, Fault The rule matching score;

[0210] The two paths complement each other through weighted fusion: when the AI ​​probability is high but the rule score is low, the overall confidence is lowered (suppressing false positives); when the AI ​​probability is moderate but the rule score is high, the overall confidence is higher (enhancing credibility). When the rule path has no output (all... The fallback mechanism relies entirely on AI probability to ensure that the system can output effective results under any circumstances.

[0211] S160-3, in the overall confidence level If the confidence level is greater than or equal to the preset confidence threshold, the fault identification result and root cause parameters are saved and sent to the target object; otherwise, the fault identification result and root cause parameters at the current moment are deleted.

[0212] The overall confidence level calculated by S160-2 Compared with the preset confidence threshold (e.g., 0.7) for comparison:

[0213] when The results of fault identification in S150-5, the root cause parameter list and ranking in S150-8, the overall confidence level in S160-2, and the contribution of each modality are combined. Package the data into a structured diagnostic report, save it to a local database, and push it to the operations and maintenance platform or mobile terminal to notify relevant personnel for processing.

[0214] when If the current diagnostic result is deemed insufficiently reliable, the fault identification result and root cause parameters at the current moment are deleted, no alarm is triggered, and subsequent data stream monitoring continues.

[0215] The following case study will fully demonstrate the entire process of fault identification, root cause localization, and physical verification.

[0216] (1) Data acquisition and preprocessing;

[0217] In the A-phase contact area of ​​a 10kV switchgear, a multi-modal monitoring system collects data in real time. At a certain moment, the infrared temperature sensor measured the A-phase contact temperature to be 87°C, the current transformer measured the A-phase current to be 1250A, and the vibration sensor measured the cabinet vibration amplitude to be 0.12m / s². All other parameters were normal.

[0218] The system extracts data segments with the current time as the right endpoint, and after tokenization, hierarchical timestamp encoding and normalization, it is input into the fault identification model for forward inference.

[0219] (2) Fault identification;

[0220] The model outputs a multi-label probability vector through forward inference. The adaptive thresholds for each category have been determined during the validation phase: 0.60 for overheating faults, 0.70 for partial discharges, 0.55 for mechanical jamming, and 0.75 for short-circuit faults.

[0221] The model outputs an overheating fault probability of 0.92, and the probabilities of other faults are all below their respective thresholds, indicating that the current equipment has a contact overheating fault.

[0222] (3) Root cause localization (interpretability);

[0223] Attention weights were extracted from the multi-head self-attention mechanism in the last layer of the model, aggregated according to the type of monitored parameter, and the contribution of each mode was calculated. The results showed that infrared temperature had the highest contribution, reaching 52%, making it the largest root cause; current contributed 31%, ranking second; vibration contributed only 9%, ranking third; and the contributions of the remaining parameters were all below 5%.

[0224] Based on this, it can be determined that the model's overheat diagnosis is primarily based on the abnormal temperature in the A-phase contact area and the A-phase current data, with vibration parameters serving only as supplementary references. This provides maintenance personnel with a clear direction for troubleshooting.

[0225] (4) Physical rule verification;

[0226] Step 1: Calculate the rule matching score;

[0227] The system acquires the measured values ​​of each mode at the current moment. It queries the statistical warning thresholds corresponding to overheating faults (determined by adding 3 times the standard deviation to the historical normal operating conditions): the temperature threshold is 75°C, the current threshold is 1100A, and the vibration threshold is 0.25m / s².

[0228] Among the current measured values, the temperature of 87°C exceeds the 75°C threshold, and the current of 1250A exceeds the 1100A threshold, both triggering over-limit conditions; the vibration of 0.12 m / s² does not exceed the 0.25 m / s² threshold. After weighting the over-limit results with the contribution of each mode, the rule matching score for the overheating fault is 0.83, while the rule scores for the other faults are all 0, and the other parameters do not exceed the limits.

