Multi-modal input electric power operation equipment safety risk identification method and multi-modal input electric power operation equipment safety risk identification system

Abstract

The invention belongs to the technical field of power equipment safety detection, and particularly relates to a multi-modal input power operation equipment safety risk identification method and system, electrical monitoring index data is converted into an oscillogram, spatial features are extracted through a TDN network, and time sequence features of the electrical monitoring index data are extracted in cooperation with a GNN network; and the time features and the space features are fused, so that the comprehensiveness and the accuracy of safety risk identification of the electric power operation equipment can be remarkably improved. According to the method, the electrical signal is converted into the oscillogram, so that the TDN network can directly perform feature extraction on the oscillogram and accurately capture a local abnormal region in the oscillogram, the limitation that the original TDN cannot directly extract features from time sequence current data is solved, and transient jump of the TDN to the electrical signal can be better utilized.

Description

Technical Field

[0001] The present invention belongs to the field of power equipment safety detection safety technology, and specifically relates to a multi-modal input power operation equipment safety risk identification method and system. Background Art

[0002] In the field of power equipment condition monitoring, especially in identifying potential equipment risks, feature extraction of current / voltage time series signals based on deep learning has become a key method for fault diagnosis. In recent years, some research has attempted to migrate spatiotemporal modeling networks (such as TDN networks) from the field of computer vision to analyze equipment status monitoring data. Equipment status monitoring data generally includes power data, environmental data, and equipment status data. However, due to the fundamental differences between the design characteristics of the original model and the physical properties of the equipment status data, practical applications have significant drawbacks: TDN relies on two-dimensional differential calculations between video frames, while current and voltage signals are one-dimensional time series and lack spatial structure. As a result, TDN's core differential module cannot directly capture the electromagnetic transient characteristics of the time series data current. In addition, high-frequency currents in power signals (such as switching arcs and harmonic interference) are linearly amplified in TDN's inter-frame differential operation, which will mask the microsecond-level characteristics of the electromagnetic signal, resulting in the system being unable to capture transient jumps in the electromagnetic signal. Summary of the Invention

[0003] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a multi-modal input power operation equipment safety risk identification method and system, which solves the problem that when TDN is applied to the feature extraction of electrical monitoring data of power equipment in the prior art, the TDN network cannot directly extract the current timing signal.

[0004] The purpose of the present invention can be achieved through the following technical solution: A multi-modal input power operation equipment safety risk identification method comprises the following steps: S1: Acquire monitoring indicator data of the device in real time, wherein the monitoring indicator data includes electrical monitoring indicator data, and the monitoring indicator data includes current and voltage data, convert the current and voltage data into waveform graphs, and use the TDN network to extract spatial features of the waveform graphs; S2: Using a 1D-CNN network to extract the temporal features of the electrical monitoring index data, and performing feature space alignment on the spatial features and the temporal features; S3: performing feature fusion on the spatial features and temporal features after spatial alignment to obtain input features of electrical monitoring index data.

[0005] Preferably, step S1 includes the following sub-steps: Before waveform conversion, the electrical monitoring index data is subjected to wavelet packet transformation to decompose the signal into different frequency bands. The high-frequency part is processed separately with a high sampling rate to retain the microsecond transient characteristics. Construct a time-frequency feature map based on short-time Fourier transform, and set up an attention mechanism to automatically identify and enhance high-frequency transient areas.

[0006] Preferably, in S3, the fusion function for implementing feature fusion is The formula is: ; in, is a one-dimensional time series signal, is a one-dimensional time series feature extraction function, is the converted waveform image, It is a two-dimensional image feature extraction function.

[0007] Preferably, S2 includes the following sub-steps: Use the fully connected layer of 1D-TDNN to output a temporal feature sequence with the same dimension as the spatial feature; Build a The distance matrix D, the element D of the distance matrix D Represents the Euclidean distance between the spatial feature vector and the temporal feature vector; Create the cumulative cost matrix C: And initialize it; Filling the cumulative cost matrix, performing recursive calculation, determining the optimal alignment path, and performing feature space alignment on the spatial features and the temporal features based on the optimal alignment path.

[0008] Preferably, after step S3, the following steps are further included: S4: Obtain the causal relationship between historical equipment monitoring indicator data and security risks to construct a causal relationship directed graph; S5: Based on the causal relationship directed graph, rules are constructed and confidence levels are initialized to obtain a causal rule base, input data is matched based on the causal rule base, and security risks of the device are identified and output.

