Transformer substation inspection method, system and equipment and storage medium

Abstract

The invention relates to the technical field of substation inspection. The invention provides a substation inspection method, system and device and a storage medium. The method comprises the following steps: controlling an inspection robot to move to the front of a target equipment cabinet, and triggering an access control state recognition mechanism; the cabinet door state is analyzed, and if the cabinet door state is recognized as a closed state, a cabinet opening instruction is generated and sent to the execution mechanism; executing a cabinet opening action, and collecting state data of target equipment in the cabinet in real time; according to the state data collected in real time, defects existing in target equipment in the cabinet are identified, and abnormal features in the real-time data are analyzed; if the abnormal features are recognized, an early warning signal is generated, and an alarm notification is triggered; and executing a cabinet closing action, updating the equipment state database, and generating a periodic inspection report and equipment operation state evaluation. The problems that a monitoring blind area exists in equipment in a cabinet due to the fact that an existing rail type inspection robot cannot open an equipment cabinet door due to lack of active operation capacity, and traditional manual inspection is low in efficiency, poor in accuracy and lagged in data are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation inspection, in particular to a substation inspection method, system, device and storage medium. BACKGROUND

[0002] In the daily operation and maintenance of substation in the industry of electrolytic aluminum and the like, the monitoring of the state of equipment and facilities highly depends on artificial periodic inspection. The staff needs to perform a large number of repetitive tasks including equipment temperature measurement, instrument transcription, defect tracking and data entry. This traditional method has significant problems: the inspection efficiency is low, it is difficult to meet the high frequency demand; artificial records are prone to missed inspection and misinspection, the data accuracy is difficult to guarantee; the collection and analysis of inspection data have serious lag, it is difficult to provide timely and effective decision support for real-time perception of equipment state, abnormal warning and accident prevention, which restricts the improvement of intelligent operation and maintenance level of substation.

[0003] In the prior art, there are track-type inspection robots to try to replace part of the artificial work, such as combined track-type indoor substation inspection robots. Such robots usually move along a preset track and collect data of exposed equipment on the path through fixed sensors. However, its core limitation is that the function is single: the robot body lacks active operation ability and cannot open the door of the equipment cabinet, which leads to the inability to directly and effectively monitor the running parameters, temperature and image information of the key instruments and meters (such as protection devices, indicating instruments, terminal state, etc.) in the cabinet. This makes the existing track inspection method have a monitoring blind area and cannot realize comprehensive and unmanned intelligent inspection of substation equipment (especially the equipment in the cabinet), and its application effect and automation degree still need to be greatly improved. SUMMARY

[0004] The purpose of the present application is to provide a substation inspection method, system, device and storage medium, which aims to solve the problem that the existing track-type inspection robot cannot open the door of the equipment cabinet due to the lack of active operation ability, resulting in a monitoring blind area in the cabinet, and the problem of low efficiency, poor accuracy and data lag of traditional artificial inspection.

[0005] The present application is realized by the following technical solutions: A substation inspection method, comprising the following steps: controlling an inspection robot to move along a preset track to the front of a target equipment cabinet, and triggering an access state recognition mechanism; analyzing the state of the cabinet door through the access state recognition mechanism, if it is identified as a closed state, generating an open cabinet instruction and sending it to the actuator of the inspection robot; in response to the open cabinet instruction, performing an open cabinet action, and collecting the state data of the target equipment in the cabinet in real time through the multi-modal collection module carried by the inspection robot; According to the real-time collected state data, the defects of the target equipment in the cabinet are identified, and the abnormal features in the real-time data are analyzed; If the abnormal features are identified, a pre-warning signal containing the equipment positioning, the abnormal type and the confidence is generated, and an alarm notification is triggered; After the data collection is completed, a closing cabinet action is performed, and the equipment state database is updated to generate a periodic inspection report and an equipment running state evaluation.

[0006] Optionally, the specific process of analyzing the cabinet door state through the access control state identification mechanism, if the closed state is identified, generating an opening cabinet instruction and sending it to the control terminal of the inspection robot's execution mechanism is as follows: The multi-modal acquisition module carried by the inspection robot acquires image streams of the cabinet door area in real time, and simultaneously acquires cabinet door distance data through a distance sensor; The image streams of the cabinet door area and the cabinet door distance data are input into a pre-trained door state classification model, which extracts cabinet door handle positioning features, door gap features and lock state features based on a convolutional neural network and a point cloud fusion algorithm, and obtains a feature fusion result; According to the feature fusion result, the cabinet door state confidence is dynamically calculated, if the confidence is greater than or equal to a first threshold value, and the closed state is identified, an opening cabinet instruction is generated; if the confidence is less than the first threshold value, the data is re-acquired; The opening cabinet instruction is associated with the spatial coordinates of the cabinet and the safety operation permission code, and is sent to the control terminal of the inspection robot's execution mechanism through an encrypted communication protocol.

[0007] Optionally, the specific process of executing the opening cabinet action in response to the opening cabinet instruction, and acquiring the state data of the target equipment in the cabinet through the multi-modal acquisition module carried by the inspection robot is as follows: After receiving the opening cabinet instruction, the control terminal of the execution mechanism performs decryption verification based on the safety operation permission code associated with the instruction; After the verification is passed, a path planning algorithm is called to generate an opening cabinet trajectory adapted to the door type of the target cabinet, and the mechanical arm end effector is positioned to the cabinet door handle; By monitoring the mechanical arm joint torque and visual feedback data in real time, the opening cabinet force is dynamically adjusted until the cabinet door reaches a preset opening angle; After the cabinet door is opened, the synchronous acquisition protocol of the multi-modal acquisition module is triggered: the infrared thermal imager acquires temperature distribution data of the equipment in the cabinet in a spatial grid scanning mode; the camera controls the adaptive focusing based on the equipment preset identification points, acquires the state images of the instrument readings and the wiring terminals; and the wireless probe reads the running parameters of the protection device; The temperature distribution data, the state images and the running parameters are input into a multi-source heterogeneous data fusion algorithm to generate a timestamp-aligned equipment state data set.

