Intelligent power grid fault monitoring method based on artificial intelligence

Abstract

The invention relates to the technical field of power system monitoring, and discloses an intelligent power grid fault monitoring method based on artificial intelligence, which comprises the following steps: step 1, acquiring multi-modal sensing data through a multi-modal sensing network deployed on power grid equipment, the multi-mode sensing data comprises gas characteristic data acquired by an olfactory chip array, temperature data acquired by a distributed optical fiber sensor and electrical data acquired by a current transformer; and step 2, sending the multi-modal sensing data to an edge computing node, wherein a first data processing model is built in the edge computing node. According to the method, the technical scheme that the lightweight model is deployed through the edge computing nodes and graded early warning is linked is adopted, the technical effects of real-time response and rapid fault isolation are achieved, and compared with the scheme depending on cloud centralized processing in the prior art, the defect of protection action lagging caused by fault diagnosis delay is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system monitoring, in particular to a smart grid fault monitoring method based on artificial intelligence. BACKGROUND

[0002] With the continuous expansion of the scale of the power system and the advancement of the process of intelligentization, the early detection and early warning of power grid faults become increasingly important. Traditional power grid fault monitoring methods rely on changes in electrical parameters or the physical state of equipment for fault diagnosis.

[0003] For example, the Chinese invention application with publication number CN119577625A discloses a smart grid fault monitoring method and system applying artificial intelligence, relating to the technical field of power system fault monitoring, which includes obtaining power internet of things monitoring data through a sensor and preprocessing it, extracting fault-related features using a machine learning algorithm based on the preprocessed data, fusing them into a unified fault feature vector, and determining a preliminary fault monitoring threshold range through statistical analysis based on historical grid data.

[0004] The deficiencies of the above-mentioned patent are:

[0005] Traditional monitoring methods rely on current and voltage fluctuation detection, which requires a complete cycle of signal acquisition, transmission, and processing, resulting in a minute-level delay in fault diagnosis. In the emergency scenario of sudden short-circuit and ground fault in the power grid, the delay may miss the best handling opportunity, causing the protection device to act late, affecting the timeliness of fault isolation, and threatening the stable operation of the power system.

[0006] Power grid equipment, especially transformers, switching devices, and charging stations, may experience gas leakage, smoke generation, and other phenomena during operation. Traditional power grid monitoring techniques have not paid sufficient attention to early warning of hidden faults, so once a device fails, it may cause environmental pollution, equipment damage, and even serious consequences such as fires.

[0007] Conventional sensor networks are limited by sampling frequency and deployment density and cannot sense internal gradual abnormalities in real time. For example, in the case of cable joint oxidation and corrosion, traditional temperature monitoring requires waiting for the contact resistance to increase and cause significant heating before it can alarm, at which point the metal conductor has already undergone irreversible sulfidation corrosion.

[0008] Existing monitoring systems focus on electrical quantity measurement and external environment monitoring, lacking multi-dimensional perception of the internal microenvironment of equipment. In the scenario of dense busbar areas in substations and underground cable corridors, traditional monitoring methods may overlook potential safety hazards.

[0009] Therefore, the present application proposes a smart grid fault monitoring method based on artificial intelligence to solve the above-mentioned problems. SUMMARY

[0010] In view of the deficiencies of the prior art, the present application provides a smart grid fault monitoring method based on artificial intelligence to solve the problems raised in the background art.

[0011] To achieve the above object, the present application is implemented by the following technical scheme: a smart grid fault monitoring method based on artificial intelligence, comprising:

[0012] Step 1, collecting multi-modal sensing data through a multi-modal sensing network deployed on a power grid device, the multi-modal sensing data including gas feature data collected by an olfactory chip array, temperature data collected by a distributed optical fiber sensor, and electrical data collected by a current transformer;

[0013] Step 2, sending the multi-modal sensing data to an edge computing node, the edge computing node having a first data processing model built-in, processing the multi-modal sensing data by the first data processing model to generate an abnormal probability score;

[0014] Step 3, comparing the abnormal probability score with a first preset threshold in the edge computing node, when the abnormal probability score is greater than the first preset threshold, sending the abnormal probability score together with the multi-modal sensing data to a cloud server, the cloud server having a second data processing model built-in;

[0015] Step 4, analyzing the received abnormal probability score and multi-modal sensing data by the second data processing model in the cloud server to generate a fault diagnosis result including fault type data and fault location data;

[0016] Step 5, inputting the fault diagnosis result into a digital twin model for verification, combining the verified fault diagnosis result with the corresponding multi-modal sensing data into a fault sample, and updating the first data processing model and the second data processing model using the fault sample.

[0017] Preferably, the olfactory chip array includes a metal oxide semiconductor gas sensor unit, an electrochemical sensor unit, and a photoionization detector unit, for detecting flammable gas, toxic gas, and smoke particles, achieving ppd-level detection of gas types and concentrations.

[0018] Preferably, the generation of the fault diagnosis result further comprises: constructing a power grid fault knowledge graph, the knowledge graph taking device entities and fault types as nodes and fault propagation paths as edges, and using a graph attention network to reason on the knowledge graph to locate the root cause of the fault.

[0019] Preferably, in step 1, the multi-modal sensor data is collected by a multi-modal sensor network deployed on the power grid equipment, further comprising:

[0020] Sub-step 1.1, time synchronization calibration is performed on all of the olfactory chip arrays, the distributed optical fiber sensors and the current transformers in the multi-modal sensor network to generate a synchronization timestamp associated with all subsequent collected data;

[0021] Sub-step 1.2, based on the synchronization timestamp, data collection and validity discrimination are performed, and the data collection and validity discrimination are specifically:

[0022] The original resistance signal R reflecting the gas environment is collected by the olfactory chip array s , the baseline resistance signal R0 in a clean air environment is obtained, when the value of the baseline resistance signal R0 is within a preset sensor health interval [R min , R max ], a gas data validity flag is generated, and a gas response value S is calculated according to the formula: S = (R s -R0) / R0;

[0023] Wherein, R min and R max are the minimum and maximum values of the baseline resistance allowed;

[0024] The Brillouin scattering light signal is collected by the distributed optical fiber sensor, and the Brillouin frequency shift amount Δν B along the optical fiber path is demodulated, and the temperature value ΔT along the line is calculated according to the linear relationship:

