A distribution network off-network device interoperability risk perception method and related device

Abstract

The application provides a power distribution network external device interoperability risk perception method and related device, and belongs to the field of power distribution network intelligent perception and device safety monitoring. The method first converts the standardized multi-modal data obtained through preprocessing into a feature sequence of a unified dimension through multi-modal feature coding. Then, through cross-modal asynchronous alignment processing, the timing precision matching of multi-modal features is completed, and fine modeling of different modal timing asynchronous characteristics is realized. Subsequently, an explicit association topology between multi-modal features is constructed through graph fusion risk representation, and the interactive fusion of multi-source information is completed, solving the core problem that traditional methods cannot quantitatively represent the association relationship between modalities and device states. Finally, based on the fusion risk representation of sufficient complementary information, the risk decision result is output, effectively improving the state recognition accuracy, risk warning reliability and emergency auxiliary decision support capability of the power distribution network external device, and enhancing the explainability and engineering application value of the risk perception result.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensing and equipment safety monitoring technology for power distribution networks, specifically relating to a method and related device for sensing interoperability risks of equipment outside the power distribution network. Background Technology

[0002] With the continuous advancement of the construction of new power systems, a large number of off-grid devices such as distributed photovoltaic systems, energy storage devices, electric vehicle charging piles, user-side distributed power terminals, and low-voltage monitoring equipment are being connected to the distribution system. These devices are distributed at the edge of the power grid or at the end of the user side, and are an important component of the power system. Their operating status has a significant impact on the stability and reliability of the entire power system.

[0003] With the integration of numerous external devices, the topology, power flow distribution, operating status, and risk transmission relationships of the power distribution system become more complex. Furthermore, traditional risk perception methods for power distribution networks largely rely on single electrical quantity data such as voltage, current, and power, and these data exhibit significant differences across different modes.

[0004] Most existing multimodal fusion methods focus on feature-level splicing or unified encoding, lacking accurate modeling of the asynchronous characteristics of different modes in time series, and also making it difficult to explicitly express the correlation between different modes and different device states. As a result, in complex multi-source data environments, the accuracy of status identification of devices outside the distribution network, the effect of risk warning, and the ability to assist emergency decision-making are insufficient. Summary of the Invention

[0005] Based on this, it is necessary to provide a method and related device for risk perception of interoperability between distribution network external equipment with multi-source asynchronous alignment capability and cross-modal correlation modeling capability to address the above-mentioned technical problems, so as to solve the problems that traditional distribution network risk perception methods lack accurate modeling of the asynchronous characteristics of different modes and have difficulty in explicitly expressing the correlation between different modes and equipment states.

[0006] In a first aspect, the present invention provides a method for sensing interoperability risks of devices outside the distribution network, comprising the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0007] Furthermore, the multimodal raw operating data of external distribution network equipment is preprocessed to obtain standardized multimodal data samples, including: Data cleaning is performed on the original multimodal operating data to obtain effective multimodal data after removing noise and outliers; Timestamp correction and data slicing are performed on the effective multimodal data to obtain standardized multimodal data samples with a unified time index.

[0008] Furthermore, multimodal feature encoding is performed on the standardized multimodal data samples to obtain a unified-dimensional multimodal feature sequence, including: For standardized multimodal data samples, single-modal feature encoding is performed on the data samples of each modality to obtain independent feature vectors for each modality. By using a linear mapping layer, the independent feature vectors of each modality are projected onto the latent feature space of the same dimension, resulting in a multimodal feature sequence of a unified dimension.

[0009] Furthermore, cross-modal asynchronous alignment processing is performed using the multimodal feature sequences to obtain temporally aligned multimodal features, including: Based on the original sampling frequency and timestamp information of the multimodal feature sequences, multimodal feature sequences with different sampling frequencies are mapped to a time window of uniform length to obtain resampled modal sequences under a uniform time window; Based on the validity of the resampled mode sequence within each time window, a frequency mask matrix is ​​generated to identify the validity of the data at each time point; By combining the resampled modal sequence with the frequency mask matrix for temporal correlation reconstruction, we obtain temporally aligned multimodal features with a unified time scale.

[0010] Furthermore, graph fusion risk characterization processing is performed on the multimodal features to obtain a fusion risk characterization, including: Heterogeneous graph structure construction is performed on the multimodal features to obtain an association graph structure with modal features and device states as nodes and modal association weights as edge weights; By performing graph neural network message passing processing on the association graph structure, a fusion risk representation after multimodal information interaction and fusion is obtained.

[0011] Furthermore, based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network, including: Based on the risk characterization of the integration, risk classification and risk level assessment are carried out respectively to obtain the risk category results and risk level results corresponding to equipment interoperability; By combining risk category results, risk level results, and equipment operation scenarios, an interpretable risk cause description is generated and matched with corresponding emergency response strategies, resulting in a risk perception result that includes risk causes and response recommendations.

[0012] Furthermore, the method also includes a joint model training optimization step, including: Based on multimodal sample data labeled with risk categories, risk levels, and risk causes, we construct temporal reconstruction loss, cross-modal alignment loss, graph structure association preservation loss, risk category identification loss, risk level assessment loss, and risk cause generation loss, respectively. The multi-task joint loss function is obtained by weighting and combining the various loss terms. By using a multi-task joint loss function to iteratively optimize model parameters, a risk perception model for interoperability of distribution network external equipment is obtained, which can perform multimodal feature encoding, cross-modal asynchronous alignment, graph fusion risk characterization, and risk decision output processing.

[0013] Secondly, the present invention provides a device for sensing interoperability risks of external equipment in a distribution network, comprising: The data preprocessing module is used to preprocess the multimodal raw operating data of the equipment outside the distribution network to obtain standardized multimodal data samples; The feature encoding module is used to perform multimodal feature encoding processing on standardized multimodal data samples to obtain a multimodal feature sequence with a unified dimension; The temporal alignment module is used to perform cross-modal asynchronous alignment processing on multimodal feature sequences to obtain temporally aligned multimodal features; The risk characterization module is used to perform graph fusion risk characterization processing on multimodal features to obtain a fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate a fused risk characterization. The results output module is used to process risk decisions based on the integrated risk characterization and obtain the risk perception results of interoperability between equipment outside the distribution network.