[0008] Optionally, the specific process of identifying defects of the target equipment in the cabinet according to the real-time collected state data and analyzing abnormal features in real-time data is as follows: Based on the equipment state data set, a pre-trained defect identification model is input, and the defect identification model is based on a multi-modal deep learning architecture, and temperature distribution data, state images and operating parameters are processed synchronously. Through a visual analysis module in the defect identification model, instrument reading deviation features and wiring terminal loosening or corrosion defect features in the state image are extracted, and a preliminary defect identification is generated based on a convolutional neural network classification algorithm. The real-time change trend of the operating parameters is analyzed, parameter deviation or fluctuation abnormal patterns are identified, and a thermal map analysis algorithm is applied to process the temperature distribution data to detect temperature hotspots exceeding a preset safety threshold. The preliminary defect identification, abnormal patterns and temperature hotspots are fused to generate a comprehensive defect probability vector and an abnormal feature vector through a feature-level fusion algorithm. According to the comprehensive defect probability vector, if the probability value exceeds a second threshold, it is determined that there is a defect in the target equipment in the cabinet. Based on the abnormal feature vector, the abnormal type, location and confidence are analyzed to provide analysis basis for generating a warning signal.

[0009] Optionally, the construction process of the defect identification model is as follows: A multi-modal sample data set of historical inspections of the substation is collected, and the sample data set includes temperature distribution data, state image samples, operating parameter time series labeled with defect labels, and corresponding defect type, location and severity label information. The multi-modal sample data set is preprocessed to generate a standardized multi-modal training data set. A multi-modal deep learning architecture is constructed, which includes a visual feature extraction branch, a parameter time series analysis branch and a temperature feature analysis branch; wherein the visual feature extraction branch adopts an improved convolutional neural network to extract instrument reading features and wiring terminal defect features in the state image through a residual module; the parameter time series analysis branch adopts a bidirectional recurrent neural network to extract trend features and fluctuation patterns from the operating parameter time series; and the temperature feature analysis branch combines a thermal map analysis algorithm and a convolutional neural network to extract hot region features and temperature gradient features in the temperature distribution data. A cross-modal feature fusion layer is set, an attention mechanism is used to dynamically allocate weights to feature vectors output by the visual feature extraction branch, the parameter time series analysis branch and the temperature feature analysis branch, and a multi-modal fusion feature is generated through a feature-level fusion algorithm. A defect classification and anomaly positioning sub-network is constructed based on the multi-modal fusion feature, the sub-network adopts a full connection layer and a softmax classifier, and outputs a defect type probability distribution, an anomaly position coordinate and a confidence score; The multi-modal training data set is divided into a training set and a validation set, the model is iteratively trained by using a joint loss function through the training set, and the weights of the branch networks are dynamically adjusted; The performance of the trained model is evaluated through the validation set, the defect recognition accuracy, the anomaly positioning accuracy and the confidence threshold stability are calculated, if the evaluation indexes do not reach the preset standard, the network depth, the fusion strategy or the loss function weight is adjusted and retrained; The optimized model is deployed to an inference engine, and online fine-tuning is performed through incremental samples of actual inspection scenes to generate a final defect recognition model.

[0010] Optionally, the device state database is updated, and the specific process of generating a periodic inspection report and a device operation state evaluation is as follows: The real-time collected state data and the historical baseline data stored in the device state database are time-aligned, and the long-term change trend features and the short-term fluctuation features of the device state are extracted; Based on the change trend features and the fluctuation features, the parameters of the defect recognition model are dynamically iteratively updated, and the recognition threshold and the feature weight of the defect recognition model for the progressive defects of the device are optimized; The device spatial positioning information, the historical defect record, the real-time anomaly feature and the early warning signal are associated and mapped through knowledge graph technology, a whole life cycle health file containing device health indicators and defect evolution paths is constructed and updated to the device state database; When generating a periodic inspection report, the multi-source data in the device health file are structurally analyzed, the key indicators, the anomaly statistics and the trend prediction results are extracted, and the device remaining life prediction value and the maintenance priority ranking are output.

[0011] Optionally, the specific process of triggering an alarm notification is as follows: Based on the anomaly type and the confidence contained in the early warning signal, the alarm level is dynamically determined; According to the alarm level and the preconfigured push rule, an alarm message containing device positioning, anomaly type, confidence and suggested treatment measures is generated; Through a multi-channel communication protocol, the alarm message is synchronously pushed to a mobile terminal of an operator, an alarm screen of a centralized control center and a device management system; The historical defect record in the device state database is associated, and the historical failure statistics and the processing plan link of the same type device are attached to the alarm message.

