Distributed graph database association analysis method and system based on TinkerPop API

Abstract

The invention relates to the technical field of graph database analysis, provides a distributed graph database association analysis method and system based on a TinkerPop API, and solves the problems of poor efficiency and low accuracy of wide area infrastructure abnormal event positioning and causal inference. The method comprises the following steps: collecting sound wave sensing signals of a monitoring area and SCADA time sequence data of key equipment; processing the sound wave signal to obtain strain data, and performing sound field reconstruction to determine space-time coordinates of a plurality of sound source events; simulated voiceprints are generated based on the sound wave fragments, and a structured event stream is formed by combining space-time coordinates after comparison; extracting multi-scale feature information of the SCADA data, and fusing the event stream and the feature information to generate a causal relationship set; and constructing a causal association network in a distributed graph database by taking a causal relationship as an edge and taking elements such as a geographic space entity and the like as vertexes. According to the invention, the efficiency and accuracy of wide-area infrastructure abnormal event positioning and causal inference are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph database analysis, and particularly relates to a distributed graph database correlation analysis method and system based on TinkerPop API. BACKGROUND

[0002] The distributed graph database correlation analysis method has important value in group enterprise operation, and can be used to integrate Supervisory Control and Data Acquisition (SCADA), Internet of Things sensors and other multi-source data to construct a correlation network of device states and abnormal events, especially in enterprises with extensive infrastructure and complex industrial equipment systems.

[0003] In the prior art, a distributed graph database correlation analysis method based on a Graph Computing Framework Application Programming Interface (TinkerPop API) has been applied; for example, by connecting device monitoring data in a time series database and event records in a business system, the data is first extracted and mapped into vertices and edges in a graph, and then a Gremlin query language is used for relationship traversal and pattern discovery to realize asset topology management or rule-based event correlation; this method can effectively organize and manage defined entity relationships in structured data and explicit correlation rules.

[0004] However, the adaptability of the existing method is challenged when dealing with complex scenarios such as wide-area continuous spatial physical monitoring, such as abnormal positioning of sound and vibration in infrastructure along the railway, comprehensive pipe gallery, etc.; these scenarios involve deep integration of unstructured physical sensing signals and structured device SCADA time series data, the signals themselves have the characteristics of high dimensionality, continuity and strong noise, and the relationship between the signals and the device state is not a simple, pre-set rule describable static correlation, but a dynamic, probabilistic spatiotemporal causal relationship; therefore, the existing technology has the technical problem that the physical sensing signals and device operation state data are not deeply integrated, and it is difficult to effectively mine dynamic spatiotemporal causal correlations. SUMMARY

[0005] The present application provides a distributed graph database correlation analysis method and system based on TinkerPop API to solve the problem of poor efficiency and low accuracy of wide-area infrastructure abnormal event positioning and causal inference in the prior art.

[0006] To solve the above technical problems, in a first aspect, the application provides a distributed graph database correlation analysis method based on TinkerPop API, comprising:

[0007] Collecting acoustic wave sensing signals in a monitoring area and SCADA time series data of key equipment in the monitoring area;

[0008] Processing the acoustic wave sensing signals to obtain strain data, and reconstructing the acoustic field of the strain data to obtain a plurality of sound source events, and determining the space-time coordinates for each sound source event;

[0009] Inputting the acoustic wave segment corresponding to each sound source event into a generative adversarial network, generating a simulated acoustic print through the generative adversarial network, comparing the simulated acoustic print with the acoustic wave segment, and combining the space-time coordinates to obtain a structured event stream;

[0010] Performing multi-scale feature extraction on the SCADA time series data to obtain feature information, the feature information being used to reflect the operating state of the key equipment;

[0011] Aligning and fusing the structured event stream with the feature information to obtain fusion data, and using a causal discovery algorithm to analyze the leading-lagging relationship and conditional independence test between a plurality of fusion records in the fusion data to generate a causal relationship set;

[0012] Using the causal relationships in the causal relationship set as edges, and using geographical space entities, sensing units, sound source events, and device sources as vertices, constructing a causal correlation network in a distributed graph database by calling the operations provided by the TinkerPop API, the causal correlation network being used to correlate physical disturbances with the operating state of the key equipment.

[0013] Optionally, the inputting of the acoustic wave segment corresponding to each sound source event into the generative adversarial network, the generation of the simulated acoustic print through the generative adversarial network, the comparison of the simulated acoustic print with the acoustic wave segment, and the combination of the space-time coordinates to obtain the structured event stream, comprises:

[0014] According to the occurrence time of the sound source event, extracting the acoustic wave segment in the corresponding time period from a data matrix;

[0015] Inputting the acoustic wave segment into the generative adversarial network, generating a simulated acoustic print based on the features of the acoustic wave segment through the generator of the generative adversarial network;

[0016] Comparing the differences in frequency domain features and time domain features between the acoustic wave segment and the simulated acoustic print through the discriminator of the generative adversarial network to determine the probability that the acoustic wave segment belongs to different event categories;

[0017] According to the probability, an event type label is assigned to the sound wave segment by using a softmax classifier;

[0018] The event type label is associated with the spatiotemporal coordinates to generate a structured event record, and all structured event records are combined in chronological order into a structured event stream.

[0019] Optionally, the generation of the simulated voiceprint by the generator of the generative adversarial network based on the features of the sound wave segment comprises:

[0020] The sound wave segment is received by an input layer of the generator;

[0021] The sound wave segment is encoded by a plurality of cascaded neural network layers of the generator to extract a feature vector of the sound wave segment in the frequency domain;

[0022] The feature vector is input into a decoding network of the generator, and the feature vector is combined with a preset voiceprint feature template by weighting in the decoding network;

[0023] The combined result is up-sampled by a plurality of deconvolution layers of the decoding network to generate a simulated voiceprint.

[0024] Optionally, the alignment and fusion of the structured event stream with the feature information to obtain fusion data, and the generation of a causal relationship set by using a causal discovery algorithm to analyze the leading-lagging relationship and conditional independence test between a plurality of fusion records in the fusion data, comprise:

[0025] According to the time recorded in the structured event stream, a feature corresponding to the time is searched from the feature information;

[0026] The feature is merged with a structured event record at the corresponding time in the structured event stream to form a fusion record, and a plurality of fusion records are arranged in time to obtain fusion data;

[0027] The fusion data is analyzed by using a causal discovery algorithm to determine the leading-lagging relationship between different variables;

[0028] According to the leading-lagging relationship, the causal relationship between different variables is judged in combination with a conditional independence test;

[0029] The confidence score of the causal relationship is calculated, the causal relationship and the confidence score are combined to obtain a causal relationship set.

[0030] Optionally, the sound field reconstruction of the strain data to obtain a plurality of sound source events, and the determination of a spatiotemporal coordinate for each sound source event, comprise:

[0031] The strain data are arranged along the spatial and temporal dimensions to form a data matrix;

[0032] Spatial spectrum analysis is performed on the data matrix to calculate the direction of arrival parameters of signals from different spatial azimuth and elevation angles;

[0033] Based on the arrival direction parameters, a beamforming algorithm is used to separate the vibration signal components with concentrated energy in time and space from the data matrix, and each vibration signal component is defined as a sound source event.