[0025] Wherein, C T is the temperature sensitivity coefficient;

[0026] The complete cycle of instantaneous current waveform data i(t) is collected by the current transformer, and the current harmonic vector I h containing the fundamental wave and each harmonic amplitude is calculated by fast Fourier transform;

[0027] Sub-step 1.3, the gas data validity flag is judged, when the gas data validity flag is true, the data quality index Q is calculated according to the following data quality evaluation formula, when the data quality index Q is greater than a quality threshold Q th , the gas response value S, the temperature value ΔT and the current harmonic vector I h are encapsulated together with the synchronization timestamp as frame multi-modal sensor data;

[0028] The calculation formula of the data quality index Q is:

[0029] Q = w g · Q g + w t · Q t + w e · Q e ,

[0030] wherein w g , w t , w e are weight coefficients, Q g , Q t , Q e are normalized quality scores calculated from signal-to-noise ratio of gas response value S, fluctuation of temperature value ΔT and stability of current harmonic vector I h , respectively.

[0031] Preferably, in step 2, the multi-modal sensor data is sent to an edge computing node, the edge computing node is built-in with a first data processing model, the multi-modal sensor data is processed by the first data processing model to generate an anomaly probability score, further comprising:

[0032] Sub-step 2.1, normalizing the gas response value S, the temperature value ΔT and the current harmonic vector I h in the multi-modal sensor data to generate a normalized feature vector V norm , inputting the normalized feature vector V norm into a feature fusion module, the feature fusion module performs weighted summation on features of different modalities by a preset weight coefficient to generate a fusion feature vector V fused ;

[0033] Sub-step 2.2, taking the fusion feature vector V fused as input of the first data processing model, the first data processing model is a one-dimensional convolutional neural network, the one-dimensional convolutional neural network performs convolution operation on the fusion feature vector V fused by a convolution kernel, processes by a nonlinear activation function, extracts a feature map M feature containing deep information of device state, and the calculation formula of the convolution operation is:

[0034]

[0035] wherein M feature (j) is the jth element in the feature map, V fused is the fusion feature vector, V fused (j+i-1) is the j+i-1th element in the fusion feature vector, K(i) is the ith weight parameter in the convolution kernel, k is the size of the convolution kernel, and bk is a bias term corresponding to the convolution kernel, f act is a nonlinear activation function.

[0036] Sub-step 2.3, obtaining the feature map M feature by dimension reduction processing through a global average pooling layer to obtain a state feature vector V state , the state feature vector V state is input into a fully connected layer, and the abnormal probability score P anomaly is calculated through a Sigmoid activation function. anomaly The calculation formula of the abnormal probability score P state is as follows:

[0037]

[0038] wherein V anomaly is a state feature vector, W is a weight matrix of the fully connected layer, c is a bias term of the fully connected layer, and P norm is an abnormal probability score.

[0039] Preferably, the gas response value S is calculated through the following formula to obtain a normalized gas feature S norm :

[0040]

[0041] The temperature value ΔT is calculated through the following formula to obtain a normalized temperature feature ΔT h :

[0042]

[0043] Each harmonic amplitude I h,n in the current harmonic vector I h,n,norm is calculated through the following formula to obtain a normalized current harmonic vector I norm :

[0044]

[0045] The normalized feature vector V norm is generated by splicing the normalized modal features in a predetermined order:

[0046] V norm = [S norm , ΔT norm , I h,1,norm , I h,2,norm , …, I h,N,norm ],

[0047] wherein S, ΔT, I h,nThe gas response value, temperature value and the amplitude of the nth current harmonic are in turn, S min 、S max They are the minimum and maximum values ​​of the gas response value S, respectively.

[0048] ΔT min , ΔT max are the minimum and maximum values ​​of the temperature ΔT, respectively.

[0049] I h,n,min , I h,n,max The nth current harmonic amplitude I h,n The minimum and maximum values ​​of

[0050] N is the highest order of current harmonics;

[0051] The fused feature vector V fused is obtained by normalizing the eigenvector V norm Different modal feature components in the model are given different weight coefficients and generated after weighted adjustment;

[0052] The fused feature vector V fused The calculation formula is:

[0053] V fused =[w g ·S norm ,w t ΔT norm ,w e I h,1,norm ,w e I h,2,norm ,…,w e ·

[0054] I h,N,norm ],

[0055] Among them, V fused is the fusion feature vector, w g is the weight coefficient of the gas mode, w t is the weight coefficient of the temperature mode, w e is the weight coefficient of the electrical mode.

[0056] Preferably, in step 3, the abnormality probability score is compared with a first preset threshold in the edge computing node. When the abnormality probability score is greater than the first preset threshold, the abnormality probability score is sent together with the multimodal sensing data to a cloud server. The cloud server has a built-in second data processing model, further comprising:

[0057] Sub-step 3.1, extracting the ambient temperature and the ambient humidity contained in the multi-modal sensor data, calling the pre-stored cumulative running time data of the power grid equipment, and generating the first preset threshold in real time according to the following adaptive threshold calculation formula, the first preset threshold is a dynamic alarm threshold T dyn ;

[0058] T dyn = T base ·(1+α·f age +β·f env ),

[0059] wherein, T dyn is a dynamic alarm threshold, T base is a reference alarm threshold, f age is an aging factor, f env is an environmental interference factor, and α and β are weight coefficients;

[0060] Sub-step 3.2, sequentially comparing the abnormal probability score P anomaly with the dynamic alarm threshold T dyn and a local early warning threshold T local less than the dynamic alarm threshold, and generating an early warning level flag L alert according to the following determination rule:

[0061] When the abnormal probability score P anomaly is greater than the dynamic alarm threshold T dyn , the early warning level flag L alert is set to a high level;

[0062] When the local early warning threshold T local is less than the abnormal probability score P anomaly and the abnormal probability score P anomaly is not greater than the dynamic alarm threshold T dyn , the early warning level flag L alert is set to a medium level;

[0063] When the abnormal probability score P anomaly is not greater than the local early warning threshold T local , the early warning level flag L alert is set to a low level;

[0064] Sub-step 3.3, determining the early warning level flag L alert , when the early warning level flag L alert is a high level, the abnormal probability score P anomaly , the multi-modal sensor data triggering the score, the synchronization time stamp, and the dynamic alarm threshold T dynThe alarm event data packet is sent to the cloud server through an encrypted channel.