[0012] Based on the same inventive concept, the application also provides a transformer substation inspection system for implementing the transformer substation inspection method, comprising: The inspection robot comprises an inspection robot body and an execution mechanism, the inspection robot body is in sliding connection with the track, and the execution mechanism is arranged on the inspection robot body; the execution mechanism comprises a mechanical arm, an end effector and a control terminal; after receiving a cabinet opening instruction, the control terminal is configured to perform security operation permission code decryption verification, call a path planning algorithm to generate a cabinet opening track suitable for the door type, drive the mechanical arm to position the end effector to the cabinet door handle, simultaneously monitor the joint torque of the mechanical arm and visual feedback data in real time, dynamically adjust the cabinet opening strength, and complete the cabinet opening / closing action; The track navigation positioning module is integrated in the inspection robot body and is configured to position the robot space position in real time based on preset track coordinates and laser ranging data, plan a moving path and control the robot to move to the target equipment cabinet body; The access control recognition module is carried on the inspection robot body and is configured to collect image streams and distance data of the cabinet door area, extract handle positioning, door gap and lock state features through a pre-trained door state classification model, dynamically calculate the cabinet door state confidence and generate a cabinet opening instruction; The cabinet door opening and closing execution module is in communication connection with the execution mechanism control terminal and is configured to perform decryption verification based on the security operation permission code of the cabinet opening instruction, drive the mechanical arm to perform the cabinet opening / closing action, and dynamically adjust the operation strength through visual feedback and joint torque monitoring; The multi-modal acquisition module is integrated in the inspection robot body and comprises an infrared thermal imager, a camera and a wireless probe, and is configured to synchronously collect temperature distribution data, instrument and terminal state images and protection device operating parameters of equipment in the cabinet after the cabinet door is opened, and generate a timestamp-aligned equipment state data set through a multi-source heterogeneous data fusion algorithm; The defect analysis module is deployed on a local or cloud server and is configured to input the equipment state data set into a pre-trained defect recognition model, synchronously analyze temperature, image and parameter data through a multi-modal deep learning architecture, generate a comprehensive defect probability vector and an abnormal feature vector, and identify the equipment defect type, position and confidence; The warning generation module is in communication with the defect analysis module and is configured to dynamically determine the alarm level based on the abnormal feature vector, generate an alarm message containing equipment positioning, abnormal type, confidence and disposal measures, and push the alarm message to a specified terminal through a multi-channel communication protocol; The database management module is configured to store equipment historical baseline data and real-time state data, construct an equipment full life cycle health file through knowledge graph technology, associate spatial positioning, defect records and warning signals, and generate a periodic inspection report and an equipment health evaluation.

[0013] Based on the same inventive concept, the present application also provides an electronic device comprising a memory for storing a computer program and a processor for running the computer program to enable the electronic device to perform the substation inspection method described above.

[0014] Based on the same inventive concept, the present application also provides a computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the substation inspection method described above.

[0015] The technical solution of the present application has at least the following advantages and beneficial effects: The repetitive inspection task is completed by the inspection robot instead of manual work, greatly reducing manual intervention, effectively avoiding the missed inspection and misinspection problems in manual recording, significantly improving the inspection efficiency and the accuracy of data collection, and reducing the cost of manual operation and maintenance.

[0016] The technical problem that the existing track-type inspection robot cannot monitor the equipment in the cabinet is solved, the monitoring range is extended to the key equipment in the cabinet through the active cabinet opening action, the monitoring blind area in the traditional inspection method is eliminated, and full-scene and dead-angle-free intelligent monitoring of the substation equipment (including the equipment in the cabinet) is realized.

[0017] The full-process active operation capabilities such as cabinet opening, data collection, and cabinet closing are integrated, the limitation of the single function of the existing track-type inspection robot is broken through, the full-chain unmanned operation from equipment positioning, cabinet door operation to data collection is realized, and the automation degree and intelligent level of substation inspection are greatly improved.

[0018] Through the multi-modal acquisition module, the equipment state data is acquired in real time, combined with the defect recognition and abnormal feature analysis mechanism, the early warning signal containing accurate positioning, abnormal type and confidence can be quickly generated, the timely discovery and alarm of equipment abnormalities are realized, the traditional inspection data lag problem is solved, and timely and effective data support is provided for equipment fault prevention and operation and maintenance decision.

[0019] Through automatic updating of the equipment state database, generation of periodic inspection reports and operation state evaluation, an equipment state dynamic tracking system is built, which helps operation and maintenance personnel to comprehensively master the equipment health status, provides a scientific basis for equipment maintenance, life evaluation and resource allocation, and promotes the intelligent and refined transformation of the substation operation and maintenance mode. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the substation inspection method of the embodiment of the present application is shown in the figure; Figure 2 The structure diagram of the substation inspection system of the embodiment of the present application is shown in the figure; Figure 3Fig. 1 is a schematic diagram of the connection structure of the inspection robot and the track of the substation inspection system according to an embodiment of the present application. Fig. 1 is a schematic diagram of the connection structure of the inspection robot and the track of the substation inspection system according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] The specific embodiments are described below with reference to the accompanying drawings.

[0022] Reference Figure 1 A substation inspection method, comprising the following steps: Step 1, controlling the inspection robot to move along the preset track to the front of the target equipment cabinet, and triggering the access state recognition mechanism; Step 2, analyzing the cabinet door state through the access state recognition mechanism, if it is identified as a closed state, generating an open cabinet instruction and sending it to the execution mechanism of the inspection robot.

[0023] In some embodiments, the specific process of analyzing the cabinet door state through the access state recognition mechanism, if it is identified as a closed state, generating an open cabinet instruction and sending it to the execution mechanism of the inspection robot is as follows: Through the multi-modal acquisition module carried by the inspection robot, real-time image stream of the cabinet door area is collected, and distance sensor is used to synchronously obtain the cabinet door distance data; The image stream of the cabinet door area and the cabinet door distance data are input into a pre-trained door state classification model, the door state classification model extracts the cabinet door handle positioning feature, the door gap feature and the lock state feature based on the convolutional neural network and the point cloud fusion algorithm, and obtains the feature fusion result; According to the feature fusion result, the cabinet door state confidence is dynamically calculated, if the confidence is greater than or equal to the first threshold value, and it is identified as a closed state, an open cabinet instruction is generated; if the confidence is less than the first threshold value, the data is re-collected; The open cabinet instruction is associated with the spatial coordinates of the cabinet and the safety operation permission code, and is sent to the control terminal of the execution mechanism of the inspection robot through the encrypted communication protocol.