[0034] Based on the arrival direction parameter, calculate the distance of each sound source event relative to each sensing unit in the sensing array;

[0035] Based on the distance and the time information corresponding to the strain data in the data matrix, the location coordinates and occurrence time of each sound source event are determined, and the location coordinates and occurrence time constitute spatiotemporal coordinates.

[0036] Secondly, this application provides a distributed graph database association analysis system based on the TinkerPop API, comprising:

[0037] The acquisition module is used to acquire acoustic wave sensor signals and SCADA time-series data of key equipment within the monitoring area.

[0038] The processing module is used to process the acoustic wave sensing signal to obtain strain data, and to reconstruct the sound field of the strain data to obtain multiple sound source events, and to determine the spatiotemporal coordinates of each sound source event.

[0039] The input module is used to input the sound wave segment corresponding to each sound source event into the adversarial generative network, generate a simulated voiceprint through the adversarial generative network, compare the simulated voiceprint with the sound wave segment, and obtain a structured event stream by combining the spatiotemporal coordinates.

[0040] The extraction module is used to perform multi-scale feature extraction on the SCADA time series data to obtain feature information, which is used to reflect the operating status of key equipment;

[0041] The fusion module is used to align and fuse the structured event stream with the feature information to obtain fused data, and to use a causal discovery algorithm to analyze the lead-lag relationship and conditional independence test among multiple fused records in the fused data to generate a causal relationship set.

[0042] The constructing module is configured to construct a causal correlation network in a distributed graph database by invoking operations provided by the TinkerPop API, taking the causal relationships in the set of causal relationships as edges, and taking the geospatial entities, the sensing units, the sound source events, and the device sources as vertices, the causal correlation network being used to correlate the physical disturbance with the operating state of the key device.

[0043] In a third aspect, the present application provides an electronic device, comprising:

[0044] a memory configured to store a computer program;

[0045] a processor configured to execute the computer program to implement the steps of the TinkerPop API-based distributed graph database correlation analysis method according to the first aspect.

[0046] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executable by a processor to implement the steps of the TinkerPop API-based distributed graph database correlation analysis method according to the first aspect.

[0047] The technical scheme provided by the present application has the following beneficial effects:

[0048] The present application provides two types of heterogeneous but complementary data sources, i.e., physical perception and device state, by collecting sound waves and SCADA data; secondly, the sound wave signal is processed and the sound field is reconstructed, converting continuous physical disturbance into discrete events with accurate spatiotemporal labels, thereby realizing preliminary positioning and separation of abnormal physical phenomena; then, the sound wave segment is processed by the generative adversarial network to generate a simulated acoustic fingerprint for comparison, and a structured event stream is formed in combination with the spatiotemporal coordinates, this step enhances the explainability of the event by improving the unstructured acoustic signal into an event description with clear semantics;

[0049] Meanwhile, multi-scale feature extraction is performed on the SCADA data, which can more comprehensively depict the operating state of the device at different time granularities, providing rich state representation for subsequent causal analysis; then, the structured event stream and the device feature information are aligned and fused, and a causal discovery algorithm is used for analysis, which can infer the potential causal direction and correlation strength between the physical event and the device state from the time series data, and surpasses the simple correlation relationship; finally, the TinkerPop API is invoked to construct a causal correlation network in a graph database, realizing unified correlation modeling and persistence of multi-source heterogeneous data, so that the dynamic causal relationship between the physical disturbance and the device operating state can be visualized, stored, queried, and reasoned.

[0050] Further, the application also realizes fine identification and classification of the original acoustic signal with high confidence through the generation and discrimination mechanism of the generative adversarial network, and combines the time and space information, so as to convert the original vibration data stream into a structured event sequence with clear semantics and complete time and space information, and provide high-quality and directly processable input for subsequent causal correlation analysis.

[0051] These aspects or other aspects of the application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0053] Figure 1 A flowchart of a distributed graph database correlation analysis method based on TinkerPop API provided by the embodiment of the application;

[0054] Figure 2 A specific implementation schematic diagram of a distributed graph database correlation analysis method based on TinkerPop API provided by the embodiment of the application;

[0055] Figure 3 A structure schematic diagram of a distributed graph database correlation analysis system based on TinkerPop API provided by the embodiment of the application. DETAILED DESCRIPTION

[0056] For the foregoing technical problems, the root cause lies in that the existing method mainly relies on pre-defined rules or simple data mapping to construct static correlation when processing physical sensing signals and device monitoring data. Such method is difficult to effectively extract "events" with business meaning when facing continuous, high-dimensional and semantically unclear physical signals such as sound wave sensing, and is unable to deeply mine the potential and dynamic causal relationship between the physical events and the complex device running state. Therefore, directly applying the existing technology to process wide-area physical monitoring scenarios often leads to that the correlation analysis between the physical disturbance and the device state stays on the surface, and is unable to realize intelligent tracing and reasoning from physical phenomena to system root cause.

[0057] To solve the above problems, the application provides a distributed graph database correlation analysis method based on TinkerPop API, the core of which is: introducing an adversarial generative network to finely identify and semantically label the sound source event, generating a structured event stream, thereby converting the unstructured physical signal into a knowledge unit with clear type, time and location information; at the same time, multi-scale feature extraction is performed on the device SCADA data to form feature information reflecting the comprehensive operation state of the device. On this basis, the structured event stream and the device feature information are spatio-temporally aligned and fused, and a causal discovery algorithm is applied to analyze the lead-lag relationship and conditional independence therebetween, so as to infer the probabilistic causal hypothesis between the physical event and the device state. Finally, by calling the TinkerPop API, these causal hypotheses are taken as dynamic edges to build and continuously update a causal correlation network in the distributed graph database together with the vertices of geographic space, sensing unit, event, device source, etc., forming a queryable and inferable causal correlation network.

[0058] Therefore, through the above cooperative working process, the scheme realizes deep semantic fusion and dynamic causal mining of physical sensing signals and device state data, fundamentally solving the problem of limited correlation analysis precision and depth caused by insufficient data fusion and causal reasoning ability in the prior art, and providing an effective technical means for abnormal positioning and root cause analysis of large-scale infrastructure.

[0059] In order to enable those skilled in the art to better understand the scheme of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0060] The core of the present application is to provide a distributed graph database correlation analysis method based on TinkerPop API, and a specific embodiment thereof is shown in the flowchart as Figure 1 The method comprises:

[0061] Step 101: collecting sound wave sensing signals in the monitoring area and SCADA time series data of key devices in the monitoring area.

[0062] In step 101, the sound wave sensing signal refers to the data collected by a distributed fiber optic sound wave sensing system, which uses a communication optical cable laid along the monitoring area such as a railway line or a comprehensive pipe gallery as a continuous sensor to convert the light signal changes caused by vibration or sound wave and other physical disturbances at each point along the optical cable into electrical signals;

[0063] SCADA time series data refers to physical quantity measurement point data reflecting the running state of equipment, which is continuously acquired at fixed time intervals from the monitoring and data acquisition system of key equipment such as a transformer or a pump station in a monitoring area.