[0065] Preferably, in step 4, the received abnormal probability score and the multi-modal sensing data are analyzed by the second data processing model in the cloud server to generate a fault diagnosis result containing fault type data and fault location data, further comprising:

[0066] Sub-step 4.1, according to the device identity information in the alarm event data packet, the adjacent devices directly connected with the device are retrieved from the pre-constructed power grid topology graph, the historical multi-modal sensing data of the adjacent devices within a preset time window before the synchronization timestamp are called, the alarm event data packet and the historical multi-modal sensing data are combined to construct a spatio-temporal data graph G st ;

[0067] Sub-step 4.2, inputting the spatio-temporal data graph G st into the second data processing model, the second data processing model being a spatio-temporal graph convolution network, the spatio-temporal graph convolution network extracting and generating a high-dimensional feature matrix H out characterizing fault propagation characteristics by performing the following spatial convolution operation and time convolution operation;

[0068] The spatial convolution operation aggregates node information using the connection relationship between devices, and the calculation formula is:

[0069]

[0070] wherein H t ′ is a new feature matrix, t is a discrete step of time, σ is a nonlinear activation function, is an adjacency matrix of the spatio-temporal data graph G st and identity matrix I is a diagonal matrix of G , X is an initial node feature matrix, and W s is a weight matrix;

[0071] The time convolution operation uses a gated recurrent unit to capture the time sequence dependence of each device's own data to update the hidden state of the node.

[0072] Sub-step 4.3, inputting the high-dimensional feature matrix H out into a fault location and classification module, the module first calculating attention weight coefficients a st of each device node in the spatio-temporal data graph G i ,

[0073] When the attention weight coefficient α of the device node is greater than a preset positioning threshold T i , the identity information of the device node is determined as the fault location data, and the feature vector of the device node is input into a Softmax classifier to generate the fault type data, and finally combined into the fault diagnosis result. loc

[0074] Preferably, the time convolution operation is used to process the time sequence features of each device node to capture dynamic change rules.

[0075] For each node, the input is the feature H t ′ after spatial convolution, and the output is the updated node hidden state h t .

[0076] The reset gate determines how much previous node information to ignore:

[0077] r t =σ g (W r H t ′ +U r h t-1 +b r ),

[0078] The update gate decides how much new node information to write into the hidden state:

[0079] z t =σ g (W z H t ′ +U z h t-1 +b z ),

[0080] The candidate hidden state calculates the candidate information at the current time:

[0081]

[0082] The final hidden state combines historical information and current candidate information to obtain the final output of the current node:

[0083]

[0084] where h t-1 is the node hidden state at the previous time step t-1, r t and z t are the output vectors of the reset gate and the update gate, respectively, W r , W z , and W h ​and U r , U z , U h is the weight matrix that the model needs to learn, b r , b z , b h is the corresponding bias vector, σ g is the Sigmoid activation function, φ h is the tanh activation function, and is the Hadamard product. is the candidate hidden state vector, h t is the final hidden state output at time step t.

[0085] The graph attention layer is used to calculate the importance of each device node in the fault event to locate the root cause of the fault.

[0086] Calculate the attention score: first, apply the shared linear transformation W att to the feature vector h i of each node, then calculate the unnormalized attention score u i using the attention vector v:

[0087] u i = v T tanh(W att h i +b att ),

[0088] Calculate the attention weight coefficient: normalize the attention scores of all nodes using the Softmax function to get the final attention weight coefficient a i :

[0089]

[0090] where h i is the feature vector of the i-th device node, W att and b att are the weight matrix and bias vector, v is the context vector, u i is the unnormalized attention score of the i-th node, u j is the unnormalized attention score of the j-th node, and a i is the final attention weight coefficient of the i-th node.

[0091] The Softmax classifier formula, after determining the fault location node through the attention mechanism, inputs the feature vector h k into the Softmax classifier to determine the specific fault type:

[0092]

[0093] z = W cls h k +b cls ,

[0094] where h k is the feature vector of the fault location node k, W cls and b cls are the weight matrix and bias vector of the classifier fully connected layer, z is the log probability vector of the fully connected layer output, z c is the c-th component for the fault type, C is the total number of fault types,

[0095] P(y = c | h k ) is the probability of predicting the fault type c given the feature vector h k of node k.

[0096] Preferably, in step 5, the fault diagnosis result is input into the digital twin model for verification, the verified fault diagnosis result is combined with the corresponding multi-modal sensor data to form a fault sample, and the first data processing model and the second data processing model are updated using the fault sample, further comprising:

[0097] Sub-step 5.1, according to the fault location data in the fault diagnosis result, locate the corresponding twin equipment model in the digital twin model, convert the fault type data into simulated fault parameters, inject the simulated fault parameters into the twin equipment model to drive physical simulation, generate simulated sensor data, and calculate the verification error score E ver between the simulated sensor data and the multi-modal sensor data according to the following formula:

[0098]

[0099] where, is the true measurement value of the i-th dimension, is the simulated value of the i-th dimension, N is the total dimension of the multi-modal sensor data, E ver is the normalized verification error score;

[0100] Sub-step 5.2, compare the verification error score E ver with the preset verification confidence threshold T ver , when the verification error score E ver is less than the verification confidence threshold T ver , mark the fault diagnosis result as verified, pair the verified fault diagnosis result with the multi-modal sensor data that triggered the diagnosis, and jointly encapsulate to form the fault sample with true value label;

[0101] Sub-step 5.3, add the fault sample to the historical fault database, trigger the online updating process of the first data processing model and the second data processing model, the online updating process is to adjust the weight parameters in the model by the gradient back propagation of the loss function L, until the numerical value of the loss function L converges to the preset stable interval, the calculation formula of the loss function L is:

[0102]

[0103] Wherein, L is the cross entropy loss function value, M is the number of fault samples, C is the total number of fault types, y ic Is a symbol function, p ic Is the prediction probability of the i-th fault sample belonging to category c.

[0104] The application provides a smart grid fault monitoring method based on artificial intelligence. The application has the following advantages:

[0105] 1. The application adopts an edge computing node to deploy a lightweight model and a hierarchical early warning linkage technical solution, achieving real-time response and rapid fault isolation technical effects, compared with the cloud centralized processing scheme in the prior art, solving the problem of protection action lag caused by fault diagnosis delay.