[0024] Step 3, in response to the open cabinet instruction, the open cabinet action is performed, and the multi-modal acquisition module carried by the inspection robot is used to real-time collect the state data of the target equipment in the cabinet.

[0025] In some embodiments, the specific process of responding to the open cabinet instruction, performing the open cabinet action, and real-time collecting the state data of the target equipment in the cabinet through the multi-modal acquisition module carried by the inspection robot is as follows: After receiving the open cabinet instruction, the control terminal of the execution mechanism performs decryption verification based on the safety operation permission code associated with the instruction; After verification, the path planning algorithm is called to generate an open cabinet trajectory adapted to the door type of the target cabinet, and the mechanical arm end effector is positioned to the cabinet door handle; By monitoring the joint torque of the robot arm and visual feedback data in real time, the opening force is dynamically adjusted until the cabinet door reaches the preset opening angle. After the cabinet door is opened, the synchronous acquisition protocol of the multi-modal acquisition module is triggered: the temperature distribution data of the equipment in the cabinet is acquired by the infrared thermal imager in a spatial grid scanning mode; the camera is controlled to adaptively focus based on the pre-set identification points of the equipment, and the state image of the instrument reading and the terminal is acquired; the running parameters of the protection device are read by the wireless probe; The temperature distribution data, state image and running parameters are input into the multi-source heterogeneous data fusion algorithm to generate a timestamp-aligned equipment state data set.

[0026] Step four, according to the real-time collected state data, identify the defects existing in the target equipment in the cabinet, and analyze the abnormal features in the real-time data.

[0027] In some embodiments, the specific process of identifying the defects existing in the target equipment in the cabinet according to the real-time collected state data and analyzing the abnormal features in the real-time data is: Based on the equipment state data set, input the pre-trained defect identification model, the defect identification model is based on a multi-modal deep learning architecture, and synchronously processes temperature distribution data, state image and running parameters; Through the visual analysis module in the defect identification model, extract the instrument reading deviation features, terminal loosening or corrosion defect features in the state image, and generate preliminary defect identification based on convolutional neural network classification algorithm; Analyze the real-time change trend of the running parameters, identify the parameter deviation or fluctuation abnormal mode, and apply the thermal map analysis algorithm to process the temperature distribution data to detect temperature hotspots exceeding the preset safety threshold; Fuse the preliminary defect identification, abnormal mode and temperature hotspots, and generate a comprehensive defect probability vector and an abnormal feature vector through a feature-level fusion algorithm; According to the comprehensive defect probability vector, if the probability value exceeds the second threshold, it is determined that the target equipment in the cabinet has defects; Based on the abnormal feature vector, analyze the abnormal type, location and confidence, and provide analysis basis for generating warning signals.

[0028] In some embodiments, the construction process of the defect identification model is: Collect multi-modal sample data sets from historical inspections of substations, including temperature distribution data, state image samples, running parameter time series labeled with defect labels, and corresponding defect type, location and severity label information. Structured data can be collected from historical inspection records of substations, including: Temperature distribution data: two-dimensional thermal map (matrix form) collected by an infrared thermal imager; State image samples: high-resolution images of instruments, wiring terminals, etc. captured by RGB cameras. Running parameter time series: time series data of voltage, current, frequency, etc. parameters of the protection device (sampling frequency ≥ 1 Hz).

[0029] Add label information to all samples: Defect label: binary label (normal / defect) or multi-class label (e.g. "wiring terminal corrosion", "instrument reading deviation").

[0030] Position annotation: bounding box coordinates of abnormal areas (e.g. ).

[0031] Severity: discrete levels (e.g. 1 to 5 levels).

[0032] Preprocess the multi-modal sample dataset to generate a standardized multi-modal training dataset; preprocessing includes: Image normalization: adjust image size to a uniform resolution and apply histogram equalization to enhance contrast; Temperature data standardization: Min-Max scaling of the heat map to [0, 1] and removal of the ambient temperature baseline; Time series data alignment: resample the running parameter sequence (fixed time window) and fill in missing values (e.g. linear interpolation); Data augmentation: apply random rotation, translation, and Gaussian noise to images.

[0033] Construct a multi-modal deep learning architecture, which includes a visual feature extraction branch, a parameter time series analysis branch, and a temperature feature analysis branch; the visual feature extraction branch uses an improved convolutional neural network to extract instrument reading features and wiring terminal defect features from state images through residual modules; an improved ResNet-50 can be used as the backbone network, removing the last fully connected layer and adding residual modules (Residual Blocks) to input state images and output feature vectors .

[0034] The parameter time series analysis branch uses a bidirectional recurrent neural network to extract trend features and fluctuation patterns from running parameter time series; a bidirectional LSTM (Bi-LSTM) network can be used, inputting running parameter sequences and outputting feature vectors .