[0014] Thirdly, the present invention provides a computer device, the device including a processor and a memory: The memory is used to store computer programs and send the instructions of the computer programs to the processor; The processor executes the following steps according to the instructions of the computer program: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0015] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0016] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0017] In summary, this invention provides a method and related apparatus for risk perception of interoperability between distribution network external devices. The method first converts preprocessed standardized multimodal data into a unified-dimensional feature sequence through multimodal feature encoding, eliminating the dimensional differences in heterogeneous modal data to establish a fusion foundation. Then, it completes precise temporal matching of multimodal features through cross-modal asynchronous alignment processing, achieving refined modeling of the asynchronous temporal characteristics of different modalities and directly compensating for the weakness of traditional methods in temporal adaptability. Subsequently, it constructs an explicit association topology between multimodal features through graph fusion risk representation and completes the interactive fusion of multi-source information, solving the core problem that traditional methods cannot quantify the association between modalities and device states. Finally, based on a fully complementary fusion risk representation, it outputs risk decision results, effectively improving the accuracy of status identification, the reliability of risk warning, and the emergency auxiliary decision support capability of distribution network external devices, enhancing the interpretability and engineering application value of risk perception results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for perceiving interoperability risks of external distribution network devices in one embodiment of the present invention. Figure 2 This is a schematic diagram of the cross-modal asynchronous alignment processing in one embodiment of the present invention; Figure 3 This is a schematic diagram of the process for graph fusion risk characterization in one embodiment of the present invention; Figure 4 This is a schematic diagram of the risk decision output processing in one embodiment of the present invention; Figure 5 A flowchart of a preferred embodiment of the method for sensing interoperability risks of external equipment in a distribution network according to the present invention; Figure 6 This is a block diagram of a distribution network external equipment interoperability risk sensing device according to one embodiment of the present invention; Figure 7 This is a block diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] It should be noted that the user information (including but not limited to user images, user portrait information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0023] In one embodiment, such as Figure 1 As shown, a method for sensing interoperability risks of devices outside the distribution network is provided, including the following steps: S101: Preprocess the multimodal raw operating data of the equipment outside the distribution network to obtain standardized multimodal data samples.

[0024] Among them, multimodal raw operating data refers to heterogeneous raw data collected from different monitoring dimensions of equipment outside the distribution network, while standardized multimodal data samples refer to a set of multimodal data that has been uniformly cleaned, corrected and formatted and has consistent data specifications and time indexes.

[0025] Optionally, raw device operation data covering five dimensions—electrical, visual, temperature, acoustic, and text logs—is collected, and noise removal, outlier handling, missing value imputation, format unification, and time base correction are performed sequentially to output standardized data samples that meet the input requirements of feature encoding.

[0026] S102: Perform multimodal feature encoding on standardized multimodal data samples to obtain a multimodal feature sequence with a unified dimension.

[0027] Among them, multimodal feature encoding refers to the process of mapping the original data of different modalities to feature vectors in a high-dimensional latent space, while a unified-dimensional multimodal feature sequence refers to a set of temporal features that have the same feature dimension after the features of each modality are aligned.

[0028] Optionally, for each modality, an adapted encoder structure is used to extract single-modal depth features, and then a dimension mapping layer is used to transform all modal features to a feature space of the same dimension to form a multimodal feature sequence.

[0029] S103: Perform cross-modal asynchronous alignment processing using multimodal feature sequences to obtain time-aligned multimodal features.

[0030] Among them, cross-modal asynchronous alignment refers to the process of unifying the time scale and filling in missing positions for the asynchronous characteristics of different modal sampling frequencies and time steps. The multimodal features after time alignment refer to the set of multimodal features that have one-to-one corresponding time steps under the same time window.

[0031] Optionally, the multimodal feature sequences are resampled based on a unified time window to identify valid samples and missing positions. The feature reconstruction of low-frequency / missing positions is completed by combining the correlation between modalities, and finally, aligned features with strict synchronization in the time dimension are output.

[0032] S104: Perform graph fusion risk characterization processing on multimodal features to obtain fusion risk characterization; graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fusion risk characterization.

[0033] Among them, graph fusion risk characterization refers to the risk feature vector generated by modeling the relationships between modalities, devices, and states using graph structures and through information interaction via graph neural networks. This vector combines single-modal anomaly features with cross-modal correlation information.

[0034] Optionally, the aligned modal features, device entities, and operating states are abstracted into graph nodes, and a heterogeneous graph is constructed with modal association strength as edge weights. Information interaction and feature updates between nodes are realized through the message passing mechanism of graph neural networks, and a risk representation that integrates multi-source association information is output.

[0035] S105: Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0036] Among them, risk decision output refers to the entire process of risk identification, level determination and disposal reasoning based on integrated risk representation, while the risk perception result of interoperability between distribution network equipment refers to a complete set of risk outputs including risk category, risk level, risk cause and disposal suggestions.

[0037] Optionally, the risk representation input can be integrated into a multi-branch decision network to complete risk category identification, risk level assessment and risk cause decoding respectively. Then, combined with equipment scenario matching emergency response strategies, the complete risk perception and auxiliary decision-making results can be output.

[0038] The method provided in this embodiment first converts the preprocessed standardized multimodal data into a unified-dimensional feature sequence through multimodal feature encoding, eliminating the differences in data dimensionality between heterogeneous modalities to build a fusion foundation. Then, it completes the precise temporal matching of multimodal features through cross-modal asynchronous alignment processing, realizing refined modeling of the asynchronous temporal characteristics of different modalities, directly making up for the weakness of traditional methods in temporal adaptation. Subsequently, it constructs an explicit association topology between multimodal features through graph fusion risk representation and completes the interactive fusion of multi-source information, solving the core problem that traditional methods cannot quantify the association between modalities and device states. Finally, it outputs risk decision results based on the fully complementary fusion risk representation, effectively improving the accuracy of status identification of equipment outside the distribution network, the reliability of risk warning, and the emergency auxiliary decision support capability, and enhancing the interpretability and engineering application value of risk perception results.

[0039] In an exemplary embodiment, the multimodal raw operating data of external distribution network devices are preprocessed to obtain standardized multimodal data samples, including: S201: Perform data cleaning on the original multimodal operating data to obtain effective multimodal data after removing noise and outliers.

[0040] Among them, data cleaning refers to the preprocessing operation of filtering out and repairing noise, abnormal spikes and missing values ​​in the original collected data, and multimodal effective data refers to the multimodal data set that retains effective information and removes invalid interference after cleaning.

