A method and system for identifying safety risks of power operation equipment with multi-modal input

Abstract

The present application 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, which converts electrical monitoring index data into a waveform graph and extracts spatial features through a TDN network, cooperates with a GNN network to extract time sequence features of the electrical monitoring index data, and fuses the time features and the spatial features, so as to significantly improve the comprehensiveness and accuracy of power operation equipment safety risk identification. The electrical signal is converted into a waveform graph, so that the TDN network can directly extract features from the waveform graph, accurately capture local abnormal areas in the waveform graph, solve the limitation that the original TDN cannot directly extract features from the time sequence current data, and better utilize the TDN for transient jump of the electrical signal.

Description

Technical Field

[0001] This invention belongs to the field of power equipment safety inspection and safety technology, specifically relating to a method and system for identifying safety risks of power operation equipment with multimodal input. Background Technology

[0002] In the field of power equipment condition monitoring, particularly in identifying potential risks, feature extraction from 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 transfer spatiotemporal modeling networks (such as TDN networks) from the field of computer vision to the analysis of equipment condition monitoring data. Equipment condition monitoring data generally includes power data, environmental data, and equipment condition data. However, due to the fundamental differences between the design characteristics of the original models and the physical properties of the equipment condition data, practical applications suffer from significant limitations.

[0003] TDN relies on two-dimensional differential calculations between video frames, while current and voltage signals are one-dimensional time series and lack spatial structure. Consequently, the core differential module of TDN cannot directly capture the electromagnetic transient characteristics of time-series data current. Furthermore, high-frequency currents in power signals (such as switching arcs and harmonic interference) are linearly amplified during the inter-frame differential operation of TDN, which masks the microsecond-level characteristics of electromagnetic signals, thus preventing the system from capturing transient changes in electromagnetic signals. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method and system for identifying safety risks of power operation equipment with multimodal input, solving the problem that in the existing technology, when applying TDN to the feature extraction of electrical monitoring data of power equipment, the TDN network cannot directly extract the current time sequence signal.

[0005] The objective of this invention can be achieved through the following technical solution: a method for identifying safety risks of power operation equipment with multimodal input, comprising the following steps:

[0006] S1: Real-time acquisition of equipment monitoring index data, including electrical monitoring index data, including current and voltage data; conversion of the current and voltage data into waveforms; extraction of spatial features of the waveforms using a TDN network.

[0007] S2: Use a 1D-CNN network to extract the temporal features of electrical monitoring index data, and perform feature space alignment between the spatial features and the temporal features;

[0008] S3: Perform feature fusion on the spatial features and temporal features after spatial alignment to obtain the input features of electrical monitoring index data;

[0009] S4: Obtain historical device monitoring index data and construct a causal relationship directed graph between the data and security risks;

[0010] S5: Construct rules based on the causal relationship directed graph and initialize confidence to obtain a causal rule base. Match the input data based on the causal rule base to identify and output the security risks of the device.

[0011] Preferably, step S1 includes the following sub-steps:

[0012] Before waveform conversion, wavelet packet transform is performed on electrical monitoring index data to decompose the signal into different frequency bands. High sampling rate is used to process the high-frequency part separately to retain microsecond-level transient characteristics.

[0013] Construct a time-frequency feature map based on short-time Fourier transform, and set an attention mechanism to automatically identify and enhance high-frequency transient regions.

[0014] Preferably, in S3, the fusion function that implements feature fusion The formula is:

[0015] ;

[0016] in, It is a one-dimensional time-series signal. This is a one-dimensional time-series feature extraction function. The converted waveform image. This is a two-dimensional image feature extraction function.

[0017] Preferably, S2 includes the following sub-steps:

[0018] The fully connected layers of a 1D-CNN output a sequence of temporal features with the same dimension as the spatial features;

[0019] Build a The distance matrix D, and the elements D of the distance matrix D. Represents the Euclidean distance between spatial feature vectors and temporal feature vectors;

[0020] Create the cumulative cost matrix C: And initialize it;

[0021] The cumulative cost matrix is ​​filled, recursive calculations are performed to determine the optimal alignment path, and feature space alignment is performed on the spatial features and temporal features based on the optimal alignment path.

[0022] Preferably, it also includes S6, which updates the confidence level of the causal rule base using a gradient update algorithm, specifically including:

[0023] 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 reverse direction of the gradient;

[0024] 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.