[0009] Preferably, the method further includes S6, using a gradient update algorithm to update the confidence level of the causal rule base, specifically including: For each rule in the causal rule base, calculate the gradient of the loss function of each rule with respect to its initial confidence, and update the confidence in the opposite direction of the gradient; Based on the set update constraints, the confidence level is iteratively updated until the preset maximum number of iterations is reached, and the updated confidence level is output.

[0010] Preferably, the cross entropy loss function is used to calculate the difference between the predicted risk of the rule base and the actual risk, and its calculation formula is: ; in, is the i-th rule in the rule base The initial confidence of the rule, is a regularization term, which is L2 regularization and is used to prevent the loss function from overfitting. is the regularization coefficient, is the true risk value at time t, is the predicted risk value at time t, and T is the maximum number of iterations of the loss function.

[0011] Preferably, in S4, the causal relationship includes multiple causal relationships corresponding to the existing security risk types, and the causal relationship corresponding to any security risk includes: multiple cause indicators and multiple effect indicators corresponding to the security risk, and changes in the cause indicators will cause corresponding security risks, and when security risks occur in the equipment, multiple effect indicators will cause changes.

[0012] A multi-modal input power operation equipment safety risk identification system includes: a spatial feature extraction module: used to obtain monitoring indicator data of the equipment in real time and convert the obtained electrical monitoring indicator data into a waveform graph, and use a TDN network to extract the spatial features of the waveform graph; Time series feature extraction module: uses 1D-CNN network to extract the time series features of electrical monitoring index data; Feature fusion module: aligns the spatial features and temporal features and performs feature fusion.

[0013] The beneficial effects of the present invention are: Converting electrical monitoring index data into waveforms and extracting spatial features through the TDN network, combining it with the GNN network to extract the temporal features of the electrical monitoring index data, and fusing the temporal features with the spatial features can significantly improve the comprehensiveness and accuracy of safety risk identification of power operation equipment.

[0014] On the one hand, converting electrical signals into waveforms allows the TDN network to directly extract features from the waveforms and accurately capture local abnormal areas in the waveforms (such as the spike shape of arc interference and waveform distortion caused by harmonics). This solves the limitation of the original TDN that it cannot directly extract features from time-series current data, and can better utilize TDN to detect transient jumps in electrical signals.

[0015] GNNs, on the other hand, excel at mining the dynamic dependencies of time-series data. They can effectively extract the temporal evolution trends of electrical signals and capture the temporal correlation characteristics of equipment states transitioning from normal to abnormal. When these two are combined, spatial features (static morphology) complement temporal features (dynamic changes). Aligning the feature space ensures precise semantic correspondence, preserving the spatial distribution details of waveform anomalies while incorporating the temporal patterns of state evolution, thus avoiding the limitations of single modal features. This significantly enhances the ability to identify early-stage latent faults and complex, multi-factor faults, ultimately improving the accuracy and robustness of risk identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0017] Figure 1 A flowchart of method steps according to an embodiment of the present invention; Figure 2 This is a sub-step block diagram of S6 in another embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0019] See also Figure 1-Figure 2 This embodiment provides a multi-modal input power operation equipment safety risk identification method, comprising the following steps: S1: Acquire multiple monitoring indicator data of the device in real time. The monitoring indicator data of the device includes electrical monitoring indicator data, which includes voltage and current data. Convert the time domain signals of the voltage and current data into frequency domain representation, analyze the frequency components of the signal, reconstruct the waveform through inverse Fourier transform (IFFT), use sine waves to represent the voltage and current data, output the waveform, and use the TDN network to extract the spatial features of the waveform. Specifically, it includes: S11: Multi-dimensional monitoring indicator data of equipment operation is obtained in real time through multiple types of sensors and monitoring devices deployed on power operation equipment. The multi-dimensional monitoring indicator data includes: Electrical monitoring indicator data: Real-time collection of current and voltage data of multiple devices through current transformers and voltage sensors; Equipment status monitoring indicator data: Vibration sensors are used to collect vibration signals from equipment bearings and windings, and infrared temperature sensors are used to collect the temperature of busbar joints and other locations where equipment needs to monitor temperature. Of course, in the monitoring of the status of specific power equipment, the multi-dimensional monitoring indicator data here is categorized and adjusted according to the monitoring indicator data actually needed. This embodiment focuses on feature extraction of electrical monitoring indicator data with large jumps.