[0041] For example, data collection covers distribution transformer areas, user-side energy consumption scenarios, new energy grid connection scenarios, and energy storage operation scenarios. The monitoring objects include distributed photovoltaic grid-connected devices, energy storage battery cabinets, electric vehicle charging piles, low-voltage distribution terminals, user-side distributed power consumption terminals, and edge monitoring devices, as well as other equipment outside the distribution network. The data dimensions include five categories: electrical quantities, images, temperature, sound, and text logs. For the raw data, noise removal, outlier handling, missing value imputation, normalization, and unified formatting are performed sequentially. For electrical quantity data, median filtering or moving average is used to remove abnormal spikes, linear interpolation, forward padding, or prediction based on historical adjacent windows is used to impute missing data, and Z-score standardization is used to normalize the values. For image data, the image is uniformly adjusted to 256×256 pixels, brightness correction and contrast enhancement are performed, and data enhancement is achieved through random cropping, rotation, and horizontal flipping. For temperature data, temperature rise features are extracted according to key parts of the equipment, and thermal imaging images are uniformly mapped to the same spatial size as visible light images. For audio data, the image is uniformly resampled to 16kHz, and a Hamming window with a window length of 25ms and a frame shift of 10ms is used to extract the Mel spectrum. For text log data, word segmentation, stop word removal, keyword extraction, and unified encoding are performed, with the maximum text length set to 128 tokens.

[0042] It should be noted that the noise characteristics of different modalities differ significantly from those of missing modes, so a modality-adaptive cleaning strategy is required. In practice, the size of the filtering window and the missing data completion strategy can be adjusted according to the actual sampling quality of the sensors on site. In scenarios with sufficient data, samples with excessively high missing data rates can also be discarded directly.

[0043] It is understandable that text log data is unstructured data, and its cleaning process is fundamentally different from that of structured time-series data. It needs to use natural language processing technology to complete semantic extraction and encoding in order to adapt to the subsequent feature fusion requirements.

[0044] S202: Perform timestamp correction and data slicing on the effective multimodal data to obtain standardized multimodal data samples with a unified time index.

[0045] Among them, timestamp correction refers to the processing operation of unifying the time base of different modal data to the same clock source, data slicing refers to the processing of segmenting time series data according to a fixed time window, and standardized multimodal data samples with unified time index refers to a set of multimodal data samples with consistent time base and slicing granularity.

[0046] For example, firstly, the timestamps of all modal data are clock-synchronized to eliminate clock skew between different sensors; then, a sliding window method is used to slice the time-series data, with a window length of 30 seconds and a sliding step size of half the window length; preliminary features are extracted for each modal data: dynamic features such as voltage fluctuations, power changes, frequency shifts, current abrupt changes, and harmonic anomalies are extracted from the electrical quantity dimension; visual features such as equipment appearance, smoke, alarm lights, wiring status, and local abnormal areas are extracted from the image dimension; features such as local temperature rise, average temperature, maximum temperature, temperature change rate, and temperature gradient are extracted from the temperature dimension; features such as spectrograms, frequency domain energy, abnormal frequency band energy, and short-term energy change features are extracted from the sound dimension; and features such as text logs are extracted from the text log dimension. The device name, alarm type, time information, risk keywords, and handling records are extracted and encoded into text semantic vectors. Finally, a multimodal dataset containing electrical quantity time series samples, image samples, temperature samples, sound samples, and text log samples is constructed and divided into training set, validation set, and test set in a 7:1:2 ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and early stop judgment, and the test set is used to evaluate the model's generalization performance. The sample library contains no less than 100,000 electrical quantity time series samples, no less than 10,000 image samples, and corresponding temperature, sound, and text samples. All samples are labeled with device status, risk category, risk level, and risk cause description tags, and a unified timestamp index, device number, and scene tag is established for each sample.

[0047] It should be noted that sliding window slicing is a common method for time series data processing. The window length and step size need to be adapted according to the time scale of the risk event. A 30-second window can cover the development process of most abnormal events of equipment outside the distribution network, while a 1 / 2 step size can increase the sample density while ensuring data independence.

[0048] In practice, the dataset division ratio can be adjusted according to the actual monitoring scenario. In small sample scenarios, cross-validation can be used to improve the stability of model training.

[0049] This embodiment can convert multi-source heterogeneous raw acquisition data into a standardized sample library with unified format, time alignment, and complete labels, eliminating data format differences and time reference deviations between different modalities.

[0050] In an exemplary embodiment, standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence of uniform dimension, including: S301: For standardized multimodal data samples, perform single-modal feature encoding on the data samples of each modality to obtain independent feature vectors for each modality.

[0051] Among them, single-modal feature encoding refers to the process of mapping the original data into high-dimensional deep feature vectors by adopting an encoder structure adapted to the characteristics of single-modal data. Independent feature vectors refer to the feature vectors that represent the semantic information of the corresponding modality, output by each modality through its own encoder.

[0052] For example, the specific implementation process of this step is as follows: (1) Electrical quantity mode coding: The sliced ​​electrical quantity time series data is normalized and then input into the time series encoder. The time series encoder can adopt a one-dimensional convolutional layer, a Transformer time series encoder or a gated cyclic unit structure to map the time series data of different dimensions such as voltage, current, active power, reactive power, frequency, and power factor into a time series feature vector of a unified dimension. .

[0053] (2) Image modality coding: The preprocessed 256×256 pixel image is input into a visual encoder. The visual encoder can be ViT, ResNet or a lightweight convolutional neural network to extract visual semantic features such as equipment appearance, local temperature rise, smoke, alarm light status and wiring abnormalities, and output image feature vectors. .

[0054] (3) Temperature modal coding: For point temperature data, statistical features are extracted first; for thermal imaging data, spatial size mapping is completed first, and then the data is input into the temperature feature encoder. The encoder can be a convolutional neural network, a visual Transformer, or a temporal coding network, and outputs a temperature feature vector. .

[0055] (4) Sound modality coding: The generated Mel spectrogram is input into the sound feature encoder. The encoder can use a convolutional neural network or a visual Transformer structure shared with the image modality for embedding modeling. After converting the one-dimensional sound signal into two-dimensional spectral features, deep coding is completed, and the sound feature vector is output. This method allows sound features and image features to be distributed in similar feature spaces, reducing the difficulty of subsequent cross-modal fusion.