[0025] Preferably, the cross-entropy loss function is used to calculate the difference between the predicted risk and the actual risk in the rule base, and the calculation formula is as follows:

[0026] ;

[0027] in, For the j-th rule in the rule base Initial confidence level of the rule This is the regularization term, an L2 regularization used to prevent overfitting of the loss function. The regularization coefficient is... It is the true risk value at time t. It is the predicted risk value at time t, where T is the total number of time steps.

[0028] Preferably, in S4, the causal relationship includes multiple causal relationships corresponding to different types of security risks. The causal relationship corresponding to any security risk includes multiple causal indicators and multiple effect indicators corresponding to the security risk. A change in the causal indicator will result in a corresponding security risk. After a security risk occurs in the device, multiple effect indicators will change.

[0029] A multimodal input power operation equipment safety risk identification system includes: a spatial feature extraction module: used to acquire monitoring index data of the equipment in real time and convert the acquired electrical monitoring index data into waveforms, and use a TDN network to extract the spatial features of the waveforms;

[0030] Temporal feature extraction module: Uses a 1D-CNN network to extract temporal features from electrical monitoring index data;

[0031] Feature fusion module: performs feature space alignment and feature fusion on the spatial features and temporal features.

[0032] The beneficial effects of this invention are as follows:

[0033] Converting electrical monitoring index data into waveforms and extracting spatial features through a TDN network, combined with a GNN network to extract temporal features from the electrical monitoring index data, and fusing the temporal and spatial features, can significantly improve the comprehensiveness and accuracy of identifying safety risks in power operation equipment.

[0034] On the one hand, converting electrical signals into waveforms allows TDN networks to directly extract features from the waveforms, accurately capturing local abnormal regions in the waveforms (such as the peak shape of arc interference and waveform distortion caused by harmonics). This solves the limitation that the original TDN cannot directly extract features from time-series current data, and can better utilize the transient jumps of electrical signals by the TDN.

[0035] On the other hand, GNN excels at uncovering dynamic dependencies in time-series data, effectively extracting the evolution trend of electrical signals over time and capturing the temporal correlation features of equipment states transitioning from normal to abnormal. When the two are fused, spatial features (static morphology) and temporal features (dynamic changes) complement each other. Feature space alignment 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 latent faults and complex multi-factor faults, ultimately improving the accuracy and robustness of risk identification. Attached Figure Description

[0036] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0037] Figure 1 This is a flowchart illustrating the method steps according to one embodiment of the present invention;

[0038] Figure 2 This is a flowchart of a sub-step S6 in another embodiment of the present invention. Detailed Implementation

[0039] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0040] Please see Figures 1-2 This embodiment provides a method for identifying safety risks of power operation equipment with multimodal input, including the following steps:

[0041] S1: Real-time acquisition of multiple monitoring index data of the equipment, including electrical monitoring index data such as voltage and current data. The time-domain signals of the voltage and current data are converted to frequency-domain representation, the frequency components of the signals are analyzed, and the waveforms are reconstructed using Inverse Fourier Transform (IFFT). Sine waves are used to represent the voltage and current data, and waveform graphs are output. The spatial features of the waveform graphs are extracted using a TDN network, specifically including:

[0042] S11: Real-time acquisition of multi-dimensional monitoring data of equipment operation through various types of sensors and monitoring devices deployed on power operation equipment. These multi-dimensional monitoring data include:

[0043] Electrical monitoring data: Real-time acquisition of current and voltage data from multiple devices via current transformers and voltage sensors;

[0044] Equipment status monitoring index 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 bus joints and other parts of the equipment where temperature monitoring is required.

[0045] Of course, depending on the specific power equipment status monitoring, the multi-dimensional monitoring indicator data here will be adjusted in categories according to the actual monitoring indicator data required. This embodiment focuses on feature extraction of electrical monitoring indicator data with large jumps.

[0046] S12: Perform feature extraction on the acquired monitoring indicator data, transforming the raw data into feature vectors, including:

[0047] The use of TDN (Temporal Difference Network) in the feature extraction of electrical monitoring index data is irreplaceable. Its core value lies in its sensitivity to transient change features, which is difficult to achieve with traditional time series models. However, since TDN is applied to image frame processing, it lacks spatial structure for one-dimensional time series such as voltage and current. Consequently, the core difference module of TDN cannot directly capture the electromagnetic transient features of current. Therefore, this solution first converts the acquired electrical monitoring index data into waveforms and uses the TDN network to extract the spatial features of the waveforms. By combining TDN's ability to capture time series features with image processing techniques, TDN can capture instantaneous jumps in voltage and current through waveforms.