[0229] Step 2: Calculate the overall confidence level;

[0230] The AI ​​probability (92%) and rule matching score were weighted and fused in a 7:3 ratio to calculate the overall confidence level. Since the rule path was effective, the fused overall confidence level was 94.4%, significantly higher than the preset 70% confidence threshold, and the validation passed.

[0231] (5) Output of diagnostic report;

[0232] The system generates a structured diagnostic report:

[0233] Diagnostic conclusion: Phase A contact overheating fault (AI probability 92.0%, overall confidence level 94.4%).

[0234] Root cause parameters: Phase A infrared temperature 87°C (contribution 52%), Phase A current 1250A (contribution 31%), vibration 0.12m / s² (contribution 9%).

[0235] Exceedance verification: Temperature and current exceedances are consistent with AI judgment.

[0236] Maintenance Recommendation: It is recommended to immediately perform infrared retesting on the A-phase contacts to check the contact spring pressure and oxidation of the contact surface. If necessary, arrange for a power outage for maintenance.

[0237] This embodiment covers a complete closed loop of data acquisition, time-series alignment, prior knowledge extraction, knowledge injection training, interpretability diagnosis, and physical rule verification, and has the following beneficial effects:

[0238] Firstly, it eliminates time mismatches in heterogeneous data, laying the foundation for data fusion. Addressing the issue of significant differences in sampling rates (Hz to MHz) among multiple sensors in switchgear, hierarchical timestamp encoding unifies the data from each modality to a common time reference. This enables the model to accurately perceive the temporal sequence of cross-modal events, laying the foundation for subsequent causal correlation matrix construction and fault diagnosis. This encoding uses a global time projection matrix and a modality-specific time projection matrix to encode absolute time and internal modal rhythms respectively, then superimposes them to achieve precise expression of time-series information.

[0239] Secondly, causal prior knowledge is extracted to suppress spurious correlations. Addressing the issue that pure data-driven models easily misclassify common environmental factors (such as simultaneous increases in summer temperature and load) as fault-related factors, Pearson correlation is used to screen correlation strength, followed by Granger causality tests with Bonferroni correction to eliminate spurious correlations. A causal correlation matrix is ​​then constructed using Hadamard product fusion, ensuring that the prior knowledge injected into the Transformer has genuine causal relationships, effectively improving diagnostic accuracy.

[0240] Third, prior guidance for attention and physical constraints to ensure consistency. The causal correlation matrix is ​​injected into the self-attention score in the form of a logarithmic bias term, guiding the model to prioritize mode combinations with significant causal correlations. At the same time, a relaxed regularization loss constraint training process covering multiple physical laws such as Joule's law, mechanical vibration characteristics, and partial discharge characteristics is introduced to ensure that the model output does not violate basic physical laws, thus making the diagnostic results physically consistent.

[0241] Fourth, attention convergence pinpoints the root cause, and dual-path collaboration ensures reliability. Multi-head self-attention weights are extracted from the model and converged modally as contribution values ​​to accurately locate the root cause parameters leading to the failure, breaking the "black box" limitation of deep learning. Furthermore, by integrating the contribution-weighted physical rule verification score with AI probability into a comprehensive confidence score, diagnostic results are output only when the confidence score meets the standard, effectively suppressing false alarms and providing maintenance personnel with traceable and reliable decision-making basis.

[0242] like Figure 4 As shown, this application also provides an adaptive fault diagnosis system for high and low voltage switchgear based on AI multimodal fusion, including:

[0243] The acquisition module is used to acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points;

[0244] The timing alignment module is used to perform timing alignment on each monitoring parameter in the multimodal monitoring data based on hierarchical timestamp encoding to obtain aligned multimodal monitoring data;

[0245] The causal association module is used to filter historical fault sample data from the aligned multimodal monitoring data, and construct a causal association matrix between various fault modes and monitoring parameters based on the historical fault sample data. The causal association matrix represents the causal relationship between monitoring parameters under various fault modes.