[0056] (5) Text log modal coding: The preprocessed text sequence is input into the text encoder. The encoder can use BERT, general Transformer encoder or power field-specific language model to perform semantic modeling on equipment alarm records, dispatch instructions, operation logs, maintenance records and emergency response records, and output a global semantic vector. It is used to characterize information such as alarm type, risk keywords, handling process and operational status description.

[0057] It should be noted that the differences in data structure among different modalities determine the differences in encoder selection: time series data focuses on time dependency modeling, image data focuses on spatial semantic extraction, and text data focuses on contextual semantic understanding.

[0058] Understandably, the audio modality adopts a spectrogram + visual encoder approach, which can reuse mature visual feature extraction architectures, reduce the number of model parameters, and improve feature compatibility.

[0059] S302: The independent feature vectors of each modality are projected onto the feature latent space of the same dimension using a linear mapping layer to obtain a multimodal feature sequence of the same dimension.

[0060] Among them, the linear mapping layer refers to the feature dimension transformation structure composed of fully connected layers, the feature latent space refers to the unified high-dimensional feature space to which all modal features are mapped, and the unified dimension multimodal feature sequence refers to the set of temporal features with the same feature dimension for all modalities.

[0061] For example, independent linear mapping layers are set up for five modalities: electrical quantity, image, temperature, sound, and text log. The independent feature vectors output by each modality are projected onto a latent space of the same dimension. In a specific implementation scenario, the unified feature dimension can be set to 256. After mapping, the features of the five modalities are represented as follows: All feature vectors have the same dimensionality.

[0062] It should be noted that linear mapping is the simplest way to align feature dimensions. It only transforms the dimensions without changing the semantic distribution of the features, and can preserve the original information of single-modal encoding to the greatest extent. In practice, nonlinear mapping layers can also be used to improve the adaptability of the feature space, but care should be taken to control the model complexity to avoid overfitting.

[0063] This embodiment, through modal coding and unified mapping processing, can extract accurate deep features for the data characteristics of each modality, and eliminate the dimensional differences between different modalities, unifying heterogeneous features into the same feature space.

[0064] In an exemplary embodiment, cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features, including: S401: Based on the original sampling frequency and timestamp information of the multimodal feature sequences, the multimodal feature sequences with different sampling frequencies are mapped to a time window of uniform length to obtain the resampled modal sequences under the uniform time window.

[0065] Adaptive frequency sampling refers to the process of resampling feature sequences with different time granularities to a unified time window based on the original sampling frequency and timestamp of each modality. The resampled modal sequence refers to the feature sequences of each modality that have the same number of time steps after being aligned by a unified time window.

[0066] For example, please refer to Figure 2 , Figure 2 This is a flowchart of cross-modal asynchronous alignment processing. The diagram illustrates the process of aligning high-frequency and low-frequency modes after data from different modalities passes through an adaptive frequency sampler, frequency mask, multi-head attention module, and time-domain reconstruction module. Specifically, electrical quantities are high-frequency sampled data, images / thermal imaging are low-frequency sampled data, temperature is low-frequency / continuous data, sound is continuous waveform data, and text logs are event-triggered data. The original sampling frequencies of each modality differ significantly. Let the... m The feature sequence after feature encoding of each modality is: (1) In the formula, m Modal numbering, t For time indexing, an adaptive frequency sampler is constructed to map the feature sequences of different modalities to a unified time window based on the timestamps, sampling frequencies, and sampling intervals of each modality's data. T The resampled modal sequence is obtained: (2) In the formula, Sampler( ) is an adaptive frequency sampling function, which ultimately outputs a resampled multimodal sequence within the same time window.

[0067] It should be noted that inconsistent sampling frequencies among multiple sensors are a common phenomenon in industrial monitoring scenarios. High-frequency data has high time resolution but large data volume, while low-frequency data has coarse time granularity but contains key status information.

[0068] In practice, the uniform time window length can be adjusted according to the evolution speed of equipment anomalies. For transient risks, the window length can be shortened, while for slowly changing anomalies, the window can be appropriately increased to include more trend information.

[0069] S402: Based on the validity of the resampled mode sequence within each time window, generate a frequency mask matrix to identify the validity of data at each time point.

[0070] The frequency mask matrix is ​​a binary matrix used to identify whether each modal data is a real and valid sample at each time step. Valid sampling positions are marked as 1, and missing / complete positions are marked as 0.

[0071] For example, such as Figure 2As shown, during the resampling process, some time steps only contain actual samples of high-frequency modes, while low-frequency modes have no original acquired data at the corresponding positions. Therefore, a frequency mask matrix is ​​constructed. M Mark the validity of the data; when the first m Each modality within the time window t When the memory is validly sampling data, When data is missing for this modality within the time window, The frequency mask matrix and the resampled data are input together into the subsequent time series prediction model to distinguish between the real sampled data and the data to be reconstructed, so as to avoid the model misjudging the missing data as the real sampled value.

[0072] It should be noted that the masking mechanism is a standard technique for handling missing values ​​in sequence modeling. By explicitly identifying the missing locations, it can guide the model to learn features based only on real sampling points, thereby improving the accuracy of the reconstruction results.

[0073] It is understandable that the frequency mask matrix only records the valid / missing state of the data and does not contain specific feature values. Therefore, it does not introduce additional semantic information interference and only serves as a position identifier.

[0074] S403: Combine the resampled modal sequence with the frequency mask matrix for time-series correlation reconstruction to obtain time-series aligned multimodal features with unified time scale.

[0075] Temporal correlation reconstruction refers to the process of predicting and completing modal features at low-frequency / missing locations based on the correlation between modalities and guided by high-frequency modal features. Temporally aligned multimodal features refer to a feature set in which all modalities have corresponding features at each time step and are completely synchronized in the temporal dimension.