[0064] In the embodiment of the present application, firstly, the original acoustic wave sensing signals reflecting the physical disturbance along the line are collected in real time through the optical fiber sensing network deployed in the monitoring area; at the same time, the SCADA time series data of the running state of each key equipment in the area are synchronously collected from the monitoring system of the equipment, thereby providing two types of heterogeneous data sources from the physical perception layer and the equipment monitoring layer for subsequent correlation analysis.

[0065] Step 102: processing the acoustic wave sensing signals to obtain strain data, and performing acoustic field reconstruction on the strain data to obtain a plurality of acoustic source events, and determining a space-time coordinate for each acoustic source event.

[0066] Among them, the strain data refers to a numerical sequence obtained by demodulating the distributed optical fiber sensing signals, which represents the physical disturbance intensity suffered by each point along the optical cable at each sampling time; the acoustic source event refers to a signal component in the data processing process, which is identified as originating from an independent physical vibration source, and is relatively concentrated in time and space; the space-time coordinate refers to the combined information for uniquely identifying the geographical position and time point at which an acoustic source event occurs, which includes the position coordinate and the occurrence time.

[0067] In this embodiment, the step 102 includes the following processes:

[0068] Step 1021: arranging the strain data along the spatial dimension and the time dimension to form a data matrix.

[0069] In step 1021, the data matrix is a mathematical structure obtained by organizing the strain data corresponding to each sensing point arranged along the spatial position of the optical cable and each sampling time arranged along the time axis in a two-dimensional table.

[0070] In the embodiment of the present application, firstly, the corresponding strain data is organized according to the spatial position sequence of each sensing point and the time sequence of each sampling time, and then filled into a two-dimensional table according to the rule that the spatial position sequence corresponds to the row and the time sequence corresponds to the column, thereby constructing a data matrix whose row and column describe the spatial distribution and the time evolution.

[0071] Step 1022: performing spatial spectrum analysis on the data matrix to calculate the direction of arrival parameters of signals from different spatial azimuth and elevation angles.

[0072] In step 1022, the direction of arrival parameter refers to a set of angle values calculated by the method, which represent the spatial orientation of the signal source, including azimuth and elevation.

[0073] In the embodiments of the present application, the correlation calculation is performed on the signals from different spatial positions in the data matrix to form a covariance matrix describing the spatial relationship between the signals, and then the eigenvalue decomposition or spectral peak search processing is performed on the covariance matrix, so as to estimate and extract the direction of arrival parameter representing the main spatial direction of the vibration signal.

[0074] Step 1023: According to the direction of arrival parameter, the beam forming algorithm is used to separate the vibration signal components with energy concentrated in time and space from the data matrix, and each vibration signal component is defined as a sound source event.

[0075] In the embodiments of the present application, a digital beam forming filter is constructed for each specific direction based on the direction of arrival parameter, and then the signals in the data matrix are input into the corresponding filter, so as to enhance the signal components from each direction respectively, and then the energy concentration degree of each enhanced signal component in the time axis and the space axis is analyzed, and the signal components with energy concentrated in time and space are extracted, and finally each such independent component is defined as a sound source event.

[0076] Step 1024: According to the direction of arrival parameter, the distance of each sound source event relative to each sensing unit in the sensing array is calculated.

[0077] In step 1024, the sensing array refers to a sequence of continuous virtual sensing points distributed along the optical cable for collecting strain data; the spatial position distribution of the sensing array determines the organization structure of the data matrix in the spatial dimension, and each row of data in the data matrix corresponds to the strain data collected by a specific sensing unit in the sensing array in the time sequence, so the sensing array is the physical basis and data source for constructing the spatial dimension of the data matrix.

[0078] In the embodiments of the present application, according to the direction of arrival parameter corresponding to each sound source event, combined with the geographical position information of each sensing unit in the sensing array and the physical orientation of the optical cable, the straight line distance between the occurrence position of each sound source event and each sensing unit in the sensing array is calculated through geometric triangular relationship.

[0079] The explanation of the above geometric triangular relationship can refer to related technologies, and the related description of the geometric triangular relationship in the present embodiment is not repeated.

[0080] Step 1025: determining the position coordinates and occurrence time of each sound source event based on the distance and the time information corresponding to the strain data in the data matrix, wherein the position coordinates and the occurrence time constitute a space-time coordinate.

[0081] In the embodiment of the present application, the distance information of each sound source event to the plurality of sensing units is calculated, and a multi-point geometric positioning method is used to solve, so as to determine the accurate position coordinates of the sound source event in space, and the specific time point of the occurrence of the sound source event is determined according to the starting position of the signal component corresponding to the sound source event on the time axis of the data matrix. Finally, the calculated position coordinates and the determined occurrence time point are combined to form the space-time coordinates of the sound source event.

[0082] Among them, the explanation of the multi-point geometric positioning method can refer to the related technology, and the specific calculation process of the method is not described in detail in the embodiment.

[0083] The present application analyzes the original continuous acoustic wave sensing signal into a series of discrete independent physical events with accurate geographical position and occurrence time through the above process, which lays a key foundation for subsequent conversion of physical disturbance into structured information that can be analyzed.

[0084] Step 103: inputting the acoustic wave segment corresponding to each sound source event into the generative adversarial network, generating a simulated voiceprint through the generative adversarial network, comparing the simulated voiceprint with the acoustic wave segment, and obtaining a structured event stream in combination with the space-time coordinates.

[0085] Among them, the acoustic wave segment refers to the waveform sequence of the original vibration data corresponding to the sound source event separated in step 102 and intercepted from the data matrix within a specific time period;

[0086] The simulated voiceprint refers to the waveform data synthesized by the generator in the generative adversarial network according to the input acoustic wave segment features, which represents the typical sound features of a certain type of device;

[0087] The structured event stream refers to a data sequence formed by a series of records with uniform format, containing event type semantic labels and space-time coordinates, arranged in chronological order.

[0088] The overall structure of the generative adversarial network can be a deep convolutional generative adversarial network, and the generator part adopts an encoder-decoder architecture. The encoder substructure is composed of three one-dimensional convolution layers in cascade. Each layer specifically includes convolution operation, batch normalization processing and LeakyReLU activation function. For example, the first layer uses a convolution kernel with a width of 64 and a step of 2 to map the input segment from a length of 1000 to a feature map with a length of 500.

[0089] The decoder substructure is composed of three one-dimensional transpose convolution layers in cascade, each of which specifically includes a transpose convolution operation, batch normalization processing and a ReLU activation function, and a Tanh activation function is used in the last layer to output a waveform, for example, a convolution kernel with a width of 64 and a step of 2 is used to upsample the feature map to an output with a length of 1000 in the last layer; the discriminator part adopts a classifier architecture, and the substructure thereof is composed of four one-dimensional convolution layers in cascade, each of which includes a convolution operation, batch normalization processing, a LeakyReLU activation function and a dropout layer, and a fully connected layer is connected at the end to output a classification probability.