[0020] S12: Feature extraction is performed on the acquired multiple monitoring indicator data, and the raw data is converted into feature vectors, including: The use of TDN (Temporal Difference Network) in the feature extraction of electrical monitoring indicator data is irreplaceable. Its core value lies in its sensitivity to transient mutation characteristics, which is difficult to achieve with traditional time series models. However, since TDN is applied to image frame processing, it lacks spatial structure in one-dimensional time series such as voltage and current, which makes TDN's core difference module unable to directly capture the electromagnetic transient characteristics of current. Therefore, this solution first converts the acquired electrical monitoring indicator data into a waveform graph, uses the TDN network to extract the spatial characteristics of the waveform graph, and combines TDN's ability to capture time series characteristics with imaging methods to enable TDN to capture the instantaneous jump of voltage and current through the waveform graph. The frequency characteristics of the vibration signal are extracted through the Fourier transform method. Then, the discrete wavelet transform (DWT) is applied to perform multi-scale decomposition. The db4 wavelet basis function is selected and decomposed into three layers. The low-frequency approximate coefficients and high-frequency detail coefficients are extracted to convert the non-stationary signal into a time-frequency feature that is easier to analyze. S2: Use the 1D-CNN network to extract the time series features of the electrical monitoring indicator data and perform feature space alignment on the spatial features and time series features. Spatial alignment here refers to the spatial alignment of the one-dimensional time series features and the two-dimensional spatial features. This can be achieved using the DTW algorithm or multi-scale modeling. S3: Perform feature fusion on the spatial features and time series features after spatial alignment to obtain the input features of the electrical monitoring index data and implement the fusion function of feature fusion The formula is: ; in, is a one-dimensional time series signal, is a one-dimensional time series feature extraction function, is the converted waveform image, It is a two-dimensional image feature extraction function; The subsequent steps are to fuse the input features with other monitoring indicator data, input the pre-set learning and recognition model, perform risk identification and output.

[0021] Converting electrical monitoring index data into waveforms and extracting spatial features through the TDN network, combining it with the GNN network to extract the temporal features of the electrical monitoring index data, and fusing the temporal features with the spatial features can significantly improve the comprehensiveness and accuracy of safety risk identification of power operation equipment.

[0022] On the one hand, the waveform diagram converts abstract electrical signals (such as transient changes in current and voltage) into intuitive spatial morphological features (such as waveform distortion and spike distribution). The TDN network, with its multi-scale convolution and attention mechanism, can accurately capture local abnormal areas in the waveform diagram (such as the spike shape of arc interference and waveform distortion caused by harmonics), solving the problem of highly hidden risk features in the original electrical data and providing rich spatial topological information for risk identification.

[0023] GNNs, on the other hand, excel at mining the dynamic dependencies of time-series data. They can effectively extract temporal trends in electrical signals (such as amplitude changes and periodic fluctuations during fault development) and capture the temporal correlation characteristics of equipment states transitioning from normal to abnormal. When these two are integrated, spatial features (static morphology) complement temporal features (dynamic changes). Aligning feature spaces ensures precise semantic correspondence, preserving the spatial distribution details of waveform anomalies while incorporating the temporal patterns of state evolution, thus avoiding the limitations of single modal features. This significantly enhances the ability to identify early-stage latent faults and complex, multi-factor faults, ultimately improving the accuracy, robustness, and early warning capabilities of risk identification.

[0024] Because high-frequency currents in power signals (such as switching arcs and harmonic interference) are linearly amplified in the inter-frame differential operation of TDN, the microsecond-level characteristics of the electromagnetic signal will be masked, resulting in the system being unable to capture transient jumps in the electromagnetic signal. Therefore, in one implementation method, before waveform conversion in S1, the electrical monitoring indicator data is subjected to wavelet packet transform (WPT) to decompose the signal into different frequency bands (such as 0-5kHz for low frequency, 5-50kHz for medium frequency, and 50kHz-1MHz for high frequency). The high-frequency part is processed separately using a high sampling rate (≥2MHz) to retain the microsecond-level transient characteristics.