[0076] For example, firstly, dynamic feature modeling is performed on different modalities using a time-series feature extraction module. For the electrical quantity modality, features such as voltage fluctuations, power surges, frequency shifts, and current anomalies are extracted. For the temperature modality, features such as temperature rise trends, local high-temperature areas, and temperature change rates are extracted. For the image modality, features such as abnormal equipment appearance, smoke, alarm light status, and wiring anomalies are extracted. For the sound modality, features such as abnormal vibrations, abnormal discharge sounds, and changes in mechanical noise are extracted. For the text log modality, semantic features such as alarm types, risk keywords, dispatch instructions, and handling records are extracted. Then, a multi-head attention mechanism is used to calculate the temporal correlation between different modalities. The calculation process is as follows: (3) In the formula, Q , K , V These are the query vector, key vector, and value vector, respectively. d kThe feature dimension is used as the input dimension. After inputting the features of each modality into the multi-head attention module, the correlation strength between any two modalities is calculated to form a modality correlation weight matrix. A Matrix elements A ij Indicates the first i The modality and the first j The association weights of each modality in the time and semantic dimensions are considered. For example, when the current of an energy storage device rises abnormally and the temperature rises rapidly, the association weight between the electrical quantity mode and the temperature mode increases; when an image shows smoke features and the text log shows an overheating alarm, the association weight between the image mode and the text mode increases. Finally, based on the modality association weight matrix and guided by high-frequency modal data, time-domain reconstruction is performed on low-frequency modes or missing modes. Let the m-th mode be within the time window... t The true characteristics within are The reconstructed features are The reconstruction process is as follows: (4) In the formula, High-frequency modal characteristics, The correlation weights between the high-frequency mode and the target mode are... The historical state features of the target modality are used to obtain multimodal features aligned under a unified time window, enabling different modal data such as electrical quantities, images, temperature, sound, and text logs to be synchronously expressed in the time dimension, avoiding the incorrect fusion of data from different time points.

[0077] It should be noted that the multi-head attention mechanism can capture fine-grained correlations between modes from multiple representation subspaces at the same time, and has a stronger correlation modeling capability compared with the single attention mechanism; the idea of ​​using high frequency to guide low frequency reconstruction is essentially to infer the intermediate state of low resolution signal by using the changing trend of high time resolution signal.

[0078] In practice, the time-domain reconstruction module can be implemented using a gated loop unit or a Transformer decoder structure, and the appropriate type can be flexibly selected according to the reconstruction accuracy requirements.

[0079] This embodiment, through adaptive sampling, mask identification, and temporal reconstruction processing, can accurately model the temporal asynchronous characteristics of different modalities, achieve precise alignment of high-frequency and low-frequency modalities within a unified time window, and fundamentally solve the feature mismatch problem caused by temporal misalignment in traditional multimodal fusion.

[0080] In an exemplary embodiment, graph fusion risk characterization processing is performed on multimodal features to obtain a fusion risk characterization, including: S501: Perform heterogeneous graph structure construction on multimodal features to obtain an association graph structure with modal features and device states as nodes and modal association weights as edge weights.

[0081] Among them, heterogeneous graph structure refers to graph data structure containing multiple types of nodes and multiple types of edges, while associative graph structure refers to graph structure used to describe the relationships between modalities, devices, and states.

[0082] For example, please refer to Figure 3 , Figure 3 This is a flowchart of the graph fusion risk representation process. The flowchart illustrates how electrical quantity features, image features, temperature features, sound features, and text features are used as graph nodes, and modal association weights are used as edge weights to obtain a fused risk representation through a graph neural network. Specifically, after completing cross-modal asynchronous alignment, the features of each modality are constructed into a graph structure: (5) In the formula, V For a set of graph nodes, E For the graph edge set, A The graph is an adjacency matrix. Nodes are divided into three categories: the first category is modal nodes, including electrical quantity nodes, image nodes, temperature nodes, sound nodes, and text nodes, which correspond to the five aligned modal features; the second category is equipment nodes, including distributed photovoltaic equipment, energy storage equipment, electric vehicle charging piles, low-voltage distribution terminals, and other equipment outside the distribution network; the third category is status nodes, including normal operation, general risk, high risk, and emergency risk operating states. The edges in the graph represent the association relationships between different modal nodes, between equipment nodes, or between status nodes. The edge weights are determined by the modal association weight matrix obtained in the cross-modal alignment stage. Through this graph structure, the interaction relationships between electrical quantity anomalies, temperature anomalies, visual anomalies, sound anomalies, and text alarms can be explicitly expressed, as well as the risk propagation relationships between devices and the evolution relationships between states.

[0083] It should be noted that heterogeneous graphs can model multiple types of entities and multiple types of relationships simultaneously, making them more suitable for depicting complex relationship scenarios involving multiple modalities, multiple devices, and multiple states compared to homogeneous graphs.

[0084] It is understandable that directly reusing the modal association weights calculated in the cross-modal alignment stage for edge weights can ensure the consistency of association calculations and avoid errors and computational losses caused by repeated calculations.

[0085] S502: Perform graph neural network message passing processing on the association graph structure to obtain the fusion risk representation after multimodal information interaction and fusion.

[0086] Among them, graph neural network message passing refers to the computation process in graph neural networks where nodes receive information from neighboring nodes through adjacent edges and update their own features, while fusion risk representation refers to the global risk feature that integrates single-modal anomaly and cross-modal correlation information after multiple rounds of message passing.

[0087] For example, in combination Figure 3 As shown, the specific implementation process of this step is as follows: input the constructed heterogeneous graph structure into the graph neural network, and update the node features through the message passing mechanism; l The node update process of a layered graph neural network is as follows: (6) In the formula, Let A be the modal correlation weight matrix, representing the features of the l-th layer nodes. For learnable parameter matrix, σ ( () is a nonlinear activation function; after propagation through a multi-layer graph neural network, electrical quantity anomalies, image anomalies, temperature anomalies, sound anomalies, and text alarm information are fully interacted and fused within the graph structure, ultimately yielding a fused risk representation. This characterization includes not only anomalous features in a single modality, but also the correlation between different modalities and the risk propagation relationship between equipment states, which can comprehensively depict the risk status in the equipment interoperability process.

[0088] It should be noted that graph neural networks achieve neighborhood information aggregation through message passing mechanisms, making them naturally adaptable to the association modeling of graph structures. Compared to fully connected networks, they are better at capturing the relationships and dependencies between entities.

[0089] In practice, graph neural networks can use different architectures such as graph convolutional networks and graph attention networks. When it is necessary to distinguish the importance of different neighboring nodes, graph attention networks can be preferred to achieve adaptive neighborhood weight allocation.

[0090] This embodiment, through the fusion processing of heterogeneous graph construction and graph neural network, can explicitly and quantitatively express the correlation between different modalities, different devices, and different states, realize the deep interactive fusion of multi-source information, solve the core defect of traditional methods that cannot characterize the risk correlation between modalities and device states, and output a fusion risk representation that is both comprehensive and interpretable.