[0090] The training process of the model first prepares a sound wave segment dataset containing labeled event types as real samples, initializes the network parameters of the generator and the discriminator, and then fixes the generator parameters and updates the discriminator in each training iteration, that is, inputs real samples and false samples synthesized by the generator into the discriminator, optimizes the discriminator to better distinguish true and false by calculating the binary cross-entropy loss and back propagation;

[0091] Then fix the discriminator parameters and update the generator, input the output of the generator into the discriminator with fixed parameters, calculate the loss of the generated sample being judged as true, and combine it with the feature matching loss of the real sample, optimize the generator to synthesize more realistic samples by back propagation; such a cycle is carried out for adversarial training until the performance of the generator and the discriminator reaches a balance and the loss function converges, and finally a stable model that can synthesize high-quality simulated voiceprints and accurately distinguish event categories is obtained.

[0092] It should be noted that the above structure is exemplary, and the internal structure design, parameter design, training process, etc. of the adversarial generative network are not specifically limited in the present application, and can be set accordingly according to the actual situation.

[0093] In this embodiment, the step 103 includes the following flow as shown in Figure 2

[0094] Step 1031: Extract the sound wave segment in the corresponding time period from the data matrix according to the occurrence time of the sound source event.

[0095] In step 1031, the time period refers to a time interval formed by extending a preset time length forward and backward from the timestamp of the sound source event.

[0096] ​In the embodiments of the present application, first, according to the occurrence timestamp of the sound source event, a certain length of data is intercepted forward and backward respectively to determine a corresponding time period, then according to the time period, the corresponding column range is located on the time column of the data matrix, and according to the sensing point range corresponding to the sound source event, the corresponding row is located on the space row of the data matrix, and finally all strain values in the row-column intersection area of the data matrix are extracted, which are arranged in time sequence to constitute the sound wave segment of the sound source event.

[0097] In practical application, assuming that a sound source event is identified by sound field reconstruction of distributed optical fiber sound wave sensing data of a comprehensive pipe gallery along a certain railway, the occurrence timestamp of the sound source event is 10:30:00 on the monitoring day, the preset interception length is 0.5 seconds forward and backward, and the determined time period is from 10:29:59.5 to 10:30:00.5 on the monitoring day; the data of all virtual sensing channels corresponding to the K100+500 to K101+000 section of the pipe gallery in the time period are extracted from the data matrix constructed by the optical fiber sensing data accessed to the enterprise unified data base Internet of Things platform, and a sound wave segment containing the physical vibration intensity change along the section in 1 second is obtained, which can be pushed to the intelligent algorithm modeling workbench of the base for further analysis.

[0098] Step 1032: inputting the sound wave segment into the generative adversarial network, generating a simulated voiceprint based on the features of the sound wave segment through the generator of the generative adversarial network.

[0099] In the embodiments of the present application, the step 1032 can specifically include the following steps:

[0100] A1: receiving the sound wave segment through the input layer of the generator.

[0101] In step A1, the input layer of the generator receives the sound wave segment data from the previous step, and the input layer is usually a fully connected layer or a one-dimensional convolutional layer, which converts the input data into a tensor format that can be processed by the neural network.

[0102] In the embodiments of the present application, the sound wave segment data is input in the form of a floating point number array, the input layer contains neuron nodes matched with the length of the sound wave segment, and each node receives the value of a sampling point.

[0103] A2: encoding the sound wave segment through a plurality of cascaded neural network layers of the generator to extract the feature vector of the sound wave segment in the frequency domain.

[0104] In step A2, the cascaded neural network layers refer to a plurality of sequentially connected fully connected layers or convolutional layers for performing nonlinear transformation and feature abstraction on the input data; and the feature vector in the frequency domain refers to a low-dimensional mathematical vector learned by the neural network and capable of representing the frequency distribution and energy concentration characteristics of the sound wave signal.

[0105] In the embodiments of the present application, the sound wave segment is sent to the subsequent multiple cascaded neural network layers of the generator after passing through the input layer. These layers first extract the local time-frequency features of the signal through convolutional operation, then reduce the data dimension and enhance the feature robustness through pooling operation, and finally integrate and compress the dispersed features into a fixed-length feature vector through fully connected layers, which represents the main characteristics of the original sound wave segment in the frequency domain.

[0106] In actual application, the encoding part includes three one-dimensional convolutional layers, each followed by an activation function and a pooling layer. After processing by the three layers, the dimension of the sound wave segment is gradually transformed from 1000x1 to 250x32, 125x64, and 62x128, and finally flattened and compressed into a feature vector with a length of 128 through a fully connected layer.

[0107] A3: inputting the feature vector into the decoding network of the generator, and performing weighted combination of the feature vector and a preset voiceprint feature template in the decoding network.

[0108] In step A3, the voiceprint feature template is a mathematical vector representing the typical acoustic characteristics of different types of devices, which is learned through training in advance.

[0109] In the embodiments of the present application, the feature vector output from the encoding part is input into the decoding network. The decoding network first maps the feature vector to the same dimension as the preset voiceprint feature template through a fully connected layer, and then performs weighted summation operation on the mapped feature vector and one or more voiceprint feature templates loaded from the pre-trained model library in this layer, wherein the weight coefficients are determined by network parameters dynamically learned or set as fixed values according to the task, so as to generate a new feature vector that integrates the features of the input segment and the typical voiceprint characteristics of the device.

[0110] A4: performing up-sampling on the weighted combination result through multiple deconvolutional layers of the decoding network to generate a simulated voiceprint.

[0111] In step A4, the deconvolutional layer is a neural network layer for up-sampling data to increase its spatial or temporal dimension; and the up-sampling refers to the process of increasing the length of the data sequence.

[0112] In the embodiments of the present application, the fused feature vector obtained after weighted combination is sent to the subsequent multiple deconvolution layers of the decoding network. Each deconvolution layer performs deconvolution operation on the input data through a specific filter, gradually increases the length of the data and reduces the number of feature channels. After several layers of such processing, the data is reconstructed into a time-domain waveform sequence with the same length as the original sound wave segment. This waveform sequence is the simulated voiceprint finally output by the generator.

[0113] In actual application, the decoding network includes three deconvolution layers. The first layer up-samples the 256-dimensional fused vector into data with a length of 500 and a channel number of 128. The second layer up-samples the data into data with a length of 1000 and a channel number of 64. The third layer up-samples the data into data with a length of 1000 and a channel number of 1. This data is the simulated voiceprint.

[0114] In step 1033, the discriminator refers to the neural network part of the adversarial generation network responsible for distinguishing whether the input data is real data or generator synthesized data. The frequency domain feature refers to the amplitude and phase distribution characteristics of the signal in the frequency domain after Fourier transform. The time domain feature refers to the waveform, amplitude and statistical characteristics of the signal in the time axis. The probability of event category refers to a numerical distribution output by the discriminator, representing the possibility of the input sound wave segment belonging to each pre-defined device event type.

