Lightweight CNN-based secondary safety measure ticket execution step checking method and system

Abstract

The invention discloses a secondary safety measure ticket execution step checking method and system based on a lightweight CNN, and belongs to the technical field of power system safety operation and maintenance, and the method comprises the steps: constructing a power plant booster station secondary equipment digital model and a safety measure logic library, constructing a lightweight CNN model, including a feature separation module, a feature processing module, a feature fusion module and a channel attention module; an equipment state image in a safety measure execution process is collected through a mobile terminal, a lightweight CNN model is used for identification, an image identification result and data acquired from a station level network and a process level network are combined, a safety measure execution state is comprehensively judged, an actual state of a safety measure object obtained through identification is compared with an expected state in a safety measure ticket, and a safety measure execution result is obtained. According to the method, the checking result is fed back in real time, automatic, real-time and intelligent checking of the safety measure execution step is realized, the checking efficiency and accuracy are improved, the human error risk is reduced, and the safe operation level of a power plant is improved.

Description

Technical Field

[0001] This invention belongs to the field of power plant safety operation and maintenance technology, specifically relating to a method and system for verifying the execution steps of secondary safety measure tickets (hereinafter referred to as "secondary safety measure tickets") based on lightweight convolutional neural networks (CNN). This invention combines deep learning image recognition technology with the management specifications for safety measures of secondary equipment in power plant substations to automate the verification of the execution steps of secondary safety measure tickets, thereby improving the safety operation level of power plants. Background Technology

[0002] With the rapid development of society and the economy, the safety requirements of power plants are increasing. Traditional secondary safety measure ticket management has the following problems: the quality of manually prepared safety measure tickets is greatly affected by subjective human factors; the implementation of safety measures requires manual verification, and unreasonable or incomplete implementation of safety measures will pose a challenge to the safe operation of the power plant. In existing technologies, the implementation of secondary safety measure tickets mainly relies on manual supervision, which has the following shortcomings: Manual verification is inefficient: During the implementation of safety measures, guardians need to check the implementation status of each measure one by one, which is time-consuming and prone to errors.

[0003] Subjective factors have a significant impact: the quality of safety measures preparation and implementation is affected by subjective factors such as personnel experience and sense of responsibility, and the degree of standardization is insufficient.

[0004] Insufficient real-time capability: The existing online monitoring devices are located at a distance from the safety control cabinet, making it impossible to notify operators of relevant statuses in real time. Many signals and statuses still require manual verification.

[0005] Insufficient technology integration: Although there are existing methods for verifying secondary safety measures at power plant substations based on digital twins and online verification systems for visualizing secondary safety measure operation tickets, these systems still have limitations in terms of mobile terminal deployment and real-time image recognition.

[0006] In recent years, lightweight CNN technology has made significant progress. Models such as SqueezeNet, ShuffleNet, and MobileNet have achieved efficient operation under limited resources by reducing the number of parameters and computational costs. These technologies have made it possible to deploy safety ticket execution verification systems on mobile devices. However, existing technologies have not effectively combined lightweight CNNs with secondary safety ticket execution verification, and there is a lack of a solution that can verify safety execution steps in real time on resource-constrained mobile devices. Summary of the Invention

[0007] This invention aims to solve the problems of manual labor, low efficiency, and error susceptibility in the verification of secondary safety measures tickets in the prior art. It provides a method and system for verifying the execution steps of secondary safety measures tickets based on lightweight CNN, so as to realize the automated, real-time, and intelligent verification of the safety measures execution steps and improve the safety operation level of power plants.

[0008] Technical solution To achieve the above objectives, this invention provides a method for verifying the execution steps of a secondary security ticket based on a lightweight CNN, comprising the following steps: S1. Digital Modeling of Safety Tickets S1.1 constructs a digital model of the secondary equipment of the power plant's booster station based on SCD, ICD, and SPCD files.

[0009] S1.2 Analyze the GOOSE / SV receiving and subscription relationships, soft pressure plate opening and closing relationships, and physical connection relationships between secondary devices, and establish a security logic library.

[0010] S1.3 structures the content of the safety measure ticket to form a digital safety measure ticket template that includes the object of the safety measure, the expected state, and the execution sequence.

[0011] S2. Lightweight CNN Model Construction S2.1 uses depthwise separable convolution (DWC) and pointwise convolution (PWC) to replace standard convolution operations, reducing the number of model parameters.