[0035] The temperature feature analysis branch combines heat map analysis algorithm and convolutional neural network to extract hot spot area features and temperature gradient features in temperature distribution data. Heat map spectral analysis is applied to extract frequency domain features: performing 2D discrete Fourier transform (DFT) on the temperature distribution matrix; inputting the frequency spectrum into a lightweight CNN (for example: 3 layers of convolution + pooling), and outputting a feature vector The expression of DFT calculation is shown in the following formula (1):

[0036] Wherein, is the temperature value of the temperature distribution matrix at the spatial position , which is a two-dimensional real matrix (heat map) representing the surface temperature distribution of the equipment in the cabinet, is the row index (vertical direction coordinate) of the matrix, with a value range of , is the column index (horizontal direction coordinate) of the matrix, with a value range of ; represents the height (number of rows) of the temperature distribution matrix, with a unit of pixels; represents the width (number of columns) of the temperature distribution matrix, with a unit of pixels; is the horizontal frequency index in the frequency domain, with a value range of , representing the horizontal spatial frequency component; is the vertical frequency index in the frequency domain, with a value range of , representing the vertical spatial frequency component; here is the imaginary unit; represents the base of natural logarithm; represents the ratio of a circle; is the output result, representing the complex value (including amplitude and phase information) at the frequency domain coordinate , which is the representation of temperature distribution data in the frequency domain.

[0037] Take the logarithmic amplitude spectrum as the CNN input, as shown in the following formula (2):

[0038] Wherein, represents the logarithmic amplitude spectrum matrix; The output of the CNN is shown in the following formula (3):

[0039] Wherein, represents the convolutional neural network model, which is used to process the input matrix , i.e. the input matrix Performing efficient feature extraction; The weight parameter set of the CNN, including convolution kernel weights, bias terms and other parameters, is learned through optimization algorithms (such as gradient descent) during model training.

[0040] A cross-modal feature fusion layer is set up, and an attention mechanism is used to dynamically weight the feature vectors output by the visual feature extraction branch, the parameter time series analysis branch and the temperature feature analysis branch. A multi-modal fusion feature is generated through a feature-level fusion algorithm. The three modal feature vectors are dynamically weighted and fused using the attention mechanism: 、 、 The output fusion feature is as shown in the following formula (4):

[0041] wherein, represents the transformed feature vector, which is the result after linear transformation and activation function processing; represents the rectified linear unit activation function; represents the modal-specific weight matrix, which is used for linear transformation of the input feature ; is the input feature vector, representing the original output feature of the th modal branch; represents the modal-specific bias vector, which is used in conjunction with to add bias in the linear transformation and adjust the feature distribution; is the modal index, representing the current processing modal type, and the value range is , wherein, corresponds to the visual feature extraction branch, corresponds to the parameter time series analysis branch, corresponds to the temperature feature analysis branch; is the attention score, representing the relative importance or relevance of the modal , and the higher the score, the greater the weight of the modal in the fusion; is the transpose of the attention score weight vector; represents the hyperbolic tangent activation function; represents the shared attention weight matrix, which is used for attention calculation; represents the shared attention bias vector, which is used in conjunction with ; represents the attention weight, which is the normalized value through the softmax function; represents the natural exponential function; represents the sum of the exponentialized attention scores of all modalities; Based on the multi-modal fusion feature, a defect classification and anomaly positioning sub-network is constructed, the sub-network adopts a full connection layer and a softmax classifier, and outputs a defect type probability distribution, an abnormal position coordinate and a confidence score. The full connection layer is shown in the following formula (5):

[0042] wherein, represents an output vector (hidden state) of the full connection layer; represents a weight matrix of the full connection layer; represents a bias vector of the full connection layer; The defect classification output is shown in the following formula (6):

[0043] wherein, is an output vector, representing a probability distribution of the defect type; is an activation function, used to convert the input vector into a probability distribution, ensuring that all elements sum to 1 and the value is in the range of [0, 1]; represents a weight matrix of the classification layer; represents a bias vector of the classification layer; The anomaly positioning output is shown in the following formula (7):

[0044] wherein, is an output vector, representing a predicted abnormal position coordinate; represents a Sigmoid activation function, used to compress the output to the range of [0, 1]; represents a weight matrix of the positioning layer; represents a bias vector of the positioning layer; The confidence output is shown in the following formula (8):

[0045] wherein, is an output scalar, representing a confidence score; represents a transpose of a weight vector; represents a bias scalar.

[0046] The multi-modal training data set is divided into a training set and a validation set. Through the training set, the model is iteratively trained using a joint loss function, and the weights of each branch network are dynamically adjusted. The joint loss function is shown in the following formula (9):

[0047] wherein, is a total loss function, i.e. a joint loss, representing the overall optimization goal of model training; The weighting coefficients represent the classification loss. The weighting coefficients representing the localization loss; The weighting coefficients representing the confidence loss; Indicates classification loss; Indicates positioning loss; Indicates confidence loss; The classification loss is shown in equation (10):

[0048] in, Indicates the total number of defect categories; Indicates a category index; The first one representing the true label Each element is a binary value (0 or 1). In one-hot encoding, the true class is 1, and other classes are 0. The model predicts the first The probability (scalar) of each category, with values ​​ranging from [0,1]; The positioning loss is shown in equation (11):

[0049] in, The model predicts the first The coordinate value; representing the actual position. Each coordinate value; This represents the coordinate dimension index, with a value range of 100. The four coordinates of the bounding box corresponding to the abnormal location (e.g.: ); The confidence loss is shown in equation (12):

[0050] in, The true confidence level label is a binary value (0 or 1), where 0 represents "normal" (no defects) and 1 represents "defect" (defined by the defect label). The weight update is shown in equation (13):

[0051] in, This indicates that the updated model parameters are at step [step]. The value; This indicates the model parameters (weights and biases) at each training iteration step. The value (vector or tensor); It represents the learning rate (a scalar) and controls the size of the update step size; is the bias-corrected first moment estimate (momentum term, vector) representing the exponentially moving average of the gradient; is the bias-corrected second moment estimate (adaptive learning rate term, vector) representing the exponentially moving average of the squared gradient; is the numerical stability constant (hyperparameter, scalar) preventing division by zero; is the first moment estimate (uncorrected, vector) at step is computed as: where is the gradient of the loss function at step ; is the second moment estimate (uncorrected, vector) at step is computed as: ; is the first moment decay rate used to control the forgetting speed of the history gradient; is the second moment decay rate used to control the forgetting speed of the history squared gradient.