[0115] In step 1033, the discriminator refers to the neural network part of the adversarial generation network responsible for distinguishing whether the input data is real data or generator synthesized data. The frequency domain feature refers to the amplitude and phase distribution characteristics of the signal in the frequency domain after Fourier transform. The time domain feature refers to the waveform, amplitude and statistical characteristics of the signal in the time axis. The probability of event category refers to a numerical distribution output by the discriminator, representing the possibility of the input sound wave segment belonging to each pre-defined device event type.

[0116] In step 1033, the discriminator refers to the neural network part of the adversarial generation network responsible for distinguishing whether the input data is real data or generator synthesized data. The frequency domain feature refers to the amplitude and phase distribution characteristics of the signal in the frequency domain after Fourier transform. The time domain feature refers to the waveform, amplitude and statistical characteristics of the signal in the time axis. The probability of event category refers to a numerical distribution output by the discriminator, representing the possibility of the input sound wave segment belonging to each pre-defined device event type.

[0117] In the embodiments of the present application, the discriminator receives the original sound wave segment and the generated simulated voiceprint as input. The discriminator extracts the frequency spectrum features of the two inputs in the frequency domain and the waveform envelope, zero-crossing rate and other features in the time domain through multiple convolutional neural networks. Then, the difference between the two inputs in each feature dimension is calculated. Finally, a probability vector is output through a fully connected layer and an activation function. Each element value in the probability vector represents the confidence of the input sound wave segment belonging to a specific event category.

[0118] In practical applications, the discriminator extracts two types of features, mel-frequency cepstral coefficients and short-time energy, from the sound wave segment and the simulated voiceprint respectively, obtains a difference score by calculating the cosine similarity of the two features, and then compares the score with the classification decision boundary learned from the historical railway equipment acoustic event training data based on the intelligent algorithm modeling platform of the enterprise unified data base. Finally, a probability distribution conforming to the railway operation and maintenance scene event classification system is output, for example, train passing 0.05, catenary arc discharge 0.85, and foreign matter intrusion 0.10.

[0119] Step 1034: According to the probability, an event type label is assigned to the sound wave segment using a softmax classifier.

[0120] In step 1034, the softmax classifier is a mathematical function that converts the original score output by the neural network into a normalized probability distribution and selects the class with the maximum probability as the final classification result.

[0121] In the embodiments of the present application, the probability vector output by the discriminator is input into a softmax classifier. The classifier first normalizes the probability vector to ensure that the sum of all elements is 1, then selects the class index corresponding to the element with the maximum value, and finally assigns the corresponding event type label to the sound wave segment according to the preset mapping relationship between the class index and the event type name.

[0122] In practical applications, the original score vector output by the discriminator is 2.1, 5.3, and 0.8. After the softmax function is calculated, the normalized probability is 0.05, 0.85, and 0.10. The second element value 0.85 is the largest, and the corresponding event type is catenary arc discharge. Therefore, the event type label assigned to the sound wave segment is catenary arc discharge.

[0123] Step 1035: Associate the event type label with the spatiotemporal coordinates to generate a structured event record, and combine all structured event records in chronological order into a structured event stream.

[0124] In step 1035, the structured event record refers to a data object containing three core fields, event type label, location coordinates, and occurrence timestamp.

[0125] In the embodiment of the present application, the event type label assigned in step 1034 is bound with the spatiotemporal coordinates of the sound source event determined in step 102 to create a data record containing five fields of event type, latitude, longitude, altitude and timestamp, i.e., a structured event record; after all sound source events are processed, all generated structured event records are sorted in ascending order according to their timestamp fields, and then these sorted records are connected in sequence to form a continuous, time-advancing structured event stream.

[0126] In actual application, the spatiotemporal coordinates of a sound source event are at the K100+750 mile along a certain railway line, corresponding to the geographical coordinates of north latitude 123 degrees and east longitude 456 degrees, with a timestamp of June 15, 2025, 10:30:00, and the event type label is pantograph arc discharge, then the generated structured event record contains the fields of event type, mile position, geographical coordinates and timestamp; it should be understood that the above east longitude, north latitude and other related values given later are all virtual, hypothetical geographical information, not real existing geographical information.

[0127] When multiple such records are generated in the same monitoring period, after sorting all records by timestamp, through the data stream processing capability of the data base, they are written into a time series database or as a streaming data source, providing for the subsequent enterprise graph database to build an associated network, thereby forming a structured event stream for analysis.

[0128] In the above process, the present application realizes intelligent recognition and high-confidence classification of unstructured acoustic signals by using a generative adversarial network, and combines them with accurate spatiotemporal information, thereby converting the original physical disturbance data into an event sequence with clear semantics and standardized format, providing high-quality, directly processable input information for subsequent deep fusion and causal analysis with device state data.

[0129] Step 104: Multi-scale feature extraction is performed on the SCADA time series data to obtain feature information, which is used to reflect the running state of the key equipment.

[0130] In step 104, the feature information refers to a set of quantitative indicators that can more comprehensively represent the comprehensive running state of the equipment in a certain time period, obtained by calculating and combining the original SCADA time series data.

[0131] In the embodiment of the present application, first, the SCADA time series data stream collected from the key equipment monitoring system is received, and then for each physical quantity measurement point data of interest in the data stream, parallel calculation and analysis are performed using a short time window and a long time window respectively, the local statistical features reflecting the rapid fluctuation and instantaneous change of the data are calculated in the short time window, and the global statistical features reflecting the overall level and trend evolution of the data are calculated in the long time window.

[0132] Then, the feature values of each measurement point calculated at different time scales are combined to form a multi-scale feature vector of the measurement point, and finally the feature vectors of all related measurement points are integrated together according to a predetermined dimension order to form a multi-dimensional feature matrix, which is the final feature information obtained for comprehensively reflecting the running state of the key equipment.

[0133] Step 105: Align and fuse the structured event stream and the feature information to obtain fused data, and use a causal discovery algorithm to analyze the lead-lag relationship and conditional independence test between multiple fused records in the fused data to generate a causal relationship set.

[0134] Among them, the fused data refers to a new data set generated after the alignment and fusion operation, each record of which contains event type, spatio-temporal information and equipment feature information; the lead-lag relationship refers to the correlation between two time series at different time offsets to determine the degree of one variable leading another variable in time;

[0135] The conditional independence test refers to a statistical test method for determining whether two variables remain statistically independent given the values of other variables; the causal relationship set refers to a data set output by the causal discovery algorithm, containing a set of inferred causal relationships and their confidence scores.

[0136] In this embodiment, the step 105 includes the following flow:

[0137] Step 1051: According to the time of the record in the structured event stream, find the corresponding feature in the feature information.

[0138] In step 1051, the time of the record in the structured event stream refers to the occurrence timestamp marked in each structured event record.