[0106] 2. The application adopts a multi-modal gas sensing and spatio-temporal correlation analysis technical solution, achieving early gas leakage accurate identification technical effects, compared with the single electrical quantity monitoring scheme in the prior art, solving the problem of safety hidden danger caused by the lack of implicit fault early warning.

[0107] 3. The application adopts a high-frequency sensing acquisition and dynamic threshold self-adaptive adjustment technical solution, achieving gradual abnormal early warning technical effects, compared with the passive threshold monitoring scheme in the prior art, solving the problem of equipment internal progressive damage that cannot be sensed in time.

[0108] 4. The application adopts a gas-temperature-electricity multi-dimensional fusion sensing technical solution, achieving three-dimensional monitoring of the internal state of the equipment, compared with the two-dimensional electrical monitoring scheme in the prior art, solving the problem of single monitoring dimension of safety hidden danger in complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0109] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0110] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings of the embodiments of the present application, and obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, other embodiments obtained by those skilled in the art without creative efforts shall fall within the scope of the present application.

[0111] The present application will be described in detail in the following with reference to the accompanying drawings:

[0112] Embodiments:

[0113] Please refer to the accompanying drawings Figure 1 , the embodiments of the present application provide a smart grid fault monitoring method based on artificial intelligence, comprising:

[0114] Step 1, collecting multi-modal sensing data through a multi-modal sensing network deployed on a power grid device, the multi-modal sensing data including gas feature data collected by an olfactory chip array, temperature data collected by a distributed optical fiber sensor, and electrical data collected by a current transformer;

[0115] The olfactory chip array includes a metal oxide semiconductor gas sensor unit, an electrochemical sensor unit, and a photoionization detector unit, for detecting flammable gas, toxic gas, and smoke particles, and realizing ppd-level detection of gas types and concentrations;

[0116] Substep 1.1, time synchronization calibration is performed on all olfactory chip arrays, distributed optical fiber sensors, and current transformers in the multi-modal sensing network, to generate a synchronization timestamp associated with all subsequent collected data;

[0117] Substep 1.2, based on the synchronization timestamp, data collection and validity discrimination are performed, and the data collection and validity discrimination specifically include:

[0118] The original resistance signal R s reflecting the gas environment is collected by the olfactory chip array, the baseline resistance signal R0 in a clean air environment is obtained, when the value of the baseline resistance signal R0 is within a preset sensor health interval [R min , R max ], a gas data validity flag is generated, and a gas response value S is calculated according to the formula: S=(R s -R0) / R0;

[0119] Wherein, R min and R max are the minimum and maximum values of the baseline resistance allowed;

[0120] The Brillouin scattering light signal is collected by the distributed optical fiber sensor, and the Brillouin frequency shift amount Δν along the optical fiber path is demodulated B According to the linear relationship:

[0121] The temperature value ΔT distributed along the line is calculated, wherein C T is the temperature sensitivity coefficient;

[0122] The instantaneous current waveform data i(t) of a complete cycle is collected by the current transformer, and the current harmonic vector I h containing the fundamental wave and each harmonic amplitude is calculated by fast Fourier transform;

[0123] Substep 1.3, judge the validity flag of the gas data, when the validity flag of the gas data is true, calculate the data quality index Q according to the following data quality evaluation formula, when the data quality index Q is greater than a quality threshold Q th , the gas response value S, the temperature value ΔT and the current harmonic vector I h are packaged together with the synchronization timestamp as frame multimodal sensing data;

[0124] The calculation formula of the data quality index Q is:

[0125] Q=w g ·Q g +w t ·Q t +w e ·Q e ,

[0126] Wherein, w g , w t , w e are weight coefficients, Q g , Q t , Q e are the normalized quality scores calculated by the signal-to-noise ratio of the gas response value S, the volatility of the temperature value ΔT and the stability of the current harmonic vector I h ;

[0127] Step 2, send the multimodal sensing data to the edge computing node, the edge computing node is built-in first data processing model, the multimodal sensing data is processed by the first data processing model, and the abnormal probability score is generated;

[0128] Substep 2.1, normalize the gas response value S, the temperature value ΔT and the current harmonic vector I h in the multimodal sensing data to generate the normalized feature vector V norm , the normalized feature vector V normInput to the feature fusion module, the feature fusion module performs weighted summation on the features of different modes through the preset weight coefficient to generate a fusion feature vector V fused ;

[0129] Sub-step 2.2, the fusion feature vector V fused As the input of the first data processing model, the first data processing model is a one-dimensional convolutional neural network, which uses the convolution kernel to fuse the feature vector V fused Perform convolution operations and extract feature maps M containing deep information about the device state through nonlinear activation function processing. feature , the calculation formula of the convolution operation is:

[0130]

[0131] Among them, M feature (j) is the jth element in the characteristic spectrum, V fused is the fusion feature vector, V fused (j+i-1) represents the j+i-1th element in the fused feature vector, K(i) represents the i-th weight parameter in the convolution kernel, k represents the size of the convolution kernel, and b k is the bias term corresponding to the convolution kernel, f act is a nonlinear activation function;

[0132] Sub-step 2.3, the feature map M feature The state feature vector V is obtained by performing dimensionality reduction processing through the global average pooling layer. state , the state feature vector V state Send it to the fully connected layer and calculate the abnormal probability score P through the Sigmoid activation function anomaly , abnormal probability score P anomaly The calculation formula is:

[0133]

[0134] Among them, V state is the state feature vector, W is the weight matrix of the fully connected layer, c is the bias term of the fully connected layer, P anomaly Score the probability of anomaly;

[0135] The gas response value S is calculated by the following formula to obtain the normalized gas characteristic S norm :

[0136]

[0137] The temperature value ΔT is calculated by the following formula to obtain the normalized temperature characteristic ΔT norm :

[0138]

[0139] The current harmonic vector I h is calculated by the following formula: h,n The normalized current harmonic vector I h,n,norm is calculated by the following formula:

[0140]

[0141] The normalized feature vector V norm is generated by splicing the normalized modal features in a predetermined order: norm

[0142] V norm = [S norm , ΔT norm , I h,1,norm , I h,2,norm , …, I h,N,norm ],

[0143] wherein S, ΔT, I h,n are the gas response value, the temperature value, and the amplitude of the nth current harmonic, respectively, S min and S max are the minimum and maximum values of the gas response value S, respectively,