[0052] The trained model is evaluated for performance by the validation set, and the defect recognition accuracy, abnormal positioning accuracy, and confidence threshold stability are calculated. If the evaluation indicators do not meet the preset standards, the network depth, fusion strategy, or loss function weight is adjusted and retrained; The optimized model is deployed to the inference engine, and online fine-tuning is performed on incremental samples of actual inspection scenes to generate the final defect recognition model. The fine-tuning loss expression is shown in the following formula (14):

[0053] where is the total loss of fine-tuning, used for online incremental learning after model deployment; is the loss on new sample data; is the regularization coefficient used to control the "memory strength" of the old model parameters. The larger the value, the smaller the parameter change during fine-tuning (to prevent catastrophic forgetting); is the model parameter in the current fine-tuning; is the parameter of the model before fine-tuning (old model), i.e., the parameter before deployment to the inference engine.

[0054] Step five, if an abnormal feature is recognized, a warning signal containing device positioning, abnormal type, and confidence is generated, and an alarm notification is triggered.

[0055] In some embodiments, the specific process of triggering the alarm notification is: Based on the abnormal type and confidence contained in the warning signal, the alarm level is dynamically determined; According to the alarm level and the preconfigured push rule, an alarm message containing device positioning, abnormal type, confidence and suggested treatment measures is generated; Through a multi-channel communication protocol, the alarm message is synchronously pushed to a mobile terminal of an operation personnel, an alarm screen of a centralized control center and a device management system; The historical defect record in the device state database is associated, and the historical fault statistics and processing plan link of the same type device are attached in the alarm message.

[0056] After the data collection is completed, a closing cabinet action is performed, the device state database is updated, and a periodic inspection report and a device operation state evaluation are generated.

[0057] In some embodiments, the specific process of updating the device state database and generating the periodic inspection report and the device operation state evaluation is as follows: The real-time collected state data and the historical baseline data stored in the device state database are time-aligned, and the long-term change trend characteristics and the short-term fluctuation characteristics of the device state are extracted; Based on the change trend characteristics and the fluctuation characteristics, the parameters of the defect identification model are dynamically iteratively updated, and the recognition threshold and the feature weight of the defect identification model for the progressive defects of the device are optimized; Through knowledge graph technology, the device spatial positioning information, the historical defect record, the real-time abnormal feature and the early warning signal are associated and mapped, a whole life cycle health file containing device health indicators and defect evolution paths is constructed and updated to the device state database; When generating the periodic inspection report, the multi-source data in the device health file are structurally analyzed, the key indicators, abnormal statistics and trend prediction results are extracted, and the device residual life prediction value and the maintenance priority ranking are output.

[0058] Based on the same inventive concept, corresponding to any of the above embodiments, with reference to Figure 2 、 Figure 3 , the present application provides a substation inspection system for implementing the foregoing substation inspection method, comprising: The inspection robot comprises an inspection robot body and an execution mechanism, the inspection robot body is in sliding connection with the track, and the execution mechanism is arranged on the inspection robot body; the execution mechanism comprises a mechanical arm 2, an end effector 3 and a control terminal; after receiving the cabinet opening instruction, the control terminal is configured to perform security operation permission coding decryption verification, call a path planning algorithm to generate an opening cabinet track adapted to the door type, drive the mechanical arm 2 to position the end effector 3 to the cabinet door handle; and simultaneously, the joint torque of the mechanical arm 2 and the visual feedback data are monitored in real time, the opening cabinet force is dynamically adjusted, and the opening cabinet / closing cabinet action is completed. An orbit navigation positioning module integrated in the inspection robot main body is configured to position the robot spatial position in real time based on preset orbit coordinates and laser ranging data, plan a moving path and control the robot to move to a target equipment cabinet; An access control recognition module carried on the inspection robot main body is configured to collect image streams and distance data of a cabinet door area, extract handle positioning, door gap and lock state features through a pre-trained door state classification model, dynamically calculate a cabinet door state confidence and generate an open cabinet instruction; A cabinet door opening and closing execution module is in communication connection with an execution mechanism control terminal and is configured to decrypt and verify a security operation permission code based on the open cabinet instruction, drive a mechanical arm to perform an open cabinet / closed cabinet action, and dynamically adjust an operation force through visual feedback and joint torque monitoring; A multi-modal acquisition module integrated in the inspection robot main body includes an infrared thermal imager, a camera and a wireless probe and is configured to synchronously acquire temperature distribution data, instrument and terminal state images and protection device operating parameters of equipment in a cabinet after the cabinet door is opened, and generate a timestamp-aligned equipment state data set through a multi-source heterogeneous data fusion algorithm; A defect analysis module deployed on a local or cloud server is configured to input the equipment state data set into a pre-trained defect recognition model, synchronously analyze temperature, image and parameter data through a multi-modal deep learning architecture, generate a comprehensive defect probability vector and an abnormal feature vector, and identify a defect type, position and confidence of the equipment; An early warning generation module in communication with the defect analysis module is configured to dynamically determine an alarm level based on the abnormal feature vector, generate an alarm message containing equipment positioning, abnormal type, confidence and disposal measures, and push the alarm message to a specified terminal through a multi-channel communication protocol; A database management module is configured to store equipment historical baseline data and real-time state data, construct an equipment full life cycle health file through knowledge graph technology, associate spatial positioning, defect records and early warning signals, and generate a periodic inspection report and an equipment health evaluation.