[0139] In the embodiment of the present application, first, the timestamp of each record in the structured event stream is read one by one, and then the timestamp is taken as the reference to find the equipment feature vector corresponding to the time point identical or closest to the event timestamp in the time series corresponding to the feature information generated by step 104, so as to realize accurate time positioning of the event.

[0140] In practical applications, assuming that the occurrence timestamp of a structured event record is June 15, 2025, 10:30:00, the event originates from the acoustic monitoring of a substation along the railway, and the time series of feature information is the data collected from the SCADA system of the key equipment of the substation and stored in the enterprise time series database, with one sampling point per minute, then the device feature vector closest to the timestamp is queried and extracted from the time series database, for example, the 10:30:00 corresponding multi-dimensional feature data including transformer oil temperature, load current, etc., and the device operating state feature at the same time as the acoustic event is taken as the device operating state feature.

[0141] Step 1052: merge the feature with the structured event record at the corresponding time in the structured event stream to form a fusion record, and arrange a plurality of the fusion records according to time to obtain fusion data.

[0142] In step 1052, the fusion record refers to a new data record formed by splicing the fields of a structured event record and the corresponding device feature vector together, which expands the information dimension.

[0143] In the embodiment of the present application, the found device feature vector is added to the corresponding structured event record as a new field, thereby generating a fusion record containing event attributes, spatial position, timestamp, and multi-dimensional device features; then all event records are processed to complete the feature merging operation for each record, and finally all generated fusion records are sorted in ascending order according to their timestamp fields, and these sorted records are arranged continuously, thereby constituting a unified multi-variable time series data arranged in time sequence, which is the fusion data.

[0144] In practical applications, a structured event record contains three fields of event type such as contact net arc discharge, position such as railway mileage K100+750, and timestamp, and the corresponding device feature vector is an array of 10 values reflecting the device operating state extracted from the substation SCADA system and obtained through multi-scale calculation, and a fusion record containing 13 fields is generated after merging; when there are 100 such acoustic event records in the same monitoring period such as 10 minutes, 100 fusion records are generated through the stream batch computing capability of the data base, and a structured table data of 100 rows and 13 columns is formed after sorting by time, which is the fusion data available for subsequent causal discovery algorithm analysis.

[0145] Step 1053: analyze the fusion data using a causal discovery algorithm to determine the leading-lag relationship between different variables.

[0146] In step 1053, the variables refer to each column of data in the fusion data, such as the event type variable and the equipment characteristic variable. The explanation of the causal discovery algorithm can refer to the related art, which will not be described here.

[0147] In the embodiment of the application, the fusion data is input into the causal discovery algorithm for processing; the algorithm first calculates the cross-correlation function or Granger causality test statistic under different time offsets for all variable pairs in the fusion data, such as the event type variable A and the equipment temperature characteristic variable B, determines the optimal time delay of variable A leading variable B or the optimal time delay of variable B leading variable A by searching for the offset that maximizes the statistic, and finally determines the leading-lag relationship between all variable pairs.

[0148] In actual application, the algorithm calculates the cross-correlation function between the event type "catenary arc discharge" and the equipment characteristic "transformer winding temperature", finds that the cross-correlation value is maximum, i.e. 0.6, when the event type leads the temperature characteristic by 3 seconds in the range of calculating time offset from -5 seconds to 5 seconds, and thus determines that the event variable leads the temperature characteristic variable by 3 seconds.

[0149] Step 1054: According to the leading-lag relationship, the causal relationship between different variables is judged in combination with the conditional independence test.

[0150] In step 1054, the conditional independence test is used to exclude the pseudo-correlation caused by the existence of other common cause variables, so as to more accurately judge the direct causal relationship.

[0151] In the embodiment of the application, based on the determined leading-lag relationship, the causal discovery algorithm further performs conditional independence test on all possible variable triplets; for example, for the variable X and the variable Y with a leading relationship, the algorithm will test whether X and Y are still correlated under the condition of a given third variable Z; if X and Y become independent under the condition of a certain variable Z, it indicates that the direct causal relationship between X and Y does not exist, and the correlation between them may be caused by Z; by systematically performing a large number of such tests, the algorithm can infer the most likely direct causal direction between variables.

[0152] In actual application, it has been determined that the event variable X leads the characteristic variable Y, and the algorithm introduces a third variable Z, i.e. the environment temperature, for conditional independence test; it is found that the partial correlation coefficient between the variable X and the variable Y is close to 0 when the specific value of the environment temperature Z is given, i.e. the two variables become independent, and thus it is judged that there is no direct causal relationship between the event variable X and the characteristic variable Y, and the correlation between them is caused by the environment temperature Z as a common cause.

[0153] Step 1055: calculating a confidence score of the causal relationship, combining the causal relationship and the confidence score to obtain a causal relationship set.

[0154] In step 1055, the confidence score is a numerical value for quantifying the reliability of the inferred causal relationship, which is usually calculated based on the statistical test level or the goodness of model fitting.

[0155] In the embodiments of the present application, for each direct causal relationship determined to be established in step 1054, a confidence score is calculated; the score can be obtained by calculating the p-value of the F-statistic of Granger causality test, or by comparing the goodness of the model containing and not containing the causal edge through Bayesian information criterion; then the description of each causal relationship, including the cause variable, the result variable, the causal direction, and the calculated confidence score, are bound to form a causal relationship entry; all the inferred causal relationship entries are collected together to constitute the final causal relationship set.

[0156] In practical applications, it is determined that the event variable "catenary arc discharge" is the cause of the increase of the device feature variable "number of protection device actions", and the p-value obtained through Granger causality test is 0.01, so the confidence score is set to 0.99; the finally generated causal relationship set contains entries such as cause catenary arc discharge, result protection device action number, direction cause pointing to result, and confidence 0.99.

[0157] The present application realizes the deep fusion of physical events and device state data through the above process, and uses statistical inference method to mine the causal relationship with clear time direction and statistics from the fused data, thereby providing direct and reliable data basis for subsequent construction of an interpretable and quantifiable causal correlation network.

[0158] Step 106: constructing a causal correlation network in a distributed graph database by calling the operations provided by TinkerPop API, taking the causal relationships in the causal relationship set as edges, and taking geographical space entities, sensing units, sound source events and device sources as vertices, the causal correlation network being used for correlating physical disturbances and the operating state of the key device.

[0159] The vertex is a basic unit in the graph data structure and is used to represent a specific entity or object. The edge is a directed line segment connecting two vertices and is used to represent the relationship between the vertices. The causal association network is a graph data structure model constructed in the present application, in which the vertex represents an entity in the physical world, such as a geographic location, a fiber sensing point, an identified acoustic event, and a source device that can cause vibration, and the directed edge represents a causal relationship from one entity to another entity inferred from the analysis data, and the network is used to systematically organize and store the dynamic association relationship between the physical disturbance event and the device operating state in a distributed graph database.

[0160] The physical disturbance can refer to a physical phenomenon that causes strain changes in a communication cable, which is monitored by a distributed fiber acoustic wave sensing system. These phenomena include, but are not limited to, vibrations caused by passing trains, abnormal sounds caused by device failures, or impacts caused by human activities. They appear as continuous waveform changes in the sensing signal and are the event sources to be analyzed in the present method.