[0246] The model building module is used to inject the causal correlation matrix as a bias term into the multi-head self-attention mechanism of the Transformer network, and to train the Transformer network that fuses the causal correlation matrix based on the aligned multimodal monitoring data and the pre-built loss function that introduces a physical regularization loss term, so as to obtain a fault identification model.

[0247] The fault and root cause identification module is used to identify the multimodal monitoring data generated by the high and low voltage switchgear during the current operation based on the fault identification model, and obtain the fault identification result; extract the multi-head self-attention weight from the fault identification model, and perform root cause localization on the fault identification result based on the multi-head self-attention weight to obtain the root cause parameters leading to the fault.

[0248] This application presents an adaptive fault diagnosis method and system for high and low voltage switchgear based on AI multimodal fusion. By unifying various monitoring parameters to a common time base through hierarchical timestamp encoding, it solves the problem of difficulty in fusing data from different sampling rates. Granger causality verification is used to eliminate statistical spurious correlations caused by environmental common factors, constructing a causal correlation matrix to represent the true causal relationships between monitoring parameters under fault modes. A bias term is injected into the Transformer self-attention mechanism to guide the model to prioritize mode combinations with physical causal relationships, avoiding false alarms caused by changes in operating conditions in purely data-driven models. Furthermore, a relaxed physical regularization loss term is introduced to constrain model training, ensuring that the model output does not violate fundamental principles such as Joule's law and mechanical dynamics, thus enhancing the reliability of the diagnostic results.

[0249] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0250] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory so that the terminal performs any of the methods in this embodiment.

[0251] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0252] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0253] In this embodiment, the memory may include random access memory (RAM) and may also include nonvolatile memory, such as at least one disk storage device.

[0254] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0255] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. The embodiments of the present application are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0256] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion, characterized in that, include: Acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points; The time sequence of each monitoring parameter in the multimodal monitoring data is aligned based on hierarchical timestamp encoding to obtain aligned multimodal monitoring data; Historical fault sample data is filtered from the aligned multimodal monitoring data, and a causal correlation matrix between various fault modes and monitoring parameters is constructed based on the historical fault sample data. The causal correlation matrix represents the causal relationship between monitoring parameters under various fault modes. The causal correlation matrix is ​​injected as a bias term into the multi-head self-attention mechanism of the Transformer network, and the Transformer network fused with the causal correlation matrix is ​​trained based on the aligned multimodal monitoring data and the pre-constructed loss function with physical regularization loss term to obtain the fault identification model. Based on the fault identification model, the multimodal monitoring data generated by the high and low voltage switchgear during the current operation are identified to obtain the fault identification result; and multi-head self-attention weights are extracted from the fault identification model, and the root cause is located based on the multi-head self-attention weights to obtain the root cause parameters leading to the fault.

2. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, Based on hierarchical timestamp encoding, each monitoring parameter in the multimodal monitoring data is time-series aligned to obtain aligned multimodal monitoring data, including: Extract a data point sequence for each monitoring parameter from the multimodal monitoring data, wherein the data point sequence includes multiple data points, and each data point includes a value and a timestamp; Multiple monitoring parameters are divided into high-frequency monitoring parameters and low-frequency monitoring parameters; Based on a pre-built sliding window, time-frequency domain statistical features of multiple time windows are extracted from the data point sequence of high-frequency monitoring parameters to obtain the token sequence corresponding to multiple time windows; The value of each time point in the data point sequence of low-frequency monitoring parameters is used as a token to obtain a token sequence corresponding to multiple time points; The token sequence is converted into a hierarchical timestamp code based on an independent linear projection layer, wherein the mathematical expression for the hierarchical timestamp code is: In the formula, Indicates layered time coding. To collect timestamps, For parameter index, For global absolute time, For parameters Exclusive timescale location, For sinusoidal position coding function, The global time projection matrix, For parameters A dedicated time projection matrix; Timing alignment is performed using hierarchical timestamp encoding based on multiple monitoring parameters to obtain aligned multimodal monitoring data.

3. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, Historical fault sample data is filtered from the aligned multimodal monitoring data, and a causal correlation matrix between various fault modes and monitoring parameters is constructed based on the historical fault sample data, including: Obtain the time points of multiple fault occurrences; A data extraction window is constructed based on the time point of the fault occurrence, and historical fault sample data is filtered from the aligned multimodal monitoring data based on the data extraction window; Extract the fault sample value sequence for each monitoring parameter from the historical fault sample data; Use any two monitoring parameters as a parameter group; Calculate the Pearson correlation between the fault sample value sequences of any two monitoring parameters in each parameter group. And based on the Pearson correlation of all parameter groups Construct a correlation coefficient matrix and normalize the correlation coefficient matrix to obtain the degree of correlation matrix. ,in, and This indicates two different monitoring parameters. Fault mode type; Granger causality check is performed based on the fault sample value sequence of any two monitoring parameters in each parameter group to obtain the causality check flag bit for each parameter group; and a causality check mask matrix is ​​constructed based on the causality check flag bits of all parameter groups. Wherein, a causal verification flag of 1 indicates the existence of a causal relationship, and a causal verification flag of 0 indicates the absence of a causal relationship; Based on the correlation matrix and the causal check mask matrix Calculate the causal relationship matrix , In the formula, It is the product of Hadama.

4. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, The multi-head self-attention mechanism of injecting the causal correlation matrix as a bias term into the Transformer network includes: Configure bias terms based on the causal correlation matrix. , wherein the bias term The mathematical expression is: In the formula, Monitoring parameters in the causal correlation matrix With monitoring parameters The correlation matrix, The set anti-crash value; The bias term By introducing the multi-head self-attention score from the Transformer network, we obtain the updated multi-head self-attention score. Among them, the updated multi-head self-attention score The mathematical expression is: In the formula, For monitoring parameters The query vector of the token. For monitoring parameters The key vector of the token, To query the dimensions of a vector / key vector, This is the prior confidence coefficient. As a time distance factor, For monitoring parameters The original collection timestamp, For monitoring parameters The original collection timestamp, This represents the maximum time delay of physical signal propagation. Based on the updated multi-head self-attention score Update the multi-head self-attention mechanism.

5. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, The mathematical expression for the loss function that introduces the physical regularization loss term is: In the formula, For the total loss, For binary cross-entropy loss, For the first The loss value of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, For the first The weights of the physical constraints, Loss constrained by Joule's law, For mechanical vibration constraint loss, For partial discharge confinement loss, This refers to the batch sample size. The change in temperature For the time change, This is the effective value of the current. Joule heat coefficient, This is a heat dissipation correction item. This is the inherent error tolerance band of Joule's law. To measure the vibration amplitude, For operating force or driving force, Indicates a given operating force or driving force The maximum vibration amplitude at that time, The inherent error tolerance band of vibration constraint, To measure the apparent discharge quantity, To apply voltage, To apply a given voltage Theoretical discharge amount at that time This is the inherent error tolerance band for partial discharge constraint.

6. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, Based on the aligned multimodal monitoring data and a pre-constructed loss function incorporating a physical regularization loss term, a Transformer network fusing the causal correlation matrix is ​​trained to obtain a fault identification model, including: Based on multiple historical fault time points, fault sample fragments are extracted from the aligned multimodal monitoring data, and normal sample fragments are extracted from the aligned multimodal monitoring data. The faulty sample fragment and the normal sample fragment are tokenized and aligned to obtain a data sample. The data samples are labeled to obtain training samples with label vectors, wherein the label vectors include flag bits for the determination results of various fault types; The training samples are input into a Transformer network that fuses causal correlation matrices to obtain prediction results; Based on a pre-constructed loss function that incorporates a physical regularization loss term, the physical constraint loss of the data sample is calculated, and the prediction loss of the prediction result and the label vector is calculated to obtain the total loss. Propagation is performed based on the total loss to adjust the internal parameters of the Transformer network that fuses the causal correlation matrix; Returning to the Transformer network that integrates the training samples and causal correlation matrix, the prediction results are obtained until training is complete, resulting in a fault identification model.

7. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 1, characterized in that, Based on the aforementioned fault identification model, multimodal monitoring data generated by high and low voltage switchgear during current operation are identified to obtain fault identification results, including: Extract the current data segment from the multimodal monitoring data generated during the current operation; The current data fragment is tokenized and aligned to obtain aligned data; The aligned data is normalized to obtain the input vector; The input vector is input into the fault identification model to obtain a fault identification vector, wherein the fault identification vector includes the probabilities of multiple fault types; Fault types with a probability greater than a preset adaptive threshold are used as the fault identification results.

8. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 7, characterized in that, Based on the multi-head self-attention weights, the root cause of the fault is determined by performing root cause localization on the fault identification results, resulting in root cause parameters, including: The contribution of each monitoring parameter is calculated based on the multi-head self-attention weights. The contribution degree The mathematical expression is: In the formula, and All are indexes of monitoring parameters. To monitor the total number of parameters, For attention head index, This represents the total number of heads of multi-head attention. For token index, For monitoring parameters The set of tokens, For monitoring parameters The set of tokens, For the first The first one in the attention. Attention weights for each token; Based on the contribution of each monitoring parameter The various monitoring parameters are sorted, and the top P monitoring parameters are extracted as the initial root cause parameters.

9. The adaptive fault diagnosis method for high and low voltage switchgear based on AI multimodal fusion according to claim 8, characterized in that, Also includes: Contribution based on all initial root cause parameters Calculate the rule matching score of the fault identification results. The rule matching score The mathematical expression is: In the formula, For indicator functions, Fault Corresponding monitoring parameters The statistical early warning threshold, To monitor the value of the parameter at the current moment; Based on the matching score of the rules Calculate the overall confidence level The overall confidence level The mathematical expression is: In the formula, These are the weighting coefficients for the AI's predicted path. For faults in the fault prediction results The probability, Fault The rule matching score; The overall confidence level If the confidence level is greater than or equal to the preset confidence threshold, the fault identification result and root cause parameters are saved and sent to the target object; otherwise, the fault identification result and root cause parameters at the current moment are deleted.

10. An adaptive fault diagnosis system for high and low voltage switchgear based on AI multimodal fusion, characterized in that, include: The acquisition module is used to acquire multimodal monitoring data generated by high and low voltage switchgear during historical operation, wherein the multimodal monitoring data includes the values ​​of various monitoring parameters at multiple historical time points; The timing alignment module is used to perform timing alignment on each monitoring parameter in the multimodal monitoring data based on hierarchical timestamp encoding to obtain aligned multimodal monitoring data; The causal association module is used to filter historical fault sample data from the aligned multimodal monitoring data, and construct a causal association matrix between various fault modes and monitoring parameters based on the historical fault sample data. The causal association matrix represents the causal relationship between monitoring parameters under various fault modes. The model building module is used to inject the causal correlation matrix as a bias term into the multi-head self-attention mechanism of the Transformer network, and to train the Transformer network that fuses the causal correlation matrix based on the aligned multimodal monitoring data and the pre-built loss function that introduces a physical regularization loss term, so as to obtain a fault identification model. The fault and root cause identification module is used to identify the multimodal monitoring data generated by the high and low voltage switchgear during the current operation based on the fault identification model, and obtain the fault identification result; extract the multi-head self-attention weight from the fault identification model, and perform root cause localization on the fault identification result based on the multi-head self-attention weight to obtain the root cause parameters leading to the fault.

Images