[0048] The frequency features of the vibration signal are extracted by Fourier transform, and then multi-scale decomposition is performed by discrete wavelet transform (DWT). The db4 wavelet basis function is selected and decomposed to 3 levels to extract low-frequency approximation coefficients and high-frequency detail coefficients, thus transforming the non-stationary signal into time-frequency features that are easier to analyze.

[0049] S2: Use a 1D-CNN network to extract the temporal features of electrical monitoring index data, and perform feature space alignment between spatial features and temporal features. Here, spatial alignment means aligning one-dimensional temporal features and two-dimensional spatial features in space. This can be achieved using the DTW algorithm or multi-scale modeling.

[0050] S3: Perform feature fusion on the spatially aligned spatial features and temporal features to obtain the input features of the electrical monitoring index data, and implement the feature fusion function. The formula is:

[0051] ;

[0052] in, It is a one-dimensional time-series signal. This is a one-dimensional time-series feature extraction function. The converted waveform image. This is a two-dimensional image feature extraction function;

[0053] The subsequent steps involve fusing the input features with other monitoring indicator data, inputting them into a pre-set learning and identification model, identifying risks, and then outputting the results.

[0054] Converting electrical monitoring index data into waveforms and extracting spatial features through a TDN network, combined with a GNN network to extract temporal features from the electrical monitoring index data, and fusing the temporal and spatial features, can significantly improve the comprehensiveness and accuracy of identifying safety risks in power operation equipment.

[0055] On the one hand, waveform diagrams transform abstract electrical signals (such as transient changes in current and voltage) into intuitive spatial morphological features (such as waveform distortion and spike distribution). TDN networks, with their multi-scale convolution and attention mechanisms, can accurately capture local abnormal regions in waveform diagrams (such as the spike shape of arc interference and waveform distortion caused by harmonics), solving the problem of strong concealment of risk features in raw electrical data and providing rich spatial topological information for risk identification.

[0056] On the other hand, GNN networks excel at mining the dynamic dependencies of time-series data, effectively extracting the evolution trend of electrical signals over time (such as amplitude changes and periodic fluctuations during fault development) and capturing the temporal correlation features of equipment states from normal to abnormal. When the two are fused, spatial features (static morphology) and temporal features (dynamic changes) complement each other. Through feature space alignment, accurate semantic correspondence is ensured, preserving the spatial distribution details of waveform anomalies while incorporating the temporal patterns of state evolution, avoiding the limitations of single-modal features. This significantly enhances the ability to identify early latent faults and complex multi-factor faults, ultimately improving the accuracy, robustness, and early warning capabilities of risk identification.

[0057] Because the high-frequency current in the power signal (such as switching arcs and harmonic interference) is linearly amplified in the inter-frame differential operation of TDN, it will mask the microsecond-level characteristics of the electromagnetic signal, thus causing the system to be unable to capture the transient jumps of the electromagnetic signal. Therefore, in one implementation, before the waveform conversion in S1, wavelet packet transform (WPT) is performed on the electrical monitoring index data 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 with a high sampling rate (≥2MHz) to retain the microsecond-level transient characteristics.

[0058] A time-frequency feature map based on STFT (Short Time Fourier Transform) is constructed, and an attention mechanism is designed to automatically identify and enhance high-frequency transient regions (such as the rising edge of a switching arc) while suppressing interference from stable power frequency signals.

[0059] Fourier transform is performed on the spatial features extracted by TDN to separate high-frequency components. Transient impact features are identified by the spectral kurtosis index. Then, the enhanced frequency domain features are re-fused with the time domain features.

[0060] It can effectively solve the problem of microsecond-level feature masking when TDN networks process high-frequency signals of electrical monitoring index data, enabling the system to accurately capture transient jumps in electromagnetic signals and improve the early warning capability for high-frequency jump risks such as switching arcs and harmonic interference.