[0161] In this embodiment, the step 106 includes the following process:

[0162] Step 1061: In the distributed graph database, for each causal relationship in the set of causal relationships, call the add edge instruction provided by the TinkerPop API to add a directed edge between the corresponding vertices.

[0163] In step 1061, the directed edge refers to an edge with a clear direction, which is used to represent the direction of the relationship, for example, from cause to result.

[0164] In the embodiments of the present application, first, each record in the set of causal relationships is read, and the vertices representing the cause variable and the result variable are found in the distributed graph database according to the specified variables in the record. Then, a specific instruction for adding a directed edge provided in the TinkerPop API is called to create a directed edge from the cause variable vertex to the result variable vertex, and other information of the causal relationship is recorded in the attributes of the directed edge, such as the confidence score obtained from the set of causal relationships.

[0165] Step 1062: Query and analyze the vertices and the directed edges by calling the graph traversal instruction provided by the TinkerPop API to form a causal association network.

[0166] In step 1062, the graph traversal instruction refers to a query operation for accessing vertices and edges in a graph data structure according to a specific path or rule. The structure of the causal association network can refer to related technologies, which will not be described here.

[0167] In the embodiments of the present application, after all the directed edges corresponding to the causal relationships are added, the graph database is queried by calling the graph traversal instruction provided by the TinkerPop API; for example, a traversal rule is written using the Gremlin query language, starting from a specified vertex, gradually accessing all other vertices and edges connected thereto along the direction of the directed edge, so as to explore the connection structure and path of the entire graph; the overall graph structure composed of vertices and directed edges revealed by the instruction query and traversal is the causal association network described in the present application.

[0168] In this embodiment, after step 106, the method further includes the following flow:

[0169] B1: input the vertex data and edge data in the causal association network into a graph neural network, and update the feature representation of the vertex data and edge data by iteratively calculating the vertex data and edge data through the graph neural network.

[0170] In the embodiments of the present application, first, the initial feature vectors of all vertices and the attribute vectors of all edges in the causal association network are input into a pre-trained graph neural network; the network allows each vertex to receive information from its adjacent vertices and connected edges through its internal message passing mechanism, and then fuses and calculates the received information and the original features of itself through an aggregation function and an update function. After multiple layers of such iterative calculations, each vertex and each edge will generate an updated new feature representation containing local graph structure information.

[0171] The structure design, parameter design, training process, etc. of the graph neural network can refer to related technologies, which will not be described here.

[0172] B2: according to the updated feature representation output by the graph neural network, the attention weight of each associated edge in the causal association network is calculated using an attention mechanism.

[0173] In step B2, the attention weight is a numerical value for measuring the importance of a certain edge in the graph in the current task, which is obtained by calculating the relevance of the features of the vertices connected to the edge.

[0174] In the embodiments of the present application, based on the updated vertex features output by the graph neural network, an attention mechanism is applied to each directed edge in the causal association network; specifically, for an edge connecting vertex A and vertex B, first, the updated feature vectors of vertex A and vertex B are spliced or dot product operation to obtain an initial attention score, then the score is input into a single-layer neural network and normalized to finally calculate a value between 0 and 1, which is the attention weight of the edge, and the higher the weight, the more important the edge in the current context.

[0175] B3: Based on the attention weight, updating the weight attribute of the corresponding association edge in the distributed graph database.

[0176] In the embodiments of the present application, the calculated attention weight of each edge is written into the weight attribute field of the corresponding edge in the distributed graph database as a new attribute value of the edge by calling the update attribute instruction of the TinkerPop API, thereby completing the dynamic optimization and update of the edge weight in the graph.

[0177] Through the above process, the present application not only dynamically constructs the mined causal relationship into a visual graph network, but also further utilizes the graph neural network and the attention mechanism to intelligently learn and dynamically adjust the relationship strength in the network, so that the causal association network can more accurately and finely reflect the complex and dynamically changing dependency relationship between the physical disturbance and the device state.

[0178] Figure 3 A structure diagram of a distributed graph database association analysis system based on TinkerPop API provided in the embodiments of the present application is shown in FIG. 1, which includes: Figure 3

[0179] The acquisition module 31 is configured to acquire the acoustic wave sensing signal in the monitoring area and the SCADA time series data of the key equipment in the monitoring area.

[0180] The processing module 32 is configured to process the acoustic wave sensing signal to obtain strain data, and reconstruct an acoustic field based on the strain data to obtain a plurality of sound source events and determine a space-time coordinate for each sound source event.

[0181] The input module 33 is configured to input an acoustic wave segment corresponding to each sound source event to a generative adversarial network, generate a simulated acoustic fingerprint through the generative adversarial network, compare the simulated acoustic fingerprint with the acoustic wave segment, and obtain a structured event stream in combination with the space-time coordinate.

[0182] The extraction module 34 is configured to perform multi-scale feature extraction on the SCADA time series data to obtain feature information, and the feature information is used to reflect the running state of the key equipment.​

[0183] The fusion module 35 is used to align and fuse the structured event stream with the feature information to obtain fused data, and to use a causal discovery algorithm to analyze the lead-lag relationship and conditional independence test among multiple fused records in the fused data to generate a causal relationship set.

[0184] The construction module 36 is used to construct a causal relationship network in a distributed graph database by calling the operations provided by the TinkerPop API, using the causal relationships in the causal relationship set as edges and geospatial entities, sensing units, sound source events and device sources as vertices. The causal relationship network is used to associate physical disturbances with the operating status of the key equipment.

[0185] The distributed graph database association analysis system based on the TinkerPop API in this application embodiment is used to implement the aforementioned distributed graph database association analysis method based on the TinkerPop API. Therefore, the specific implementation of the distributed graph database association analysis system based on the TinkerPop API can be found in the embodiment section of the distributed graph database association analysis method based on the TinkerPop API above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0186] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described distributed graph database association analysis methods based on the TinkerPop API.

[0187] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described distributed graph database association analysis methods based on the TinkerPop API.

[0188] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0189] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the distributed graph database association analysis method based on the TinkerPop API.

[0190] Those skilled in the art will further realize that the mere concepts, teachings, and embodiments described herein are merely meant to be illustrative and that changes can be made therefrom, by way of modification or improvement, within the scope of the present application which is defined in the appended claims. Accordingly, the drawings and descriptions are to be regarded as illustrative in nature and not restrictive.

[0191] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present specification are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation entrances for users to choose authorization or refusal.

[0192] The above provides a kind of based on TinkerPop API's distributed graph database association analysis method and system of the present application in detail.This paper applies specific example to the principle and implementation of the present application are described, the above example is only for helping to understand the method of the present application and its core idea.It should be pointed out, for the ordinary skilled person in the art, under the premise of not departing from the principle of the present application, the present application can be improved and modified, these improvements and modifications also fall within the scope of the present application.