[0144] ΔT min and ΔT max are the minimum and maximum values of the temperature value ΔT, respectively,

[0145] I h,n,min and I h,n,max are the minimum and maximum values of the nth current harmonic amplitude I h,n , respectively,

[0146] N is the highest order of the current harmonic;

[0147] The fusion feature vector V fused is generated by weighting and adjusting the different modal feature components in the normalized feature vector V norm by assigning different weight coefficients to them;

[0148] The calculation formula of the fusion feature vector V fused is:

[0149] V fused = [w g ·S norm , w t ·ΔT norm , w e ·I h,1,norm , w e ·I​h,2,norm ,…,w e ·

[0150] I h,N,norm ],

[0151] wherein, V fused is the fusion feature vector, w g is the weight coefficient of the gas modal, w t is the weight coefficient of the temperature modal, w e is the weight coefficient of the electrical modal;

[0152] Step 3, compare the abnormal probability score with the first preset threshold value in the edge computing node, when the abnormal probability score is greater than the first preset threshold value, send the abnormal probability score together with the multi-modal sensing data to the cloud server, the cloud server is built-in second data processing model;

[0153] Substep 3.1, extract the ambient temperature and ambient humidity contained in the multi-modal sensing data, call the pre-stored cumulative running time data of the power grid equipment, according to the following adaptive threshold calculation formula, real-time generate the first preset threshold value, the first preset threshold value is a dynamic alarm threshold T dyn ;

[0154] T dyn =T base ·(1+α·f age +β·f env ),

[0155] wherein, T dyn is the dynamic alarm threshold, T base is the reference alarm threshold, f age is the aging factor, f env is the environmental interference factor, and α and β are weight coefficients;

[0156] Substep 3.2, compare the abnormal probability score P anomaly with the dynamic alarm threshold T dyn and the local early warning threshold T local whose value is less than the dynamic alarm threshold in sequence, generate the early warning level flag L alert according to the following determination rule:

[0157] When the abnormal probability score P anomaly is greater than the dynamic alarm threshold T dyn , set the early warning level flag L alert to high level;

[0158] When the local early warning threshold T local is less than the abnormal probability score P anomaly and the abnormal probability score P anomalynot greater than the dynamic alarm threshold T dyn the warning level flag L alert is set to the medium level;

[0159] when the abnormal probability score P anomaly is not greater than the local alarm threshold T local the warning level flag L alert is set to the low level;

[0160] Sub-step 3.3, determining the warning level flag L alert , when the warning level flag L alert is the high level, the abnormal probability score P anomaly , the multi-modal sensor data triggering the score, the synchronization timestamp, and the dynamic alarm threshold T dyn are jointly encapsulated into an alarm event data packet, and the alarm event data packet is sent to a cloud server through an encrypted channel;

[0161] Step 4, in the cloud server, a second data processing model is used to analyze the received abnormal probability score and multi-modal sensor data, to generate a fault diagnosis result containing fault type data and fault location data;

[0162] The generation of the fault diagnosis result further includes: constructing a power grid fault knowledge graph, taking device entities and fault types as nodes, and taking fault propagation paths as edges, using a graph attention network to perform reasoning on the knowledge graph to locate the root cause of the fault;

[0163] Sub-step 4.1, according to the device identity information in the alarm event data packet, retrieving the adjacent devices directly connected to the device from a pre-constructed power grid topology structure diagram, calling the historical multi-modal sensor data of the adjacent devices within a preset time window before the synchronization timestamp, combining the alarm event data packet and the historical multi-modal sensor data, and constructing a spatio-temporal data graph G st ;

[0164] Sub-step 4.2, inputting the spatio-temporal data graph G st into a second data processing model, and the second data processing model is a spatio-temporal graph convolution network, which extracts and generates a high-dimensional feature matrix H out characterizing fault propagation characteristics by performing the following spatial convolution operation and temporal convolution operation;

[0165] The spatial convolution operation aggregates node information using the connection relationship between devices, and the calculation formula is:

[0166]

[0167] H t ′is the new feature matrix, t is the discrete step of time, and σ is a nonlinear activation function, is the adjacency matrix of the spatio-temporal data graph G st and the identity matrix, and is the diagonal matrix of , X is the initial node feature matrix, and W s is the weight matrix;

[0168] The time convolution operation captures the temporal dependency of each device's own data using a gated recurrent unit to update the hidden state of the node.

[0169] Sub-step 4.3: input the high-dimensional feature matrix H out into the fault location and classification module. The module first calculates the attention weight coefficient α st of each device node in the spatio-temporal data graph G i ,

[0170] When the attention weight coefficient α i of the device node is greater than a preset positioning threshold T loc , the identity information of the device node is determined as the fault location data, and the feature vector of the device node is input into the Softmax classifier to generate the fault type data, which is finally combined into the fault diagnosis result.

[0171] The time convolution operation is used to process the time series features of each device node to capture the dynamic change rule.

[0172] For each node, the input is the spatial convolution feature H t ′ , and the output is the updated node hidden state h t .

[0173] The reset gate determines how much previous node information to ignore:

[0174] r t = σ g (W r H t ′ + U r h t-1 + b r ),

[0175] The update gate decides how much new node information to write into the hidden state:

[0176] z t = σ g (W z H t ′ + U z h t-1 + bz ),

[0177] candidate hidden state, calculate the candidate information at the current time:

[0178]

[0179] final hidden state, combine the historical information and the current candidate information to obtain the final output of the current node:

[0180]

[0181] where h t-1 is the node hidden state at the previous time step t-1, r t and z t are the output vectors of the reset gate and the update gate in turn, W r , W z , W h and U r , U z , U h are weight matrices that need to be learned by the model, b r , b z , b h are the corresponding bias vectors, σ g is the Sigmoid activation function, φ h is the tanh activation function, and is the Hadamard product, is the candidate hidden state vector, and h t is the final hidden state output at time step t.

[0182] The graph attention layer is used to calculate the importance of each device node in the fault event to locate the root cause of the fault;

[0183] Calculate the attention score: first, apply the shared linear transformation W att to the feature vector h i of each node, and then calculate the unnormalized attention score u i using the attention vector v:

[0184] u i = v T tanh(W att h i +b att ),

[0185] Calculate the attention weight coefficient: normalize the attention scores of all nodes using the Softmax function to obtain the final attention weight coefficient a i :

[0186]

[0187] where h i is the feature vector of the i-th device node, W att is the weight matrix and b att is the bias vector, v is the context vector, u i is the un-normalized attention score of the i-th node, u j is the un-normalized attention score of the j-th node, and a i is the final attention weight coefficient of the i-th node.