[0091] In an exemplary embodiment, risk decision output processing is performed based on the fused risk characterization to obtain the interoperability risk perception result of external equipment in the distribution network, including: S601: Based on the risk characterization of the integration, risk classification and risk level assessment are carried out respectively to obtain the risk category results and risk level results corresponding to equipment interoperability.

[0092] Risk classification refers to the process of identifying the specific category to which the current equipment risk belongs, while risk level assessment refers to the process of determining the severity of the current risk.

[0093] For example, please refer to Figure 4 , Figure 4 The flowchart illustrates the process of risk decision-making output processing. It shows how the fused risk representation, after passing through a risk classifier and a state decoder, outputs risk category, risk level, risk cause description, and emergency response recommendations. Combined with... Figure 4 As shown, the specific implementation process of this step is as follows: Integrating risk characterization... Input the risk classification branch and the risk level assessment branch respectively; the risk classification branch uses the Softmax classifier to output risk categories, specifically including distributed photovoltaic overload risk, reverse power backflow risk, electric vehicle charging overheating risk, energy storage device abnormal temperature rise risk, equipment misoperation risk, personal electric shock risk, power quality abnormality risk, and local voltage over-limit risk; the risk level assessment branch outputs the corresponding risk level, divided into four levels: normal, general risk, high risk, and emergency risk, to assist operation and maintenance personnel in judging the priority of risk handling.

[0094] It should be noted that the multi-branch output structure is a typical architecture for multi-task learning. It can share underlying fusion features and complete multiple prediction tasks at the same time. Compared with single-task models, it has higher parameter efficiency and task synergy gains.

[0095] In practice, the classification categories can be expanded or reduced according to the risk type of the actual distribution network scenario to adapt to the equipment configuration and risk characteristics of different distribution areas.

[0096] S602: Combining risk category results, risk level results, and equipment operation scenarios, generate an interpretable risk cause description and match corresponding emergency response strategies to obtain risk perception results that include risk causes and response suggestions.

[0097] Among them, risk cause description refers to the natural language explanation of the causes of risk, emergency response strategy refers to the standardized response plan for the corresponding risk, and risk perception result refers to the complete output result including category, level, cause, and response recommendations.

[0098] For example, in combination Figure 4As shown, the specific implementation process of this step is as follows: The risk characterization is input into the state decoder to generate risk cause description information. The state decoder outputs an interpretable risk description based on the contribution characteristics of different modes and the correlation in the graph structure. For example, when the current of the electric vehicle charging pile is continuously rising, the interface temperature is rising, and abnormal color changes appear in the on-site image, the output is "The charging interface may have poor contact or abnormal heat dissipation, leading to an increased risk of local overheating". When the voltage of the distributed photovoltaic grid connection point is rising, the active power flows in reverse, and the inverter log shows an alarm, the output is "The output of the distributed photovoltaic is higher than the local load absorption capacity, and there is a risk of reverse power backflow and voltage over-limit". When the current of the energy storage device is abnormal, the temperature rises rapidly, and is accompanied by smoke image features, the output is "The energy storage device may have early signs of thermal runaway, and immediate safety isolation and on-site verification are required". Subsequently, based on the risk category, risk level, equipment type, and operating scenario, corresponding solutions from the emergency response strategy library are matched. The strategy library includes a rule base, an expert knowledge base, and historical operation and maintenance samples, outputting specific handling suggestions. For the risk of distributed photovoltaic overload, the suggestion is to "reduce inverter output, strengthen grid connection point voltage monitoring, and activate active power limiting strategies if necessary." For the risk of electric vehicle charging overheating, the suggestion is to "suspend charging, check charging interface temperature, and investigate heat dissipation and line contact status." For the risk of abnormal temperature rise in energy storage devices, the suggestion is to "switch the energy storage system to safe mode, activate fire alarm linkage, and notify operation and maintenance personnel for on-site verification." For the risk of electric shock, the suggestion is to "immediately cut off the power supply to relevant equipment, set up a safe isolation area, notify on-site personnel to evacuate, and carry out emergency response." Ultimately, a complete closed loop from risk perception to decision support is formed.

[0099] It should be noted that the risk cause description can significantly improve the interpretability of the method, helping operations and maintenance personnel to quickly understand the risk triggers. Compared with black box models that only output classification results, it has more engineering application value. The emergency response strategy library integrates rules, expert knowledge and historical cases, which can ensure the professionalism and practicality of the response suggestions.

[0100] Understandably, the state decoder is essentially a text generation model that can convert high-dimensional risk features into natural language descriptions, and is a core component for achieving risk interpretability.

[0101] This embodiment, through multi-branch risk decision-making and response strategy matching, can simultaneously output accurate risk categories, clear risk levels, explainable risk causes, and actionable response suggestions, achieving a complete closed loop from risk perception to emergency auxiliary decision-making, effectively improving the practicality and engineering support capabilities of risk perception for equipment outside the distribution network.

[0102] In one exemplary embodiment, the method further includes a model joint training optimization step, comprising: S701: Based on multimodal sample data labeled with risk categories, risk levels, and risk causes, construct temporal reconstruction loss, cross-modal alignment loss, graph structure association preservation loss, risk category identification loss, risk level assessment loss, and risk cause generation loss respectively.

[0103] The joint loss function refers to the total loss function composed of a weighted combination of the losses of multiple sub-tasks. It is used to simultaneously optimize the model parameters of multiple sub-tasks and achieve multi-task collaborative training.

[0104] The specific implementation process of this step is as follows: To achieve collaborative optimization of asynchronous alignment, graph fusion risk identification, and risk cause generation tasks, six sub-loss functions are constructed: (1) Temporal reconstruction loss This is used to constrain the consistency between reconstructed modal features and true modal features, and its calculation method is as follows: (7) In the formula, M Indicates the number of modes. T Indicates the number of time windows. Indicates frequency mask, Represents true modal features, This represents the reconstructed modal features.

[0105] (2) Cross-modal alignment loss This is used to constrain different modal features to maintain semantic consistency under the same risk event, and its calculation method is as follows: (8) In the formula, Ω represents the set of modal pairs with temporal correlation. K sim( represents the number of modal pairs) ) represents the cosine similarity function.

[0106] (3) Loss of preserving graph structure association This is used to ensure consistency of the modal association structure before and after the constraint graph neural network update, and its calculation method is as follows: (9) In the formula, A This represents the original modal correlation weight matrix. This represents the graph association matrix updated by the graph neural network. This represents the Frobenius norm.