[0061] In one implementation, the method for assessing the risks of power equipment based on the aforementioned multi-dimensional monitoring indicator data can be achieved by using a pre-set learning and identification model or a causal relationship rule base to learn the causal relationship between monitoring indicators and risks. The method of using a causal relationship graph for risk identification includes:

[0062] S4: Construct a directed causal graph to obtain the causal relationships between monitoring indicators and security risks in the device. This graph includes multiple causal relationships corresponding to different security risk types. Each security risk's causal relationship includes multiple causal indicators and multiple effect indicators. A change in one of the causal indicators will result in a corresponding security risk. After a security risk occurs in the device, multiple effect indicators will change. Specifically:

[0063] S41: Based on the power equipment risk case database, such as the State Grid equipment defect records and industry standards (DL / T 722-2019), identify the types of safety risks present in typical power equipment and clarify the risk boundaries, for example:

[0064] Electrical risks: open circuit, short circuit, overload, insulation breakdown, and partial discharge, etc.

[0065] Mechanical risks: loose parts, worn bearings, and jamming of operating mechanisms, etc.

[0066] S42: For each possible risk of multiple electrical devices, the “causal indicators” (risk triggers) and “effect indicators” (risk consequences) are broken down by using expert experience, fault tree analysis and machine learning combined with historical risk types.

[0067] Because the indicators are the precursors that directly lead to the occurrence of risks, that is, changes in the indicators cause the risks to occur in the equipment. For example, the indicators for "overload risk" are: current continuously exceeding the rated value by 1.2 times (electrical monitoring indicator data), cooling system fan speed below the threshold (equipment status indicator), and the ambient temperature of the equipment exceeding the set temperature; the indicators for "insulation aging risk" are: the equipment's service life exceeding the set service life and partial discharge exceeding the standard.

[0068] The result indicator is the subsequent change in the indicator 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 vibration frequency abnormally increases the energy within the set range.

[0069] S43: Construct and verify a directed graph of causal relationships:

[0070] After identifying the indicators and risk types, a directed graph of causal relationships is constructed using directed graphs to establish the causal relationships between the indicators, risk types, and the causal relationships between them. Specifically:

[0071] Using "risk type" as the core node, directed edges are used to connect causal and effect indicators, and causal strength values ​​are labeled.

[0072] The probability of "cause indicator combination → risk occurrence → effect indicator change" is statistically analyzed by historical data, and the causal weights are optimized by Bayesian network.

[0073] During the causal screening process, if the causal strength between certain premise attributes and conclusions is less than a set threshold, edges with this causal strength are deleted, forming a directed graph of causal relationships with optimized structure.

[0074] S5: Construct rules and initialize confidence based on the causal relationship directed graph to obtain a causal rule base. Match the input data based on the causal rule base to identify and output the security risks of the device.

[0075] S51: Define the rule format: Each rule corresponds to a security risk judgment logic, with the structure "[due to indicator condition 1] and [due to indicator condition 2], "[Due to indicator conditions n], where n is the total number of indicator conditions, it is determined as [safety risk type], and the confidence level is [initial value]". For example, Rule 1 (overload risk): "If the current is greater than the set value and the rated current and the fan speed is less than the set value, it is determined as an overload risk, and the confidence level is the calculated value". The set value is determined using the monitoring parameters of the power equipment on site. The set value represents an uncertain value and is confirmed according to the specific usage scenario and specific on-site indicators.

[0076] Confidence initialization: The initial confidence level is determined by combining expert scores and historical data statistics. For example, if historical data shows that "when condition 1 is met, the actual probability of overload risk occurring is a set value", then the initial confidence level can be set to the set value.

[0077] S52: Match the input data against the rule base to identify potential risks:

[0078] The input data is iterated through the causal rule base to verify the causal indicator conditions of the rules;

[0079] The results indicators are used to assist in the verification. If the corresponding results indicator is synchronously abnormal after the rule determines the risk, the overall confidence level will be increased by a certain amount. The amount here is determined based on on-site experience.

[0080] Based on the set threshold, risk indicators with confidence levels higher than the set threshold are filtered out, and the risk identification results are output:

[0081] The output risk information includes: risk type, occurrence time, related causal indicators and effect indicators, and overall confidence level;

[0082] Furthermore, based on the set confidence level warning threshold, when the overall confidence level exceeds the set warning threshold, a real-time warning is triggered and pushed to the control terminal.