[0025] Construct a time-frequency feature map based on STFT (Short-time Fourier Transform), design an attention mechanism to automatically identify and enhance high-frequency transient areas (such as the rising edge of the switching arc), and suppress the interference of the stable power frequency signal. The spatial features extracted by TDN are subjected to Fourier transform to separate the high-frequency components. The transient impact features are identified by the spectral kurtosis index, and then the enhanced frequency domain features are re-integrated with the time domain features.

[0026] It can effectively solve the problem of microsecond-level feature masking when the TDN network processes high-frequency signals of electrical monitoring index data, enabling the system to accurately capture transient jumps of electromagnetic signals and improve the early warning capability of high-frequency jump risks such as switching arcs and harmonic interference.

[0027] In one embodiment, based on the multi-dimensional monitoring indicator data obtained above, the method for assessing the risks of power equipment can use a preset learning recognition model or a causal relationship rule library to learn the causal relationship between monitoring indicators and risks. The method of using a causal relationship diagram to identify risks includes: S4: Obtain the causal relationship between the monitoring indicators and security risks in the device and construct a directed causal relationship graph. The causal relationship includes multiple causal relationships corresponding to the existing security risk types. The causal relationship corresponding to any security risk includes: multiple cause indicators and multiple effect indicators corresponding to the security risk. Changes in the cause indicators will cause corresponding security risks. When the device has a security risk, multiple effect indicators will change. Specifically: S41: Combined with the power equipment risk case library, such as the State Grid Equipment Defect Record and Industry Standard (DL / T 722-2019), sort out the types of security risks existing in typical power equipment and clarify the risk boundaries, such as: Electrical risks: open circuit, short circuit, overload, insulation breakdown and partial discharge, etc.; Mechanical risks: loose components, bearing wear, and operating mechanism jamming, etc.; S42: For each possible risk of multiple power equipment, we use expert experience, fault tree analysis, and machine learning combined with historical risk types to break down the "cause indicators" (risk inducements) and "effect indicators" (risk consequences); Cause indicators are pre-indicators that directly lead to risk occurrence. That is, changes in these indicators cause equipment risks to occur. For example, the cause indicators for "overload risk" are: the current continuously exceeds 1.2 times the rated value (electrical monitoring indicator data), the cooling system fan speed is lower than the threshold (equipment status indicator), and the equipment's ambient temperature exceeds XX°C; the cause indicators for "insulation aging risk" are: the equipment has been in operation for more than XX years and the partial discharge exceeds the standard. The result indicator is the subsequent indicator change caused by the occurrence of the risk, that is, the indicator changes due to the occurrence of the risk, for example, the result indicator of "overload risk"; the result indicator of "bearing wear risk": the energy of the vibration frequency in the XXHz range increases abnormally; S43: Construct and verify the causal relationship directed graph: After clarifying the indicators and risk types, use a directed graph to build a causal relationship directed graph of the indicators, risk types, and the causal relationship between them. Specifically: Taking "risk type" as the core node, use directed edges to connect cause indicators and effect indicators, and mark the causal strength value; The probability of "cause indicator combination → risk occurrence → effect indicator change" is calculated through historical data, and the causal weight is optimized through Bayesian network; In the causal screening process, if the causal strength between certain premise attributes and conclusions is less than the set threshold, the edges with such causal strength will be deleted to form a causal directed graph with optimized structure.

[0028] S5: Constructing rules based on the causal relationship directed graph and initializing confidence levels to obtain a causal rule base, matching input data based on the causal rule base, and identifying and outputting security risks of the device; S51: Define the rule format: Each rule corresponds to a security risk judgment logic, and the structure is "[due to indicator condition 1] and [due to indicator condition 2], , [due to indicator conditions n], n is the total number of due indicator conditions, then it is judged as [safety risk type], and the confidence level is [initial value]", for example: Rule 1 (overload risk): "If the current > XX and the rated current and the fan speed If the speed is XXr / min, it is determined to be an overload risk with a confidence level of XX. XX is determined using the monitoring parameters of the on-site power equipment. XX represents the uncertainty value, which is confirmed based on the specific usage scenario and specific on-site indicators. Confidence initialization: The initial confidence value is determined by combining expert scoring and historical data statistics. For example, if the historical data shows that "when the conditions of Rule 1 are met, the actual probability of overload risk is XX", the initial confidence value can be set to XX; S52: Match input data with the rule base to identify potential risks: Iterate the input data in the causal rule base to verify the cause and effect conditions of the rules; Result indicators assist in verification. If, after a rule determines a risk, the corresponding result indicator is simultaneously abnormal, the overall confidence level will be increased by a certain amount. The amount here is determined based on field experience. Based on the set threshold, filter the risk indicators with confidence levels higher than the set threshold and output the risk identification results: The output risk information includes: risk type, occurrence time, associated cause and effect indicators, and comprehensive confidence level; And according to the set warning threshold of confidence, when the comprehensive confidence is higher than the set warning threshold, a real-time warning is triggered and pushed to the control terminal.