[0107] (4) Risk category identification loss Cross-entropy loss can be used. When the sample class distribution is unbalanced, Focal Loss is preferred. Its calculation method is as follows: (11) In the formula, This represents the predicted probability corresponding to the true category. α Represents the category balance coefficient. γ This represents the adjustment factor for the difficulty level of the samples. In this embodiment... α Set to 0.25, γ Set to 2.

[0108] (5) Risk level assessment loss Cross-entropy loss is used, and its calculation method is as follows: (12) In the formula, R Indicates the number of risk levels. Indicates the true risk level label. The model predicts the number of... r The probability of each risk level.

[0109] (6) Risk cause description generates loss The negative log-likelihood loss is used, and its calculation method is as follows: (13) In the formula, L represents the length of the generated text. This represents the t-th word. This indicates the sequence of words preceding the current word. This indicates the risk profile of the integration process.

[0110] It should be noted that the core of the multi-task loss function design is to balance the optimization weights of each sub-task and avoid a single task dominating parameter updates. The six losses correspond to the six core sub-tasks of temporal reconstruction, modality alignment, graph structure preservation, classification, rating evaluation, and text generation, which can comprehensively cover all optimization objectives of the model.

[0111] Understandably, the temporal reconstruction loss introduces a frequency mask, calculating the reconstruction error only for valid sampling locations, which can avoid the interference of pseudo-labels at missing locations with model training.

[0112] S702: The multi-task joint loss function is obtained by weighting and combining the various loss terms.

[0113] For example, the joint loss function for multiple tasks is defined as: (14) In the formula, to These are the loss weight coefficients. In this embodiment, the weights of each loss term are set as follows: λ 1 = 0.20 λ 2 = 0.20 λ 3 = 0.15 λ 4 = 0.25 λ 5 = 0.10 λ =0.10, which correspond to the temporal reconstruction loss, cross-modal alignment loss, graph structure association preservation loss, risk category identification loss, risk level assessment loss, and risk cause description generation loss, respectively.

[0114] It should be noted that the allocation of loss weights needs to be adjusted according to the importance and convergence difficulty of each task. Risk category identification, as a core task, is assigned a higher weight; time series alignment and reconstruction, as basic supporting tasks, are assigned a medium weight; and level assessment and cause generation branches are assigned relatively lower weights.

[0115] In practice, the weights of each loss function can be dynamically adjusted according to the convergence of the model training. When a certain subtask converges poorly, the corresponding weight can be appropriately increased.

[0116] S703: The model parameters are iteratively optimized using a multi-task joint loss function to obtain a risk perception model for interoperability of distribution network external equipment that can perform multimodal feature encoding, cross-modal asynchronous alignment, graph fusion risk representation, and risk decision output.

[0117] For example, the specific implementation process of this step is as follows: Model training uses the Adam optimizer, and the initial learning rate is set to 1×10. -4 The batch size was set to 32, and the training epochs were set to 100. To prevent overfitting, Dropout was implemented in the multimodal feature fusion layer and the hidden layer of the graph neural network, with a Dropout ratio of 0.3 and a weight decay coefficient of 1×10⁻⁶. -5 An early stopping mechanism is adopted to prevent overfitting, and the patience rounds are set to 10 rounds. Training stops when the validation set loss does not decrease for 10 consecutive rounds. After training, an end-to-end risk perception model for interoperability between distribution network external devices is obtained, which can sequentially complete the entire process of multimodal feature encoding, cross-modal asynchronous alignment, graph fusion risk representation and risk decision output.

[0118] It should be noted that the Adam optimizer is a mainstream optimization algorithm for deep learning training. It has the ability to adaptively adjust the learning rate, and it has a fast convergence speed and good stability. Dropout, weight decay, and early stopping are all common regularization methods in deep learning, which can effectively suppress model overfitting and improve generalization ability.

[0119] This embodiment, through multi-task joint loss construction and collaborative training optimization, can simultaneously optimize multiple sub-tasks such as temporal alignment, graph fusion, risk identification, and cause generation, achieving synergistic improvement of each module and ensuring optimal overall performance of the model in terms of temporal alignment accuracy, risk identification accuracy, and decision support capability.

[0120] Please see Figure 5 , Figure 5 This paper presents a preferred implementation scheme for a risk perception method for interoperability of equipment outside the distribution network. The scheme's process includes: collecting multimodal data, covering five modalities: electrical quantities, images, temperature, sound, and text logs; sequentially processing the data through a preprocessing module to perform denoising, completion, normalization, and timestamp correction; completing a unified latent space representation through a multimodal feature encoding module; and completing a unified time window mapping and frequency mask generation through an adaptive frequency sampling module; subsequently, entering a time series prediction and reconstruction module, using high-frequency data to guide low-frequency data to complete missing modalities; and outputting time series alignment features through a cross-modal asynchronous alignment module; then generating a heterogeneous graph containing modal nodes, device nodes, and status nodes through a graph structure construction module; completing message passing and outputting a fused risk representation through a graph neural network fusion module; and finally, outputting the risk category, level, cause, and disposal recommendations through a risk decision output module. The joint loss function and model training optimization module at the bottom of the process are integrated throughout the entire process. Through a weighted combination of reconstruction loss, alignment loss, graph structure loss, classification loss, level loss, and generation loss, end-to-end model training optimization is achieved, ensuring that all modules work collaboratively and that overall performance is optimal.

[0121] Through the above specific implementation methods, this invention can achieve unified time-series modeling and cross-modal alignment of multi-source asynchronous data in interoperability scenarios of external distribution network equipment, effectively solving the problems of inconsistent sampling frequencies, inconsistent time scales, and information fragmentation of data from different modalities; by constructing modal association graphs and equipment state graphs, it enhances the ability to express the association between different modalities and different equipment states; by using graph neural networks to achieve multimodal fusion risk representation learning, it improves the accuracy of risk identification and anomaly warning in complex power distribution scenarios; at the same time, by combining risk cause descriptions and emergency response suggestions, it enhances the interpretability and engineering application value of the method, providing effective technical support for the safety monitoring, risk warning, and intelligent operation and maintenance of external equipment in new power distribution systems.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a distribution network external equipment interoperability risk sensing device for implementing the above-mentioned distribution network external equipment interoperability risk sensing method. The solution provided by this device is similar to the solution described in the above-mentioned method. Therefore, the specific limitations in the following embodiments of the distribution network external equipment interoperability risk sensing device can be found in the limitations of the distribution network external equipment interoperability risk sensing method above, and will not be repeated here.