Claims

1. A distributed graph database association analysis method based on the TinkerPop API, characterized in that, include: Collect acoustic sensor signals and SCADA timing data of key equipment within the monitoring area; The acoustic wave sensing signal is processed to obtain strain data, and the strain data is reconstructed into a sound field to obtain multiple sound source events, and the spatiotemporal coordinates of each sound source event are determined. The sound wave segment corresponding to each sound source event is input into the Generative Adversarial Network (GAN), which generates a simulated voiceprint. The simulated voiceprint is compared with the sound wave segment, and the structured event stream is obtained by combining the spatiotemporal coordinates. Multi-scale feature extraction is performed on the SCADA time-series data to obtain feature information, which is used to reflect the operating status of key equipment; The structured event stream is aligned and fused with the feature information to obtain fused data. A causal discovery algorithm is then used to analyze the lead-lag relationship and conditional independence test among multiple fused records in the fused data to generate a causal relationship set. By calling the operations provided by the TinkerPop API, using the causal relationships in the set of causal relationships as edges and geospatial entities, sensing units, sound source events, and device sources as vertices, a causal association network is constructed in a distributed graph database. The causal association network is used to associate physical disturbances with the operating status of the key equipment.

2. The distributed graph database association analysis method based on the TinkerPop API according to claim 1, characterized in that, The process of inputting the sound wave segment corresponding to each sound source event into a generative adversarial network (GAN), generating a simulated voiceprint through the GAN, comparing the simulated voiceprint with the sound wave segment, and combining the spatiotemporal coordinates to obtain a structured event stream includes: Based on the occurrence time of the sound source event, the sound wave segments within the corresponding time period are extracted from the data matrix; The sound wave segment is input into the Generative Adversarial Network (GAN), and the generator of the GAN generates a simulated voiceprint based on the characteristics of the sound wave segment. The discriminator of the adversarial generative network compares the differences in frequency domain features and time domain features between the sound wave fragment and the simulated voiceprint to determine the probability that the sound wave fragment belongs to different event categories. Based on the probability, an event type label is assigned to the sound wave segment using a softmax classifier; The event type labels are associated with the spatiotemporal coordinates to generate structured event records, and all structured event records are combined into a structured event stream in chronological order.

3. The distributed graph database association analysis method based on the TinkerPop API according to claim 2, characterized in that, The generator of the adversarial generative network generates simulated voiceprints based on the features of the sound wave fragments, including: The sound wave fragment is received through the input layer of the generator; The sound wave segment is encoded by multiple cascaded neural network layers of the generator to extract the feature vector of the sound wave segment in the frequency domain; The feature vector is input into the decoding network of the generator, where the feature vector is weighted and combined with a preset voiceprint feature template. The weighted combination result is upsampled through multiple deconvolution layers of the decoding network to generate a simulated voiceprint.

4. The distributed graph database association analysis method based on the TinkerPop API according to claim 1, characterized in that, The structured event stream is aligned and fused with the feature information to obtain fused data. A causal discovery algorithm is then used to analyze the lead-lag relationships and conditional independence tests among multiple fused records in the fused data, generating a causal relationship set, including: Based on the time recorded in the structured event stream, the feature corresponding to the time is found from the feature information; The features are merged with the structured event records of the corresponding time in the structured event stream to form a fused record, and multiple fused records are arranged according to time to obtain fused data; The fused data was analyzed using a causal discovery algorithm to determine the lead-lag relationships between different variables. Based on the aforementioned lead-lag relationship, and combined with the conditional independence test, determine the causal relationship between different variables; Calculate the confidence score of the causal relationship, and combine the causal relationship and the confidence score to obtain a set of causal relationships.

5. The distributed graph database association analysis method based on the TinkerPop API according to claim 1, characterized in that, The step of reconstructing the sound field from the strain data to obtain multiple sound source events, and determining the spatiotemporal coordinates for each sound source event, includes: The strain data are arranged along the spatial and temporal dimensions to form a data matrix; Spatial spectrum analysis is performed on the data matrix to calculate the direction of arrival parameters of signals from different spatial azimuth and elevation angles; Based on the arrival direction parameters, a beamforming algorithm is used to separate the vibration signal components with concentrated energy in time and space from the data matrix, and each vibration signal component is defined as a sound source event. Based on the arrival direction parameter, calculate the distance of each sound source event relative to each sensing unit in the sensing array; Based on the distance and the time information corresponding to the strain data in the data matrix, the location coordinates and occurrence time of each sound source event are determined, and the location coordinates and occurrence time constitute spatiotemporal coordinates.

6. The distributed graph database association analysis method based on the TinkerPop API according to claim 1, characterized in that, The process of constructing a causal relationship network in a distributed graph database by calling the operations provided by the TinkerPop API, using causal relationships in the causal relationship set as edges and geospatial entities, sensing units, sound source events, and device sources as vertices, includes: In the distributed graph database, for each causal relationship in the set of causal relationships, the edge addition command provided by TinkerPopAPI is called to add directed edges between the corresponding vertices; By calling the graph traversal instructions provided by the TinkerPop API, the vertices and directed edges are queried and analyzed to form a causal relationship network.

7. The distributed graph database association analysis method based on the TinkerPop API according to claim 1, characterized in that, After constructing the causal relationship network, the following is also included: The vertex data and edge data in the causal association network are input into the graph neural network. The graph neural network iteratively calculates the vertex data and edge data to update the feature representation of the vertex data and edge data. Based on the updated feature representation output by the graph neural network, the attention weight of each associated edge in the causal association network is calculated using an attention mechanism. Based on the attention weights, update the weight attributes of the corresponding associated edges in the distributed graph database.

8. A distributed graph database association analysis system based on the TinkerPop API, characterized in that, include: The acquisition module is used to acquire acoustic wave sensor signals and SCADA time-series data of key equipment within the monitoring area. The processing module is used to process the acoustic wave sensing signal to obtain strain data, and to reconstruct the sound field of the strain data to obtain multiple sound source events, and to determine the spatiotemporal coordinates of each sound source event. The input module is used to input the sound wave segment corresponding to each sound source event into the adversarial generative network, generate a simulated voiceprint through the adversarial generative network, compare the simulated voiceprint with the sound wave segment, and obtain a structured event stream by combining the spatiotemporal coordinates. The extraction module is used to perform multi-scale feature extraction on the SCADA time series data to obtain feature information, which is used to reflect the operating status of key equipment; The fusion module is used to align and fuse the structured event stream with the feature information to obtain fused data, and to use a causal discovery algorithm to analyze the lead-lag relationship and conditional independence test among multiple fused records in the fused data to generate a causal relationship set. The construction module is used to construct a causal relationship network in a distributed graph database by calling the operations provided by the TinkerPop API, using causal relationships in the causal relationship set as edges and geospatial entities, sensing units, sound source events and device sources as vertices. The causal relationship network is used to associate physical disturbances with the operating status of the key equipment.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the distributed graph database association analysis method based on the TinkerPop API as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the distributed graph database association analysis method based on the TinkerPop API as described in any one of claims 1 to 7.

Images