[0059] Based on the same inventive concept, the present application provides an electronic device, including a memory and a processor, the memory is used for storing a computer program, and the processor runs the computer program to make the electronic device execute the substation inspection method of the embodiments.

[0060] Optionally, the electronic device can be a server.

[0061] In addition, the present embodiment also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the substation inspection method of the embodiments.

[0062] It is appreciated that the processor in the embodiments of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0063] The method steps in the embodiments of the present application can be realized by hardware or by the processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically-erasable programmable read-only memory (EEPROM), a register, a hard disk, a mobile hard disk, a CD-ROM or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor, so that the processor can read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0064] In the embodiments described above, all or some of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or some of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded into and executed by a computer, all or some of the procedures or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in or transmitted from a storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. including one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

Claims

1. A method of substation inspection, characterized by, The method comprises the following steps: controlling the inspection robot to move along a preset track to the front of a target equipment cabinet and triggering a door state recognition mechanism; analyzing the door state through the door state recognition mechanism, and if the door state is identified as closed, generating an opening cabinet instruction and sending it to the actuator of the inspection robot; in response to the opening cabinet instruction, performing an opening cabinet action and collecting real-time state data of the target equipment in the cabinet through the multi-modal collection module carried by the inspection robot; identifying defects existing in the target equipment in the cabinet according to the real-time collected state data, and analyzing abnormal features in the real-time data; if the abnormal features are identified, generating a warning signal containing the location of the equipment, the type of the abnormality and the confidence, and triggering an alarm notification; after completing data collection, performing a closing cabinet action, updating the equipment state database, and generating a periodic inspection report and an equipment operation state evaluation.

2. The substation patrol method of claim 1, wherein, The specific process of analyzing the door state through the door state recognition mechanism, and if the door state is identified as closed, generating an opening cabinet instruction and sending it to the actuator of the inspection robot is as follows: collecting image streams of the door area in real time through the multi-modal collection module carried by the inspection robot, and synchronously acquiring door distance data through a distance sensor; inputting the image streams of the door area and the door distance data into a pre-trained door state classification model, which extracts door handle positioning features, door gap features and lock state features based on a convolutional neural network and a point cloud fusion algorithm, and obtains a feature fusion result; dynamically calculating the door state confidence according to the feature fusion result, if the confidence is greater than or equal to a first threshold value, and the door state is identified as closed, an opening cabinet instruction is generated; if the confidence is less than the first threshold value, data is re-collected; associating the opening cabinet instruction with the spatial coordinates of the cabinet and the safety operation permission code, and sending it to the control terminal of the actuator of the inspection robot through an encrypted communication protocol.

3. The substation patrol method of claim 2, wherein, The specific process of responding to the opening cabinet instruction, performing an opening cabinet action, and collecting real-time state data of the target equipment in the cabinet through the multi-modal collection module carried by the inspection robot is as follows: after receiving the opening cabinet instruction, the control terminal of the actuator decrypts and verifies based on the safety operation permission code associated with the instruction; after verification, a path planning algorithm is called to generate an opening cabinet trajectory adapted to the door type of the target cabinet, and the end effector of the mechanical arm is positioned to the door handle; by monitoring the joint torque of the mechanical arm and the visual feedback data in real time, the opening cabinet force is dynamically adjusted until the door reaches a preset opening angle; after the door is opened, the synchronous collection protocol of the multi-modal collection module is triggered: the temperature distribution data of the equipment in the cabinet is collected in a spatial grid scanning mode through an infrared thermal imager; controlling the camera to focus adaptively based on the preset identification points of the equipment, collecting the state images of the instrument readings and the terminal; reading the operating parameters of the protection device using a wireless probe; inputting the temperature distribution data, the state images and the operating parameters into a multi-source heterogeneous data fusion algorithm to generate a timestamp-aligned equipment state data set.

4. The substation patrol method of claim 3, wherein, The specific process of identifying defects existing in the target equipment in the cabinet according to the real-time collected state data, and analyzing abnormal features in the real-time data is as follows: Based on the equipment state dataset, a pre-trained defect recognition model is input, the defect recognition model is based on a multi-modal deep learning architecture, and temperature distribution data, state images and operating parameters are synchronously processed; Through a visual analysis module in the defect recognition model, instrument reading deviation features, connection terminal loosening or corrosion defect features in the state images are extracted, and a preliminary defect identification is generated based on a convolutional neural network classification algorithm; Real-time change trends of the operating parameters are analyzed, parameter deviation or fluctuation anomaly patterns are identified, and a thermal map analysis algorithm is applied to process the temperature distribution data to detect temperature hotspots exceeding a preset safety threshold; The preliminary defect identification, anomaly patterns and temperature hotspots are fused, a comprehensive defect probability vector and an anomaly feature vector are generated through a feature-level fusion algorithm; According to the comprehensive defect probability vector, if the probability value exceeds a second threshold, it is determined that the target equipment in the cabinet has a defect; Based on the anomaly feature vector, the anomaly type, location and confidence are analyzed to provide an analysis basis for generating a warning signal.