[0029] Since the operating status and risk factors of power equipment have dynamic characteristics, the initial confidence rule base is difficult to adapt to these changes in the long term. Specifically, during long-term operation, power equipment will be affected by aging, wear, load fluctuations, or changes in environmental factors (such as seasonal changes, temperature and humidity fluctuations), which will cause the causal relationship between risks and monitoring indicators to shift. For example, the initial confidence level of the association between "current exceeding the standard" and "overload risk" of new equipment may be high, but as the equipment insulation ages, the risk probability under the same current conditions will increase, and the confidence level of the initial rule will not be able to reflect the actual risk level; in addition, new fault cases, operation and maintenance strategy adjustments, or changes in sensor accuracy will also cause data distribution drift. For example, a newly added vibration sensor may collect more subtle bearing wear signals, making the original "vibration index and wear risk" rule confidence level no longer accurate. Therefore, it is necessary to use the new collected data to update the confidence level in the rule base, so that the causal rule base can capture these changes. Therefore, step S6 is also included: S6: Using the gradient update algorithm to update the confidence of the causal rule base, step S6 specifically includes: With the accuracy of risk identification as the goal, define the loss function gradient, set the learning rate, maximum number of iterations, and convergence threshold. For each rule in the causal rule base, calculate the gradient of the loss function with respect to the initial confidence level, and update the confidence level in the opposite direction of the gradient. Input data, multi-dimensional time series data of equipment monitoring indicators , are the d monitoring indicator features at time t; The rule base contains N rules , is the i-th rule The initial confidence of the rule, , is a rule condition; The actual risk label at time t ; The specific calculation process is: Step 1: Calculate the activation weight of each rule in the rule base: For the input at time t , calculate the activation weight of each rule ; = ,in, for Function, mapping similarity to ; Step 2: Risk prediction output: predict the risk probability of the data in the input data rule base. The prediction value is the weighted confidence of each rule. The specific calculation is: ; Step 3: Use the cross entropy loss function to measure the difference between the predicted risk and the actual risk. The calculation formula is: ; in, is a regularization term, which is L2 regularization and is set as a penalty term for model complexity to prevent overfitting of the loss function during parameter update. is the regularization coefficient. It is set based on the expert experience value according to the aging of the equipment during its operation and the adjustment of the risk value in the rule base. Step 4: For each rule, calculate the loss function L with respect to its initial confidence The partial derivative (gradient) of is simplified as follows: ; Step 5: Perform cyclic iterative updates according to the learning rate Update the confidence level with the maximum number of iterations set , and its updated value confidence is: , output the updated confidence; By continuously updating the confidence level of the rule base with new data, the rule base can dynamically capture the changes that occur during the long-term use of the equipment, correct deviations caused by equipment status evolution, environmental interference or data drift, and ensure that the confidence level of risk identification always matches the actual operating scenario, thereby maintaining high-accuracy risk warning capabilities in the long term, avoiding misjudgments or missed judgments due to rigid rules, and providing continuous and reliable technical support for the safe operation of power equipment.