[0083] Because the operating status and risk factors of power equipment are dynamically changing, the initial confidence rule base is difficult to adapt to these changes over a long period. Specifically, during long-term operation, power equipment may experience aging, wear, load fluctuations, or changes in environmental factors (such as seasonal changes and temperature and humidity fluctuations), leading to a shift in the causal relationship between risks and monitoring indicators. For example, the initial confidence level of the association between "excessive current" and "overload risk" for new equipment may be high, but as the equipment insulation ages, the probability of risk under the same current conditions will increase, and the confidence level of the initial rule will no longer reflect the actual risk level. In addition, new failure cases, adjustments to operation and maintenance strategies, or changes in sensor accuracy can also cause data distribution drift. For example, a newly added vibration sensor may collect more subtle bearing wear signals, making the original confidence level of the "vibration indicator and wear risk" rule no longer accurate. Therefore, it is necessary to update the confidence level in the rule base with new collected data so that the causal rule base can capture these changes. Therefore, step S6 is also included.

[0084] S6: Update the confidence level of the causal rule base using the gradient update algorithm. Step S6 specifically includes:

[0085] With the goal of improving the accuracy of risk identification, we define the gradient of the loss function, set the learning rate, the maximum number of iterations, and the convergence threshold. For each rule in the causal rule base, we calculate the gradient of the loss function of each rule with respect to the initial confidence, and update the confidence in the opposite direction of the gradient.

[0086] Input data: multi-dimensional time-series data of equipment monitoring indicators , Let d be the monitoring index features at time t;

[0087] The rule base contains N rules. , For the j-th rule The initial confidence level of the rule, , These are rules and conditions;

[0088] The true risk label at time t ;

[0089] The specific calculation process is as follows:

[0090] Step 1: Calculate the activation weight of each rule in the rule base:

[0091] For the input at time t Calculate the activation weight for each rule. ;

[0092] = ,in, For the sigmoid function, map the similarity to ;

[0093] Step 2: Risk prediction output, predict the risk probability of the input data against the data in the rule base, and its predicted value is the weighted sum of the confidence levels of each rule. The specific calculation is as follows: = ;

[0094] Step 3: Use the cross-entropy loss function to measure the difference between the predicted risk and the true risk. Its calculation formula is:

[0095] ;

[0096] where is the regularization term, which is L2 regularization and is set as the penalty term for the model complexity to prevent overfitting in the process of parameter update. is the regularization coefficient. According to the specific operation process of the device, as the device ages over the years, the risk values in the rule base are adjusted, and it is set according to the expert experience value. is the j-th rule in the rule base The initial confidence level of the rule, is the regularization coefficient, is the true risk value at time t, is the predicted risk value at time t, and T is the total number of time steps;

[0097] Step 4: For each rule, calculate the partial derivative (gradient) of the loss function L with respect to its initial confidence level . The simplified calculation formula is:

[0098] ;

[0099] Step 5: Perform iterative updates. Update the confidence level according to the learning rate and the set maximum number of iterations. The updated value of the confidence level is: , and output the updated confidence level;

[0100] By continuously updating the confidence level of the rule base with new data, the rule base can dynamically capture these changes generated during the long-term use of the device, correct the deviations caused by the evolution of the device state, environmental interference, or data drift, ensure that the confidence level of risk identification always matches the actual operation scenario, and thus maintain a high-accuracy risk warning ability in the long term, avoiding misjudgment or missed judgment caused by rule固化, and providing continuous and reliable technical support for the safe operation of power equipment.

[0101] In one implementation, the elements of the temporal feature sequence and the spatial feature sequence are regulated in both time and space to facilitate subsequent feature fusion. In the prior art, when the DTW algorithm is used to adjust the temporal sequence, the feature vector data extracted by TDN is a two-dimensional vector feature, while the temporal feature sequence is a one-dimensional vector. Therefore, it is necessary to unify the dimensions of the spatial feature and the temporal feature vector sequence before feature fusion. In this case, the dimension of the temporal feature is expanded by using 1D-CNN or fully connected layers. For example, a 3×3 convolution kernel is used to extract the local temporal pattern from the pseudo-two-dimensional matrix, and the number of output channels is set to d (consistent with d of the spatial feature).