5. The substation patrol method of claim 4, wherein, The construction process of the defect recognition model is as follows: A multi-modal sample dataset of historical inspections of a substation is collected, the sample dataset includes temperature distribution data, state image samples, operating parameter time series, and corresponding defect type, location and severity annotation information; The multi-modal sample dataset is preprocessed to generate a standardized multi-modal training dataset; A multi-modal deep learning architecture is constructed, the multi-modal deep learning architecture includes a visual feature extraction branch, a parameter time series analysis branch and a temperature feature analysis branch; wherein the visual feature extraction branch adopts an improved convolutional neural network, instrument reading features and connection terminal defect features in the state images are extracted through a residual module; the parameter time series analysis branch adopts a bidirectional recurrent neural network to extract trend features and fluctuation patterns from the operating parameter time series; the temperature feature analysis branch combines a thermal map analysis algorithm and a convolutional neural network to extract hotspot area features and temperature gradient features from the temperature distribution data; A cross-modal feature fusion layer is set, an attention mechanism is used to dynamically allocate weights to the feature vectors output by the visual feature extraction branch, the parameter time series analysis branch and the temperature feature analysis branch, and a multi-modal fusion feature is generated through a feature-level fusion algorithm; Based on the multi-modal fusion feature, a defect classification and anomaly positioning subnetwork is constructed, the subnetwork adopts a fully connected layer and a softmax classifier to output a defect type probability distribution, an abnormal location coordinate and a confidence score; The multi-modal training dataset is divided into a training set and a validation set, the model is iteratively trained using a joint loss function through the training set, and the weights of the branch networks are dynamically adjusted; The performance of the trained model is evaluated through the validation set, the defect recognition accuracy, the anomaly positioning precision and the confidence threshold stability are calculated, if the evaluation indexes do not reach the preset standard, the network depth, the fusion strategy or the loss function weight are adjusted and retrained; The optimized model is deployed to an inference engine, online fine-tuning is performed through incremental samples of actual inspection scenes, and a final defect recognition model is generated.

6. The substation patrol method of claim 4, wherein, The updating device state database, and generating periodic inspection report and device running state evaluation specific process is: The real-time collected state data and the historical baseline data stored in the device state database are time series aligned, the long-term change trend characteristics and short-term fluctuation characteristics of the device state are extracted; Based on the change trend characteristics and fluctuation characteristics, the parameters of the defect recognition model are dynamically iterated and updated, and the recognition threshold and feature weight of the defect recognition model to the progressive defect of the device are optimized; Through the knowledge graph technology, the device space positioning information, historical defect record, real-time abnormal feature and early warning signal are correlated and mapped, the full life cycle health file containing device health index and defect evolution path is constructed and updated to the device state database; When generating the periodic inspection report, the multi-source data in the device health file is structured and analyzed, the key indicators, abnormal statistics and trend prediction results are extracted, and the device remaining life prediction value and maintenance priority ranking are output.

7. The substation patrol method of claim 1, wherein, The specific process of triggering the alarm notification is: Based on the abnormal type and confidence contained in the early warning signal, the alarm level is dynamically determined; According to the alarm level and the preconfigured push rule, the alarm message containing the device positioning, abnormal type, confidence and suggested treatment measures is generated; Through multi-channel communication protocol, the alarm message is pushed to the mobile terminal of the operator, the alarm screen of the centralized control center and the device management system; The historical defect record in the device state database is associated, and the historical fault statistics and processing plan link of the same type device are attached to the alarm message.

8. A substation inspection system for implementing the substation inspection method according to any one of claims 1 to 7, characterized by It includes: The inspection robot includes an inspection robot body and an execution mechanism. The inspection robot body is in sliding connection with a track, and the execution mechanism is arranged on the inspection robot body. The execution mechanism includes a mechanical arm, an end effector and a control terminal. After receiving the cabinet opening instruction, the control terminal is configured to perform security operation permission code decryption verification, call a path planning algorithm to generate a cabinet opening trajectory adapted to the door type, drive the mechanical arm to position the end effector to the cabinet door handle, and simultaneously monitor the joint torque of the mechanical arm and the visual feedback data in real time, dynamically adjust the cabinet opening force, and complete the cabinet opening / closing action. The track navigation positioning module is integrated in the inspection robot body and is configured to position the robot space in real time based on the preset track coordinates and laser ranging data, plan a moving path and control the robot to move to the target device cabinet. The access control recognition module is carried on the inspection robot body and is configured to collect image flow and distance data of the cabinet door area, extract handle positioning, door gap and lock state features through a pre-trained door state classification model, dynamically calculate the cabinet door state confidence and generate a cabinet opening instruction. The cabinet door opening and closing execution module is in communication connection with the control terminal of the execution mechanism and is configured to perform decryption verification based on the security operation permission code of the cabinet opening instruction, drive the mechanical arm to perform cabinet opening / closing action, and dynamically adjust the operation force through visual feedback and joint torque monitoring. A multi-modal acquisition module integrated in the main body of the inspection robot includes an infrared thermal imager, a camera, and a wireless probe, and is configured to synchronously acquire temperature distribution data of equipment in the cabinet, images of instrument and terminal states, and operation parameters of protection devices after the cabinet door is opened, and generate a timestamp-aligned equipment state data set through a multi-source heterogeneous data fusion algorithm; A defect analysis module deployed on a local or cloud server is configured to input the equipment state data set into a pre-trained defect recognition model, synchronously analyze temperature, image, and parameter data through a multi-modal deep learning architecture, generate a comprehensive defect probability vector and an abnormal feature vector, and identify the type, location, and confidence of the equipment defect; An early warning generation module communicates with the defect analysis module and is configured to dynamically determine the alarm level based on the abnormal feature vector, generate an alarm message containing the equipment location, abnormal type, confidence, and disposal measures, and push the alarm message to a specified terminal through a multi-channel communication protocol; A database management module is configured to store historical baseline data and real-time state data of the equipment, construct a full life cycle health record of the equipment through knowledge graph technology, associate spatial positioning, defect records, and early warning signals, and generate periodic inspection reports and equipment health assessments.

9. An electronic device, comprising: The electronic device includes a memory and a processor. The memory is configured to store a computer program. The processor is configured to execute the computer program to cause the electronic device to perform the substation inspection method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is configured to cause the processor to perform the substation inspection method of any one of claims 1-7.

Images