[0030] In one embodiment, each element of the temporal feature sequence and the spatial feature sequence is regulated in time and space to facilitate subsequent feature fusion. In the prior art, when the DTW algorithm is used for temporal adjustment, the feature vector data extracted by TDN is a two-dimensional vector feature, and the temporal feature sequence is a one-dimensional vector. Therefore, it is necessary to unify the spatial feature and the temporal feature vector sequence in terms of dimension before performing feature fusion. In particular, when the dimension of the temporal feature is expanded by using a 1D-CNN or a fully connected layer, the dimension of the temporal feature is expanded by using a 1D-CNN or a fully connected layer. For example, a 3×3 convolution kernel is used to extract a local temporal pattern from a pseudo-two-dimensional matrix, and the number of output channels is set to d (consistent with d of the spatial feature). Ensure that the time series features after dimensionality increase are strictly uniform in dimension through linear projection or semantic embedding; The specific steps to achieve spatiotemporal alignment using the DWT algorithm include: Step 1: Feature vectorization: Represent the waveform spatial features extracted by the TDN network as a sequence X , where each element It is an n-dimensional vector, which represents the characteristics of the waveform at different positions; The temporal features extracted by 1D-CNN are represented as sequence Y , where each element Is an m-dimensional vector, representing the time series features within the time window, since; Step 1: Calculate the distance matrix: Construct a The distance matrix D, the element D of the distance matrix D Representation space feature vector With the time series feature vector The Euclidean distance between Step 3: Create the cumulative cost matrix C: And initialize it; Step 4: Dynamic programming to fill the cumulative cost matrix: For i = 1 to m, j = 1 to n, fill in the cumulative cost matrix C, where the calculation formula for any element of the cumulative cost matrix C is: C D ; By recursive calculation, find the starting point To current point The minimum cumulative path cost; Step 5: Backtrack the optimal path, starting from the lower right corner of the matrix and backtracking to the upper left corner to determine the optimal path; The backtracking rule is: If And j ,choose The direction corresponding to the minimum value; If you reach the boundary, trace back along the boundary to the starting point; Finally, we get path P , used to represent the optimal alignment between spatial features and temporal features; Step 6: According to the optimal path P, the spatial vector sequence X and the time series feature vector Y are mapped into the same space to generate an aligned feature sequence.

[0031] By dynamically unifying the temporal feature data and spatial feature data on the time axis and using a specific optimal alignment path, the time vector and the space vector are unified in time and space, making it easier to perform feature fusion later.

[0032] When using the DTW algorithm to spatially align spatial features and temporal features, DTW may over-distort the sequence (i.e., mapping one time point to multiple time points), resulting in distortion of the temporal information. In power signal detection, this distortion may change the temporal relationship of transient events (such as a spike signal being elongated), thereby affecting subsequent risk identification judgments based on power signal detection.

[0033] Because regional power grid data in the power industry is difficult to share directly due to restrictions such as privacy protection and management boundaries, the original rule base is built only based on single-region data and is easily affected by regional specificity (such as climate, equipment model, and load characteristics). This leads to reduced accuracy when rules are applied across regions (for example, rules for the humid southern region have a high misjudgment rate in the dry northern region). Therefore, to improve the regional applicability of this solution, in one implementation, the local rule parameters of multiple regional power grids are aggregated through a federated learning framework, and a weighted average algorithm is used to generate the relationship between monitoring indicators and security risks. Specifically, the following steps are involved: Initialize and train local rule parameters. Each regional power grid independently builds an initial causal rule library based on local historical fault data, equipment operation records, and expert experience, and determines the parameters of local rules (such as the association weight between causal indicators and risks, and the initial confidence value). Power grids across the country encrypt local rule parameters through federated learning clients. These encrypted rule parameters need to be uploaded to ensure data privacy in each region. The federated learning server receives the encrypted parameters from each region and forms a parameter set. Federated learning is a distributed machine learning framework that allows multiple participants to jointly train a model without sharing the original data. In power safety risk operations and maintenance, the federated learning mechanism can be used to expand risk-inducing rules for more risk types. The server determines weighting coefficients based on the data weights, device type coverage, and historical rule accuracy of each region, and aggregates the encrypted parameters using a weighted average algorithm. The specific weight ratios are calculated based on the specific relationships between the power grids in each region, and are determined specifically for the parameter values ​​of each power grid in each region. For example, the correlation weight for "temperature and equipment overheating risk" is 0.3 in North China (small data volume but high accuracy), and 0.5 in East China (large data volume). After aggregation, the global correlation weight is generated = 0.3 × North China parameter + 0.5 × East China parameter + weighted parameters of other regions; After decryption, the aggregated global parameters are distributed to the power grid clients in each region. Each regional power grid client adjusts the local rule base based on the global parameters to generate monitoring indicators and security risk relationships that integrate multi-regional characteristics. Through federated learning of power grids in various regions, aggregation of rule parameters can be achieved without sharing data between regions, complying with the power industry's data security regulations and avoiding the leakage of sensitive information in the region. After aggregating parameters from multiple regions, the rule base can be compatible with the equipment characteristics, environmental differences, and operating conditions of different regions. For example, the unified "overload risk" rule can adapt to both the high-load scenarios in northern winter and the load characteristics in the humid environment in southern summer, thereby improving the accuracy of cross-regional applications.