[0012] S2.2 performs batch standardization and sorting reconstruction on the input feature maps, dividing the feature maps into high-variance feature groups and low-variance feature groups along the channel dimension.

[0013] S2.3 performs target edge feature enhancement processing on high variance feature groups by performing grouped convolution (GWC) and point convolution (PWC) operations; and performs depthwise separable convolution (DWC) and point convolution (PWC) operations on low variance feature groups.

[0014] S2.4 performs grouped complementary fusion processing on the two sets of feature maps and uses a channel attention mechanism to dynamically adjust the channel weights.

[0015] S3. Image Acquisition and Recognition in Anco execution S3.1 uses a mobile terminal camera module to capture images of the equipment status during the implementation of safety measures.

[0016] S3.2 uses a lightweight CNN model to perform real-time identification of the acquired images and extract the safety objects and their status information.

[0017] S3.3 Combines the image recognition results with the MMS, GOOSE, and SV messages obtained from the station control layer network and the process layer network to comprehensively determine the execution status of the safety measures; S4. Verification of Safety Measures Implementation Steps S4.1 compares the actual state of the identified safety measures object with the expected state in the safety measures ticket.

[0018] S4.2 If the comparison is consistent, the verification will be displayed on the mobile terminal interface; if they are inconsistent, the corresponding safety measure ticket will be marked in red and an alarm will be issued through the audio output module.

[0019] S4.3 generates verification reports in real time, recording any abnormal situations and handling suggestions during the implementation of safety measures.

[0020] S5. Safety Measures Ticket Execution Process Management S5.1 displays the verification results of each item on the safety ticket in real time on the mobile terminal interface, and reads the text of the safety ticket items aloud through the audio output module.

[0021] S5.2 Electronic signature confirmation of executed steps ensures traceability of the execution process.

[0022] S5.3 uploads the execution results and verification reports to the management system, realizing full lifecycle management of safety measures tickets.

[0023] Furthermore, the specific execution steps of step S2.1 described above are as follows: Combining depthwise separable convolution and pointwise convolution can significantly reduce the number of parameters and computational complexity of standard convolution operations. The formula for calculating the number of parameters in standard convolution is: in: The kernel size; Input the number of channels; This represents the number of output channels.

[0024] Depthwise separable convolution decomposes standard convolution into two steps: Depthwise Convolution (DWC): A separate convolution kernel is applied to each input channel, with the following number of parameters: Pointwise convolution (PWC): Uses a 1×1 convolution to perform channel fusion on the output of DWC, with the following parameters: Therefore, the total number of parameters is: contrast: when When the number of parameters is large, the number of parameters can be reduced to that of standard convolution. This significantly reduces model complexity by a fraction of the order of magnitude.

[0025] Furthermore, the specific execution steps of step S2.2 described above are as follows: S2.2.1 Batch Standardization: Standardizing the input feature map The standardization process is performed, and the formula is as follows: in: Let be the mean of the c-th channel; Let be the variance of the c-th channel; ϵ is a small constant to prevent division by zero.

[0026] S2.2.2 Channel Variance Calculation: Calculate the variance of each channel. : S2.2.3 Channel sorting and grouping: based on variance Sort the channels in descending order and divide the first α⋅C channels into high variance feature groups. The remaining channels are divided into low-variance feature groups. , where α is the grouping ratio (e.g., 0.5).

[0027] Furthermore, the specific execution steps of step S2.3 described above are as follows: S2.3.1. High variance feature group processing (edge ​​feature enhancement + GWC + PWC); Edge feature enhancement: Edge features are extracted using Sobel or Canny operators to enhance spatial details in high-variance groups. For example, the Sobel operator's convolution kernel is: The edge strength is: Connect the edge feature E with Fusion: Where λ is the fusion weight.

[0028] Grouped convolution (GWC): Divide into G groups, and perform convolution independently on each group: in is the convolution kernel for the g-th group.

[0029] Pointwise convolution (PWC): Performs channel fusion on the output of GWC. in This is for channel splicing operations.

[0030] S2.3.2 Low-variance feature group processing (DWC+PWC) Depthwise Separable Convolution (DWC): Point convolution (PwC): Furthermore, the specific execution steps of step S2.4 described above are as follows: S2.4.1 Feature stitching: Grouping high variance groups and low variance group Stitching along the channel dimension: S2.4.2 Channel Attention Mechanism: Global Average Pooling (GAP): Calculates a global average for each channel. Fully connected layer activation: Channel weights are generated through a two-layer fully connected network. in: , This is the weight matrix; δ is the ReLU activation function; σ is the Sigmoid function; r represents the dimensionality reduction ratio.