[0188] Softmax classifier formula, after determining the fault location node through the attention mechanism, the feature vector h k is input into the Softmax classifier to determine the specific fault type:

[0189]

[0190] z = W cls h k + b cls ,

[0191] where h k is the feature vector of the fault location node k, W cls is the weight matrix and b cls is the bias vector of the classifier fully connected layer, z is the log probability vector output by the fully connected layer, and z c is the c-th component of the fault type, and C is the total number of fault types.

[0192] P(y = c|h k ) is the probability of predicting the fault type as c given the feature vector h k of the node k.

[0193] Step 5, input the fault diagnosis result into the digital twin model for verification, combine the verified fault diagnosis result with the corresponding multi-modal sensor data into a fault sample, and update the first data processing model and the second data processing model using the fault sample.

[0194] Sub-step 5.1, according to the fault location data in the fault diagnosis result, locate the corresponding twin device model in the digital twin model, convert the fault type data into simulated fault parameters, inject the simulated fault parameters into the twin device model to drive physical simulation, generate simulated sensor data, and calculate the verification error score E ver between the simulated sensor data and the multi-modal sensor data according to the following formula:

[0195]

[0196] where, is the true measurement value of the i-th dimension, is the analog value of the i-th dimension, N is the total dimension of the multi-modal sensor data, E ver is the normalized verification error score;

[0197] Sub-step 5.2, comparing the verification error score E ver with a preset verification confidence threshold T ver , when the verification error score E ver is less than the verification confidence threshold T ver , marking the fault diagnosis result as verified, and pairing the verified fault diagnosis result with the multi-modal sensor data that triggered the diagnosis to form a fault sample with a true value label.

[0198] Sub-step 5.3, adding the fault sample to the historical fault database, triggering the online updating process of the first data processing model and the second data processing model, and adjusting the weight parameters inside the model through the gradient backpropagation of the following loss function L until the numerical value of the loss function L converges to a preset stable interval, and the calculation formula of the loss function L is:

[0199]

[0200] where L is the cross-entropy loss function value, M is the number of fault samples, C is the total number of fault types, y ic is the symbol function, p ic is the predicted probability of the i-th fault sample belonging to class c.

[0201] Step 1 integrates gas, temperature, and electrical multi-dimensional data collection through a multi-modal sensor network, combines time synchronization calibration and data validity discrimination mechanism, and realizes all-around perception of equipment state. Compared with single-dimensional monitoring, this scheme breaks through the traditional data island restriction, significantly improves the coverage rate of hidden danger detection through multi-source heterogeneous data fusion. Time stamp synchronization technology ensures the spatio-temporal consistency of cross-modal data, providing accurate input for subsequent analysis, while dynamic quality evaluation mechanism effectively filters abnormal sensor data, ensuring the reliability of input data.

[0202] Step 2 deploys lightweight neural network models on edge computing nodes to realize local data processing and real-time feature extraction. Through normalization processing and multi-modal feature weighted fusion, the influence of dimension difference on the model is eliminated, and the feature expression ability is enhanced.

[0203] Step 3 introduces a dynamic threshold adjustment algorithm and a three-level early warning mechanism, breaking through the limitations of traditional fixed thresholds. By dynamically correcting the alarm threshold based on environmental parameters and equipment aging factors, the false alarm suppression capability under complex working conditions is significantly improved. The hierarchical early warning strategy realizes the cooperation of local light alarm and cloud deep analysis, optimizes the allocation of computing resources, ensures the priority processing of high threat faults, and builds a flexible emergency response system.

[0204] Step 4 realizes dynamic modeling of fault propagation path by integrating device topology relationship and historical time series data through spatio-temporal graph convolution network. The combination of knowledge graph and graph attention mechanism breaks through the limitations of traditional isolated node analysis, accurately locates the root cause device of the fault. Multi-dimensional feature reasoning technology synchronously analyzes the fault type and location, forming a three-dimensional diagnosis result of device-fault-impact, providing stereoscopic data support for operation and maintenance decision-making.

[0205] Step 5 builds a closed-loop verification system driven by digital twin, and verifies the credibility of the diagnosis result through physical simulation. The error evaluation mechanism selects high-confidence samples to inject into model training, forming a self-evolution link of data collection-diagnosis-verification-optimization. Dynamic incremental learning strategy continuously improves the model's adaptability to new types of faults, realizes the autonomous iterative upgrade of the monitoring system's intelligence level, and breaks through the generalization bottleneck of traditional static models.

[0206] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart grid fault monitoring method based on artificial intelligence, characterized in that: include: Step 1: Collect multimodal sensor data through a multimodal sensor network deployed on power grid equipment. The multimodal sensor data includes gas characteristic data collected by an olfactory chip array, temperature data collected by distributed optical fiber sensors, and electrical data collected by current transformers. Step 2: Send the multimodal sensor data to an edge computing node. The edge computing node has a built-in first data processing model. The first data processing model processes the multimodal sensor data to generate an abnormality probability score. Step 3: Compare the abnormality probability score with a first preset threshold in the edge computing node. When the abnormality probability score is greater than the first preset threshold, send the abnormality probability score together with the multimodal sensor data to a cloud server, wherein the cloud server has a built-in second data processing model. Step 4: In the cloud server, the second data processing model analyzes the received abnormality probability score and the multimodal sensing data to generate a fault diagnosis result including fault type data and fault location data; Step 5: Input the fault diagnosis result into the digital twin model for verification, combine the verified fault diagnosis result with the corresponding multimodal sensing data into a fault sample, and use the fault sample to update the first data processing model and the second data processing model.

2. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: The olfactory chip array includes a metal oxide semiconductor gas sensor unit, an electrochemical sensor unit and a photoionization detector unit, which are used to detect combustible gases, toxic gases and smoke particles, and achieve ppd-level detection of gas types and concentrations.

3. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: The generation of the fault diagnosis results further includes: constructing a power grid fault knowledge graph, wherein the knowledge graph uses equipment entities and fault types as nodes, and fault propagation paths as edges, and uses a graph attention network to perform reasoning on the knowledge graph to locate the root cause of the fault.

4. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: In step 1, collecting multimodal sensor data through a multimodal sensor network deployed on power grid equipment further includes: Sub-step 1.1, performing time synchronization calibration on all the olfactory chip arrays, the distributed optical fiber sensors, and the current transformers in the multimodal sensing network to generate a synchronization timestamp associated with all subsequent collected data; Sub-step 1.2: performing data collection and validity determination based on the synchronization timestamp. The data collection and validity determination are specifically as follows: The olfactory chip array collects the original resistance signal R reflecting the gas environment s , obtain the pre-stored baseline resistance signal R0 in the clean air environment, when the value of the baseline resistance signal R0 is within the preset sensor health range [R min ,R max ], the gas data validity flag is generated, and according to the formula: S=(R s -R0) / R0, calculate the gas response value S; Among them, R min and R max is the minimum and maximum value allowed for the baseline resistance; The distributed optical fiber sensor collects the Brillouin scattered light signal and demodulates the Brillouin frequency shift Δν along the optical fiber path. B , according to the linear relationship: Calculate the temperature value ΔT along the line, where C T is the temperature sensitivity coefficient; The instantaneous current waveform data i(t) of a complete cycle is collected by the current transformer, and the current harmonic vector I containing the fundamental wave and each harmonic amplitude is obtained by fast Fourier transform calculation. h ; Sub-step 1.3, determining the gas data validity flag, when the gas data validity flag is true, calculating the data quality index Q according to the following data quality evaluation formula, when the data quality index Q is greater than a quality threshold Q th When the gas response value S, the temperature value ΔT and the current harmonic vector I h Together with the synchronization timestamp, they are encapsulated into frame multimodal sensing data; The calculation formula of the data quality index Q is: Q=w g ·Q g +w t ·Q t +w e ·Q e , Among them, w g 、w t 、w e is the weight coefficient, Q g , Q t , Q e They are the signal-to-noise ratio of the gas response value S, the fluctuation of the temperature value ΔT, and the current harmonic vector I h The normalized mass score obtained by calculating the stability of 5. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: In step 2, the multimodal sensor data is sent to an edge computing node, the edge computing node has a built-in first data processing model, and the first data processing model processes the multimodal sensor data to generate an abnormality probability score, further comprising: Sub-step 2.1, the gas response value S, the temperature value ΔT and the current harmonic vector I in the multimodal sensing data h Perform normalization to generate normalized feature vector V norm , the normalized feature vector V norm Input to the feature fusion module, the feature fusion module performs weighted summation on the features of different modes through the preset weight coefficient to generate a fused feature vector V fused ; Sub-step 2.2, the fused feature vector V fused As the input of the first data processing model, the first data processing model is a one-dimensional convolutional neural network, which uses a convolution kernel to process the fused feature vector V fused Perform convolution operations and extract feature maps M containing deep information about the device state through nonlinear activation function processing. feature , the calculation formula of the convolution operation is: Among them, M feature (j) is the jth element in the characteristic spectrum, V fused is the fusion feature vector, V fused (j+i-1) represents the j+i-1th element in the fused feature vector, K(i) represents the i-th weight parameter in the convolution kernel, k represents the size of the convolution kernel, and b k is the bias term corresponding to the convolution kernel, f act is a nonlinear activation function; Sub-step 2.3, the characteristic map M feature The state feature vector V is obtained by performing dimensionality reduction processing through the global average pooling layer. state , the state feature vector V state Send it to the fully connected layer and calculate the abnormal probability score P through the Sigmoid activation function anomaly , the abnormal probability score P anomaly The calculation formula is: Among them, V state is the state feature vector, W is the weight matrix of the fully connected layer, c is the bias term of the fully connected layer, P anomaly Score the probability of anomaly.

6. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 5, characterized in that: The gas response value S is calculated by the following formula to obtain the normalized gas characteristic S norm : The temperature value ΔT is calculated by the following formula to obtain the normalized temperature characteristic ΔT norm : The current harmonic vector I h The amplitude of each harmonic in I h,n The normalized current harmonic vector I is calculated by the following formula: h,n,norm : Splicing generates normalized feature vector V norm , the normalized modal features are spliced ​​in a predetermined order to generate the normalized feature vector V norm : V norm =[S norm ,ΔT norm ,I h,1,norm ,I h,2,norm ,…,I h,N,norm ], Among them, S, ΔT, I h,n The gas response value, temperature value and the amplitude of the nth current harmonic are in turn, S min 、S max They are the minimum and maximum values ​​of the gas response value S, respectively. ΔT min , ΔT max are the minimum and maximum values ​​of the temperature ΔT, respectively. I h,n,min , I h,n,max The nth current harmonic amplitude I h,n The minimum and maximum values ​​of N is the highest order of current harmonics; The fused feature vector V fused is obtained by normalizing the eigenvector V norm Different modal feature components in the model are given different weight coefficients and generated after weighted adjustment; The fused feature vector V fused The calculation formula is: V fused =[w g ·S norm ,w t ·ΔT norm ,w e ·I h,1,norm ,w e ·I h,2,norm ,…,w e · I h,N,norm ]， Among them, V fused is the fusion feature vector, w g is the weight coefficient of the gas mode, w t is the weight coefficient of the temperature mode, w e is the weight coefficient of the electrical mode.

7. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: In step 3, the abnormality probability score is compared with a first preset threshold in the edge computing node. When the abnormality probability score is greater than the first preset threshold, the abnormality probability score is sent together with the multimodal sensor data to a cloud server. The cloud server has a built-in second data processing model, further comprising: Sub-step 3.1, extracting the ambient temperature and ambient humidity contained in the multimodal sensor data, calling the pre-stored power grid equipment cumulative operating time data, and generating the first preset threshold in real time according to the following adaptive threshold calculation formula. The first preset threshold is the dynamic alarm threshold T dyn ; T dyn =T base ·(1+α·f age +β·f env ), Among them, T dyn is the dynamic alarm threshold, T base is the baseline alarm threshold, f age is the aging factor, f env is the environmental interference factor, α and β are weight coefficients; Sub-step 3.2, the abnormal probability score P anomaly With the dynamic alarm threshold T dyn And the local warning threshold T whose value is less than the dynamic alarm threshold local Perform sequential comparison and generate the warning level mark L according to the following judgment rules alert : When the abnormal probability score P anomaly Greater than the dynamic alarm threshold T dyn When the warning level mark L alert Set to high level; When the local warning threshold T local Less than the abnormal probability score P anomaly And the abnormal probability score P anomaly Not greater than the dynamic alarm threshold T dyn When the warning level mark L alert Set to medium level; When the abnormal probability score P anomaly Not greater than the local warning threshold T local When the warning level mark L alert Set to low level; Sub-step 3.3, the warning level mark L alert Make a judgment, when the warning level mark L alert When the level is high, the abnormal probability score P anomaly , the multimodal sensor data that triggers scoring, the synchronization timestamp, and the dynamic alarm threshold T dyn The alarm event data packet is encapsulated together and sent to the cloud server through an encrypted channel.

8. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: In step 4, in the cloud server, the second data processing model analyzes the received abnormality probability score and the multimodal sensor data to generate a fault diagnosis result including fault type data and fault location data, further comprising: Sub-step 4.1: Based on the device identity information in the alarm event data packet, retrieve the neighboring devices directly connected to the device from the pre-constructed power grid topology diagram, retrieve the historical multimodal sensor data of the neighboring devices within a preset time window before the synchronization timestamp, combine the alarm event data packet with the historical multimodal sensor data, and construct a spatiotemporal data graph G. st ; Sub-step 4.2, the spatiotemporal data graph G st Input the second data processing model, which is a spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network extracts and generates a high-dimensional feature matrix H that can characterize the fault propagation characteristics by performing the following spatial convolution operation and temporal convolution operation: out ; The spatial convolution operation aggregates node information using the connection relationship between devices. The calculation formula is: Among them, H′ t is the new feature moment, t is the discrete step of time, σ is the nonlinear activation function, is the spatiotemporal data graph G st The adjacency matrix of and the identity matrix, for The diagonal matrix, X is the initial node feature matrix, W s is the weight matrix; The temporal convolution operation uses a gated recurrent unit to capture the temporal dependencies of each device's own data and update the hidden state of the node; Sub-step 4.3, the high-dimensional feature matrix H out Input to the fault location and classification module, the module first calculates the spatiotemporal data graph G through the graph attention layer st The attention weight coefficient α of each device node i , When the attention weight coefficient α of the device node i Greater than a preset positioning threshold T loc When the device node's identity information is determined as the fault location data, the feature vector of the device node is input into the Softmax classifier to generate the fault type data, which are finally combined into the fault diagnosis result.

9. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 8, characterized in that: The temporal convolution operation is used to process the time series characteristics of each device node and capture the dynamic change rules; For each node, the input is the feature H′ after spatial convolution t , the output is the updated node hidden state h t ; Reset the gate to determine how much previous node information to ignore: r t =σ g (W r H′ t +U r h t-1 +b r ), Update gate, which determines how much new node information is written to the hidden state: z t =σ g (W z H′ t +U z h t-1 +b z ), Candidate hidden state, calculate the candidate information at the current moment: The final hidden state combines historical information with the current candidate information to obtain the final output of the current node: Among them, h t-1 is the node hidden state at the previous time step t-1, r t and z t The output vectors of the reset gate and update gate are W r 、W z 、W h and U r 、U z 、U h is the weight matrix that the model needs to learn, b r 、b z 、b h is the corresponding bias vector, σ g is the Sigmoid activation function, φ h is the tanh activation function, ⊙ is the Hadamard product, is the candidate hidden state vector, h t is the final hidden state output at time step t; The graph attention layer is used to calculate the importance of each device node in a fault event to locate the root cause of the fault; Calculate the attention score: First, through the shared linear transformation W att The characteristic vector h acting on each node i , and then calculate the unnormalized attention score u through the attention vector v i : u i =v T fishy(W att h i +b att ), Calculate the attention weight coefficient: Use the Softmax function to normalize the attention scores of all nodes to obtain the final attention weight coefficient α i : Among them, h i is the characteristic vector of the i-th device node, W att and b att is the weight matrix and bias vector, v is the context vector, u i is the unnormalized attention score of the i-th node, u j is the unnormalized attention score of the j-th node, α i is the final attention weight coefficient of the i-th node; The Softmax classifier formula, after determining the fault location node through the attention mechanism, converts the feature vector h k Enter the Softmax classifier to determine the specific fault type: z=W cls h k +b cls , Among them, h k is the eigenvector of the fault location node k, W cls and b cls is the weight matrix and bias vector of the fully connected layer of the classifier, z is the logarithmic probability vector output by the fully connected layer, z c is the cth fault type component, C is the total number of fault types, P(y=c|h k ) is the feature vector h of a given node k k Under the condition of , the probability of predicting the fault type is c.

10. The method for monitoring faults in a smart grid based on artificial intelligence according to claim 1, characterized in that: In step 5, the fault diagnosis result is input into the digital twin model for verification, the verified fault diagnosis result is combined with the corresponding multimodal sensor data into a fault sample, and the first data processing model and the second data processing model are updated using the fault sample, further comprising: Sub-step 5.1, based on the fault location data in the fault diagnosis result, locate the corresponding twin device model in the digital twin model, convert the fault type data into simulated fault parameters, inject the simulated fault parameters into the twin device model to drive physical simulation, generate simulated sensor data, and calculate the verification error score E between the simulated sensor data and the multimodal sensor data according to the following formula ver ; in, is the true measurement value of the i-th dimension, is the analog value of the i-th dimension, N is the total dimension of the multimodal sensor data, E ver is the normalized validation error score; Sub-step 5.2: the verification error score E ver and the preset verification credibility threshold T ver When compared, the validation error score E ver Less than the verification credibility threshold T ver When the fault diagnosis result is marked as verified, the verified fault diagnosis result is paired with the multimodal sensing data that triggers the diagnosis, and the two are encapsulated together to form the fault sample with a true value label; Sub-step 5.3: Add the fault sample to the historical fault database, triggering the online update process of the first data processing model and the second data processing model. The online update process adjusts the weight parameters within the model by backpropagating the gradient of the following loss function L until the value of the loss function L converges to a preset stable range. The calculation formula of the loss function L is: Among them, L is the cross entropy loss function value, M is the number of fault samples, C is the total number of fault types, and y ic is the symbolic function, p ic is the predicted probability that the i-th fault sample belongs to category c.

Images