[0102] Linear projection or semantic embedding is used to ensure that the temporal features after dimensionality increase are strictly uniform in dimensionality;

[0103] The specific steps for achieving spatiotemporal alignment using the DWT algorithm include:

[0104] Step 1: Feature Vectorization: Represent the waveform spatial features extracted by the TDN network as a sequence X. Each element It is an n-dimensional vector representing the characteristics of the waveform at different locations;

[0105] The temporal features extracted by 1D-CNN are represented as sequence Y. Each element It is an m-dimensional vector representing the temporal features within a time window, from;

[0106] Step 1: Calculate the distance matrix: Construct a The distance matrix D, and the elements D of the distance matrix D. Representing spatial eigenvectors With time series feature vectors The Euclidean distance between them;

[0107] Step 3: Create the cumulative cost matrix C: And initialize it;

[0108] Step 4: Dynamically program to fill in the cumulative cost matrix:

[0109] For i=1 to m, j=1 to n, fill the cumulative cost matrix C, where the formula for calculating any element of the cumulative cost matrix C is:

[0110] C D ;

[0111] By recursively calculating, we find the starting point. To the current point The minimum cumulative path cost;

[0112] Step 5: Backtracking the optimal path, starting from the bottom right corner of the matrix and backtracking to the top left corner to determine the optimal path;

[0113] Its backtracking rule is: if And j ,choose The direction corresponding to the minimum value;

[0114] If the boundary is reached, then trace back along the boundary to the starting point;

[0115] Finally, path P is obtained. This is used to represent the optimal alignment between spatial and temporal features;

[0116] Step 6: Based on the optimal path P, map the spatial vector sequence X and the temporal feature vector Y into the same space to generate the aligned feature sequence.

[0117] By dynamically unifying the temporal and spatial feature data along the time axis and using a specific optimal alignment path, the temporal and spatial vectors are spatiotemporally unified, thus facilitating subsequent feature fusion.

[0118] When using the DTW algorithm to spatially align spatial and temporal features, DTW may over-distort the sequence (i.e., mapping one time point to multiple time points), resulting in distortion of temporal information. In the detection of power signals, this distortion may change the temporal relationship of transient events (e.g., a spike signal is stretched), thereby affecting the risk identification judgment derived from subsequent power signal detection.

[0119] Because data from different regional power grids in the power industry is difficult to share directly due to privacy protection and management boundary restrictions, the original rule base is built only based on single-region data. This makes it susceptible to regional specificities (such as climate, equipment model, and load characteristics), leading to a decrease in the accuracy of rules when applied across regions (e.g., rules from humid southern regions have a high misjudgment rate in dry northern regions). Therefore, to improve the regional applicability of this solution, in one implementation, local rule parameters from multiple regional power grids are aggregated using a federated learning framework, and a weighted average algorithm is used to generate the relationship between monitoring indicators and security risks. Specifically, this includes:

[0120] The local rule parameters are initialized and trained. Each regional power grid independently constructs an initial causal rule base based on local historical fault data, equipment operation records and expert experience, and determines the parameters of the local rules (such as the correlation weight between causal indicators and risks, and the initial confidence value).

[0121] Local power grids encrypt local rule parameters through federated learning clients, requiring the upload of encrypted rule parameters to ensure data privacy and security in each region. The federated learning server receives the encrypted parameters from each region, forming a parameter set. Federated learning is a distributed machine learning framework that allows multiple participants to collaboratively train a model without sharing raw data. In power safety risk operation and maintenance, the federated learning mechanism can be used to expand the rules governing the occurrence of more risk types.

[0122] The server determines the weighting coefficients based on the data weights of each region, equipment type coverage, and historical rule accuracy. It then uses a weighted average algorithm to aggregate the encrypted parameters. The specific weight ratios are calculated according to the specific relationships of the power grids in each region, and are determined specifically for the parameter values ​​of each power grid. For example, for the association weight of "temperature and equipment overheating risk", the weight of North China is 0.3 (small data volume but high accuracy), and the weight of East China is 0.5 (large data volume). After aggregation, the global association weight is generated as follows: 0.3 × North China parameter + 0.5 × East China parameter + weighted parameters of other regions.

[0123] After being decrypted, the aggregated global parameters are distributed to the power grid clients in each region. Each regional power grid client adjusts its local rule base based on the global parameters to generate monitoring indicators and security risk relationships that integrate features from multiple regions.