[0124] Please see Figure 6 This invention provides a device for sensing the interoperability risk of external equipment in a distribution network, comprising: The data preprocessing module is used to preprocess the multimodal raw operating data of the equipment outside the distribution network to obtain standardized multimodal data samples; The feature encoding module is used to perform multimodal feature encoding processing on standardized multimodal data samples to obtain a multimodal feature sequence with a unified dimension; The temporal alignment module is used to perform cross-modal asynchronous alignment processing on multimodal feature sequences to obtain temporally aligned multimodal features; The risk characterization module is used to perform graph fusion risk characterization processing on multimodal features to obtain a fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate a fused risk characterization. The results output module is used to process risk decisions based on the integrated risk characterization and obtain the risk perception results of interoperability between equipment outside the distribution network.

[0125] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0126] Reference Figure 7This invention also provides a computer device, including: a memory and a processor, and a computer program stored in the memory. When the computer program is executed on the processor, it performs the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0127] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 7 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0128] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, 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, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0129] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0130] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0131] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0132] This invention provides a computer program product, including a computer program that, when executed by a processor, performs the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; Standardized multimodal data samples are subjected to multimodal feature encoding to obtain a multimodal feature sequence with a unified dimension; Cross-modal asynchronous alignment processing is performed using multimodal feature sequences to obtain temporally aligned multimodal features; Graph fusion risk characterization processing is performed on multimodal features to obtain fused risk characterization. Graph fusion risk characterization processing is the process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between equipment outside the distribution network.

[0133] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0134] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0135] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0136] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for sensing interoperability risks of external equipment in a distribution network, characterized in that, Includes the following steps: The multimodal raw operating data of external equipment in the distribution network are preprocessed to obtain standardized multimodal data samples; The standardized multimodal data samples are subjected to multimodal feature encoding processing to obtain a multimodal feature sequence of uniform dimension; The multimodal feature sequence is used to perform cross-modal asynchronous alignment processing to obtain time-aligned multimodal features; The multimodal features are subjected to graph fusion risk characterization processing to obtain a fusion risk characterization; The graph fusion risk characterization process is a process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate the fused risk characterization. Based on the integrated risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between distribution network external equipment.

2. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, The raw multimodal operating data of external distribution network equipment is preprocessed to obtain standardized multimodal data samples, including: The original multimodal operating data is cleaned to obtain effective multimodal data after removing noise and outliers; The effective multimodal data is timestamped and sliced ​​to obtain standardized multimodal data samples with a unified time index.

3. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, The standardized multimodal data samples are subjected to multimodal feature encoding to obtain a unified-dimensional multimodal feature sequence, including: For the standardized multimodal data samples, single-modal feature encoding is performed on the data samples of each modality to obtain independent feature vectors corresponding to each modality; By using a linear mapping layer, the independent feature vectors of each modality are projected onto a feature latent space of the same dimension to obtain a multimodal feature sequence of a unified dimension.

4. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, Cross-modal asynchronous alignment processing is performed using the multimodal feature sequence to obtain temporally aligned multimodal features, including: Based on the original sampling frequency and timestamp information of the multimodal feature sequences, the multimodal feature sequences with different sampling frequencies are mapped to a time window of uniform length to obtain a resampled modal sequence under a uniform time window; Based on the validity of the resampled mode sequence within each time window, a frequency mask matrix is ​​generated to identify the validity of data at each time point; By combining the resampled modal sequence with the frequency mask matrix for temporal correlation reconstruction, a time-aligned multimodal feature with a unified time scale is obtained.

5. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, The multimodal features are subjected to graph fusion risk characterization processing to obtain a fusion risk characterization, including: The multimodal features are processed by heterogeneous graph structure construction to obtain an association graph structure with modal features and device states as nodes and modal association weights as edge weights; The graph structure is processed by graph neural network message passing to obtain a fusion risk representation after multimodal information interaction and fusion.

6. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, Based on the fused risk characterization, risk decision output processing is performed to obtain the risk perception results of interoperability between distribution network external equipment, including: Based on the aforementioned risk characterization, risk classification and risk level assessment are performed respectively to obtain the risk category and risk level results corresponding to equipment interoperability; By combining the risk category results, risk level results, and equipment operation scenarios, an interpretable risk cause description is generated and matched with the corresponding emergency response strategy, resulting in a risk perception result that includes the risk cause and response suggestions.

7. The method for sensing interoperability risks of external equipment in a distribution network according to claim 1, characterized in that, The method also includes a joint model training and optimization step, comprising: Based on multimodal sample data labeled with risk categories, risk levels, and risk causes, we construct temporal reconstruction loss, cross-modal alignment loss, graph structure association preservation loss, risk category identification loss, risk level assessment loss, and risk cause generation loss, respectively. The multi-task joint loss function is obtained by weighting and combining the various loss terms. The model parameters are iteratively optimized using the multi-task joint loss function to obtain a distribution network external equipment interoperability risk perception model that can perform the multimodal feature encoding process, the cross-modal asynchronous alignment process, the graph fusion risk characterization process, and the risk decision output process.

8. A risk sensing device for interoperability of external equipment in a distribution network, characterized in that, include: The data preprocessing module is used to preprocess the multimodal raw operating data of the equipment outside the distribution network to obtain standardized multimodal data samples; The feature encoding module is used to perform multimodal feature encoding processing on the standardized multimodal data samples to obtain a multimodal feature sequence of uniform dimension; The temporal alignment module is used to perform cross-modal asynchronous alignment processing on the multimodal feature sequence to obtain temporally aligned multimodal features; The risk characterization module is used to perform graph fusion risk characterization processing on the multimodal features to obtain a fused risk characterization. The graph fusion risk characterization process is a process of constructing the correlation between multimodal features and performing multi-source information interaction and fusion to generate the fused risk characterization. The result output module is used to perform risk decision output processing based on the fused risk characterization to obtain the risk perception result of interoperability between distribution network external equipment.

9. A computer device, characterized in that, The device includes a processor and a memory: The memory is used to store computer programs and send the instructions of the computer programs to the processor; The processor executes the risk perception method for interoperability of external equipment in the distribution network as described in any one of claims 1-7 according to the instructions of the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for perceiving interoperability risks of external distribution network equipment as described in any one of claims 1-7.

Images