[0031] Weight recalibration: Multiply the channel weights s by the concatenated features: Where ⊙ represents channel-by-channel multiplication.

[0032] S2.4.3 Output Feature Map: The final output is the feature map after dynamic adjustment by channel attention. , for use in subsequent tasks.

[0033] Furthermore, the specific execution steps of step S3.2 described above are as follows: S3.2.1 Model Architecture Design: Depthwise separable convolution: Decomposes standard convolution into depthwise convolution and pointwise convolution, reducing the computational cost to one-third that of standard convolution. (k is the kernel size).

[0034] Depthwise convolution: on the input feature map Each channel is independently convolutional with a k×k kernel, and the output is... : Pointwise convolution: Uses a 1×1 convolution to fuse channel information and output... : Channel attention mechanism: Dynamically adjust channel weights through the SE module to enhance key features (such as on / off status indicators): Compression: Global average pooling generates channel descriptors : Incentives: Weights are generated through a two-layer fully connected network. : Recalibration: Weights multiplied by feature maps: .

[0035] S3.2.2 Real-time recognition process: Input preprocessing: The image is scaled to the model input size (e.g., 224×224) and normalized to the [0,1] interval.

[0036] Feature extraction: Lightweight CNN extracts device state features and outputs class probability distribution. N represents the number of categories of safety measures objects.

[0037] State resolution: Locate the safety measure object and resolve its state through non-maximum suppression and threshold filtering; This invention also provides a secondary security ticket execution step verification system based on lightweight CNN, comprising: The data acquisition device connects to the station control layer network and the process layer network to acquire safety measure verification data such as MMS messages, GOOSE messages, and SV messages.

[0038] Data is forwarded to the security verification device via a secure data communication bastion host to ensure secure data transmission.

[0039] The safety measure verification device stores the safety measure logic library, which includes information such as the name of each power plant's step-up substation, SCD files, and safety measure tickets for each interval.

[0040] Analyze the data acquired by the data acquisition device to obtain the safety measures object and its status.

[0041] Deploy a lightweight CNN model to recognize images of objects requiring safety measures.

[0042] The combined data collection and image recognition results are compared and verified with the security ticket.

[0043] Ancolic verification terminal, human-computer interaction module: used to edit ancolic tasks and display verification results.

[0044] Camera module: Takes photos of the subjects.

[0045] Wireless communication module: Communicates with the safety verification device.

[0046] Audio output module: issues alarm signals and reads out the safety ticket text.

[0047] The lightweight CNN model module and feature separation module perform batch standardization and sorting reconstruction on the input feature maps, dividing them into high-variance and low-variance feature groups.

[0048] Feature processing module: performs edge enhancement and grouped convolution on high variance feature groups, and depthwise separable convolution on low variance feature groups.

[0049] Feature fusion module: performs complementary fusion of two sets of feature maps.

[0050] Channel attention module: dynamically adjusts channel weights to obtain the final output feature map; Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: Improve verification efficiency: Automated verification of safety measures execution steps is achieved through a lightweight CNN model, reducing manual intervention and improving verification efficiency.

[0051] Enhance verification accuracy: Combine image recognition with real-time data collection to verify the implementation status of safety measures from multiple dimensions and reduce the risk of human error.

[0052] Enables mobile deployment: Lightweight CNN models have fewer parameters and lower computational cost, enabling them to run efficiently on resource-constrained mobile devices.

[0053] Improved real-time performance: Images and data are collected in real time via mobile terminals, and verification results are fed back instantly, solving the problem of distance limitations in traditional online monitoring.

[0054] Standardize safety measure management: Achieve digital management of safety measure tickets throughout their entire lifecycle, including preparation, execution, verification, and archiving, thereby improving management standardization.

[0055] Enhanced traceability: Electronic signature confirmation and verification report generation ensure that the implementation of safety measures is traceable and auditable. Attached Figure Description

[0056] Figure 1 System architecture diagram.

[0057] Figure 2 A schematic diagram of a lightweight CNN model structure.

[0058] Figure 3 Flowchart for the verification process of Ancuo. Detailed Implementation

[0059] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0060] Example 1: System Architecture like Figure 1 As shown, the secondary security measure ticket execution step verification system based on lightweight CNN in this embodiment includes a data acquisition device, a secure data communication bastion host, a security measure verification device, and multiple security measure verification terminals.