[0124] By using federated learning of power grids in various regions, rule parameters can be aggregated without sharing data between regions, which complies with the power industry's data security standards and avoids the leakage of sensitive information in each 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 the high load scenario in the northern winter and also cover the load characteristics in the humid environment of the southern summer, improving the accuracy of cross-regional applications.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for identifying safety risks of power work equipment with multimodal input, characterized in that: Comprise the following steps: S1: real-time acquisition of the monitoring index data of the equipment, the monitoring index data comprising electrical monitoring index data, the electrical monitoring index data comprising current and voltage data, the electrical monitoring index data being converted into a waveform graph, and spatial features of the waveform graph being extracted using a TDN network; S2: time sequence features of the electrical monitoring index data being extracted using a 1D-CNN network, and the spatial features and the time sequence features being subjected to feature space alignment; S3: the spatial features and the time sequence features after the space alignment being subjected to feature fusion to obtain input features of the electrical monitoring index data; S4: a causal relationship between historical equipment monitoring index data and safety risks being acquired to construct a causal relationship directed graph; S5: rules being constructed based on the causal relationship directed graph and confidence being initialized to obtain a causal rule base, input data being matched based on the causal rule base, safety risks of the equipment being identified and output; Step S1 comprises the following sub-steps: Before the waveform graph conversion, wavelet packet transformation is performed on the electrical monitoring index data, the signal is decomposed into different frequency bands, a high sampling rate is used for the high-frequency part to be processed separately, and microsecond-level transient features are retained; A time-frequency feature graph based on short-time Fourier transform is constructed, and an attention mechanism is set to automatically identify and enhance the high-frequency transient region; S2 comprises the following sub-steps: A fully connected layer of the 1D-CNN is used to output time sequence feature sequences of the same dimension as the spatial features; A distance matrix D of dimension is constructed, the elements D of which represent the Euclidean distance between a spatial feature vector and a temporal feature vector; A cumulative cost matrix C is created, C has dimensions and is initialized; A cumulative cost matrix is filled, recursive calculation is performed to determine an optimal alignment path, and the spatial features and the time sequence features are subjected to feature space alignment based on the optimal alignment path. 2.The power operation equipment safety risk identification method of multi-modal input according to claim 1, characterized in that: In S3, a fusion function implementing feature fusion The formula is: ； wherein, is a one-dimensional time series signal, is a one-dimensional time series feature extraction function, is a converted waveform image, is a two-dimensional image feature extraction function. 3.The power operation equipment safety risk identification method of multi-modal input according to claim 1, characterized in that: S6 is further included, a gradient update algorithm is used to update the confidence of the causal rule base, and specifically comprises: For each rule in the causal rule base, the gradient of the loss function of each rule with respect to its initial confidence is calculated, and the confidence is updated in the opposite direction of the gradient; Based on the set update constraint condition, the confidence is iteratively updated until the maximum number of iterations is reached, and the updated confidence is output. 4.The power operation equipment safety risk identification method of multi-modal input according to claim 1, characterized in that: A cross-entropy loss function is used to calculate the difference between the predicted risk and the real risk of the rule base, and the calculation formula is: ； where, is the initial confidence of the jth rule in the rule base is the regularization term, L2 regularization, to prevent overfitting of the loss function, is the regularization coefficient, is the true risk value at time t, is the predicted risk value at time t, T is the total number of time steps.​ 5. The method of claim 1, wherein: In S4, the causal relationship comprises a plurality of causal relationships corresponding to a plurality of safety risk types, and the causal relationship corresponding to any safety risk comprises a plurality of cause indicators and a plurality of effect indicators corresponding to the safety risk, a change in the cause indicator causing the corresponding safety risk, and a plurality of effect indicators changing after the safety risk occurs.

6. A multi-modal input power operation equipment safety risk identification system, applying the multi-modal input power operation equipment safety risk identification method of any one of claims 1-4, comprising: a spatial feature extraction module for real-time acquisition of the monitoring index data of the equipment and conversion of the acquired electrical monitoring index data into a waveform graph, and extraction of spatial features of the waveform graph using a TDN network; a time sequence feature extraction module for extraction of time sequence features of the electrical monitoring index data using a 1D-CNN network; a feature fusion module for feature space alignment and feature fusion of the spatial features and the time sequence features.

Images