[0034] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A multi-modal input method for identifying safety risks of power operation equipment, characterized by: The following steps are involved: S1: Acquire monitoring indicator data of the device in real time, wherein the monitoring indicator data includes electrical monitoring indicator data, and the electrical monitoring indicator data includes current and voltage data, convert the electrical monitoring indicator data into a waveform graph, and use a TDN network to extract spatial features of the waveform graph; S2: Using a 1D-CNN network to extract the temporal features of the electrical monitoring index data, and performing feature space alignment on the spatial features and the temporal features; S3: performing feature fusion on the spatial features and temporal features after spatial alignment to obtain input features of electrical monitoring index data.

2. The multi-modal input safety risk identification method for electric power equipment according to claim 1, characterized in that: Step S1 includes the following sub-steps: Before waveform conversion, the electrical monitoring index data is subjected to wavelet packet transformation to decompose the signal into different frequency bands. The high-frequency part is processed separately with a high sampling rate to retain the microsecond transient characteristics. Construct a time-frequency feature map based on short-time Fourier transform, and set up an attention mechanism to automatically identify and enhance high-frequency transient areas.

3. The multi-modal input safety risk identification method for electric power equipment according to claim 1, characterized in that: In S3, the fusion function that realizes feature fusion The formula is: ； in, is a one-dimensional time series signal, is a one-dimensional time series feature extraction function, is the converted waveform image, It is a two-dimensional image feature extraction function.

4. The method for identifying safety risks of electric power equipment using multi-modal input according to claim 1, characterized in that: S2 includes the following sub-steps: Use the fully connected layer of 1D-TDNN to output a temporal feature sequence with the same dimension as the spatial feature; Build a The distance matrix D, the element D of the distance matrix D Represents the Euclidean distance between the spatial feature vector and the temporal feature vector; Create the cumulative cost matrix C: And initialize it; Filling the cumulative cost matrix, performing recursive calculation, determining the optimal alignment path, and performing feature space alignment on the spatial features and the temporal features based on the optimal alignment path.

5. The method for identifying safety risks of electric power equipment using multi-modal input according to claim 1, characterized in that: After step S3, the following steps are also included: S4: Obtain the causal relationship between historical equipment monitoring indicator data and security risks to construct a causal relationship directed graph; S5: Based on the causal relationship directed graph, rules are constructed and confidence levels are initialized to obtain a causal rule base, input data is matched based on the causal rule base, and security risks of the device are identified and output.

6. The multi-modal input safety risk identification method for electric power equipment according to claim 5, characterized in that: It also includes S6, which uses the gradient update algorithm to update the confidence of the causal rule base, including: For each rule in the causal rule base, calculate the gradient of the loss function of each rule with respect to its initial confidence, and update the confidence in the opposite direction of the gradient; Based on the set update constraints, the confidence level is iteratively updated until the preset maximum number of iterations is reached, and the updated confidence level is output.

7. The multi-modal input safety risk identification method for electric power equipment according to claim 5, characterized in that: The cross entropy loss function is used to calculate the difference between the predicted risk of the rule base and the actual risk. The calculation formula is: ； in, is the i-th rule in the rule base The initial confidence of the rule, is a regularization term, which is L2 regularization and is used to prevent the loss function from overfitting. is the regularization coefficient, is the true risk value at time t, is the predicted risk value at time t, and T is the maximum number of iterations of the loss function.

8. The multi-modal input safety risk identification method for electric power equipment according to claim 7, characterized in that: In S4, the causal relationship includes multiple causal relationships corresponding to the existing security risk types. The causal relationship corresponding to any security risk includes: multiple cause indicators and multiple effect indicators corresponding to the security risk. When the cause indicator changes, the corresponding security risk will occur. When the equipment has a security risk, it will cause multiple effect indicators to change.

9. A multi-modal input power operation equipment safety risk identification system, applying the multi-modal input power operation equipment safety risk identification method according to any one of claims 1 to 4, comprising: Spatial feature extraction module: used to obtain the monitoring index data of the equipment in real time and convert the obtained electrical monitoring index data into a waveform graph, and use the TDN network to extract the spatial features of the waveform graph; Time series feature extraction module: uses 1D-CNN network to extract the time series features of electrical monitoring index data; Feature fusion module: aligns the spatial features and temporal features and performs feature fusion.

Images