[0061] 1. The data acquisition device connects to the station control layer network and the process layer network to acquire data related to safety measure verification, including: Obtain information about the protection device's soft pressure plate, control word, hard pressure plate, and fiber optic link breakage from the MMS message.

[0062] Obtain information from the GOOSE message regarding the location of the smart terminal's switch and disconnector, the hard plate information, the fiber optic link breakage information, and the fiber optic link breakage information of the merging unit.

[0063] Obtain analog quantity information from SV messages.

[0064] The data acquisition device forwards data to the security verification device through a secure data communication bastion host to ensure data transmission security and prevent the security verification device from accessing other secondary devices besides the data acquisition device.

[0065] 2. The safety check device is the core of the system, including: Safety Measures Logic Library: Stores information such as the name of each power plant's step-up substation, SCD files, and bay safety measures tickets. Each safety measures ticket includes a safety measures ticket header, safety measures entries, and implementation and recovery status.

[0066] Data parsing module: Parses the data acquired by the data acquisition device to obtain the safety measures object and its status.

[0067] Lightweight CNN model module: Recognizes images of objects in the field of safety.

[0068] Verification and comparison module: Combines data collection and image recognition results with the security ticket.

[0069] 3. The verification terminal is a mobile terminal device, including: Human-computer interaction module: Edit safety measures tasks and display verification results.

[0070] Camera module: Takes photos of the subjects.

[0071] Wireless communication module: Communicates with the safety verification device.

[0072] Audio output module: issues alarm signals and reads out the safety ticket text.

[0073] Example 2: Lightweight CNN Model like Figure 2 As shown, the lightweight CNN model in this embodiment includes a feature separation module, a feature processing module, a feature fusion module, and a channel attention module.

[0074] The feature separation module performs batch standardization and sorting reconstruction on the input feature maps, dividing them into high-variance and low-variance feature groups along the channel dimension. Specific steps include: The input feature map x is batch standardized using a normalization formula.

[0075] The feature maps are sorted and reconstructed based on the variance of the standardized feature maps.

[0076] The sorted and reconstructed feature maps are divided into high-variance feature group x1 and low-variance feature group x2 along the channel dimension.

[0077] The feature processing module processes the two sets of feature maps separately: The high-variance feature group x1 is subjected to target edge feature enhancement processing (using a Laplacian filter), and then grouped convolution (GWC) and point convolution (PWC) operations are performed to obtain the second high-variance feature group.

[0078] The low-variance feature group x2 is subjected to depthwise separable convolution (DWC) and pointwise convolution (PWC) operations in sequence to obtain the second low-variance feature group.

[0079] The feature fusion module performs complementary fusion processing on the feature maps of the second high-variance feature group and the second low-variance feature group: The two sets of feature maps are grouped along the channel dimension, and each group has the same shape (n×g×h×w).

[0080] Calculate the cosine similarity of each subgroup in the two groups.

[0081] Based on cosine similarity, the Hungarian algorithm is used to perform unique matching of groups, and then the matched groups are merged to obtain the fused output feature map y.

[0082] The channel attention module processes the fused output feature map and dynamically adjusts the channel weights: Perform average pooling on the output feature map y to generate the pooled output feature map v.

[0083] The channel location information p is embedded into the pooled output feature map v to obtain vp.

[0084] Performing a circular convolution operation yields the convolution result vout.

[0085] The final output feature map xout is obtained by performing a dot product between the convolution result vout and the output feature map y.

[0086] Example 3: Verification Process for Safety Measures like Figure 3 As shown, the safety measure execution verification process in this embodiment includes the following steps: System initialization Configure the Ancuo logic library in the Ancuo verification device through the human-computer interaction module of the Ancuo verification terminal.

[0087] Input photos of safety measures at the power plant's substation and train the image recognition module in the safety measure verification module.

[0088] Ancuo Mission Selection On the human-computer interaction module of the safety check verification terminal, select the safety check task and retrieve the corresponding safety check ticket from the safety check logic library.

[0089] The safety measure ticket contains information such as the target of the safety measure, the expected state, and the order of implementation.

[0090] Data acquisition and image recognition The safety check verification device acquires safety check-related data from the data acquisition device and analyzes it to obtain the safety check object and status.

[0091] The camera module of the verification terminal takes a photo of the verification object and transmits it to the image recognition module.

[0092] A lightweight CNN model is used to identify photos of safety objects and output the safety object and its status.

[0093] Verification and comparison The safety check module comprehensively compares the safety check objects and their status obtained through the data acquisition device and image recognition module with the corresponding safety check tickets in the safety check logic library.

[0094] The comparison includes whether the identity and status of the security measures target meet expectations, and whether the execution order is correct.

[0095] Results Feedback The human-computer interaction module of the Ancuo verification terminal displays the verification results of each item on the Ancuo ticket in real time.

[0096] The audio output module reads aloud the text of each item on the Ancuo Ticket.

[0097] When the safety measures tickets do not match, the corresponding safety measures ticket will be marked in red, and an alarm audio signal will be output through the audio output module.

[0098] Execution confirmation and archiving Electronic signature confirmation is required for the executed steps.

[0099] Generate a verification report to record any abnormal situations and handling suggestions during the implementation of safety measures.

[0100] The execution results and verification reports are uploaded to the management system to achieve full lifecycle management of safety measures tickets.

Claims

1. A method for verifying the execution steps of a secondary security ticket based on a lightweight CNN, characterized in that, Includes the following steps: Based on SCD, ICD and SPCD files, a digital model of the secondary equipment of the power plant's step-up substation is constructed, and a safety measure logic library is established. Construct a lightweight CNN model, including feature separation, feature processing, feature fusion, and channel attention modules; The device status images during the implementation of safety measures are collected via mobile terminals, and a lightweight CNN model is used for recognition. By combining the image recognition results with data obtained from the station control layer network and the process layer network, the execution status of safety measures is comprehensively judged. The actual state of the identified safety measures object is compared with the expected state in the safety measures ticket, and the verification results are fed back in real time.

2. The method according to claim 1, characterized in that, The construction of the lightweight CNN model includes: The input feature maps are batch standardized and sorted for reconstruction, and the feature maps are divided into high variance feature groups and low variance feature groups along the channel dimension; For high-variance feature groups, target edge feature enhancement processing is performed by grouped convolution and point convolution operations; Perform depthwise separable convolution and pointwise convolution operations on low-variance feature sets; The two sets of feature maps are grouped and complemented for fusion, and the channel weights are dynamically adjusted using a channel attention mechanism.

3. The method according to claim 1, characterized in that, The comprehensive assessment of the implementation status of the safety measures includes: Obtain information about the protection device's soft pressure plate, control word, hard pressure plate, and fiber optic link breakage from the MMS message; Obtain the switch location information, hard plate information, fiber optic link breakage information, and fiber optic link breakage information of the merging unit from the GOOSE message; Obtain analog quantity information from SV messages; Based on the above data and image recognition results, the implementation status of the safety measures is comprehensively judged.

4. The method according to claim 1, characterized in that, The verification results feedback includes: The verification results of each item on the safety check ticket are displayed in real time on the mobile terminal interface; The audio output module reads aloud the text of each item on the Ancuo ticket. When the safety measures tickets do not match, the corresponding safety measures ticket will be marked in red, and an alarm audio signal will be output through the audio output module.

5. A secondary security ticket execution step verification system based on lightweight CNN, characterized in that, include: The data acquisition device is used to access the station control layer network and the process layer network to obtain relevant data for safety measure verification; A secure data communication bastion host is used to forward data acquired by the data acquisition device to the security verification device; The Ancolic Verification Device is used to store the Ancolic logic library, parse the data acquired by the data acquisition device, deploy a lightweight CNN model for image recognition, and compare and verify the combined data acquisition and image recognition results with the Ancolic ticket. The safety check terminal includes a human-computer interaction module, a camera module, a wireless communication module, and an audio output module. It is used to edit safety check tasks, take photos of safety check objects, communicate with safety check devices, display check results, and output alarm signals.

6. The system according to claim 5, characterized in that, The lightweight CNN model in the safety verification device includes: The feature separation module is used to perform batch standardization and sorting reconstruction of the input feature maps, dividing them into high-variance and low-variance feature groups; The feature processing module is used to perform edge enhancement and grouped convolution on high variance feature groups, and depthwise separable convolution on low variance feature groups. The feature fusion module is used to perform complementary fusion of two sets of feature maps. The channel attention module is used to dynamically adjust the channel weights to obtain the final output feature map.

7. The system according to claim 5, characterized in that, The security logic library includes: Names of each power plant substation, SCD files associated with each power plant substation, and safety tickets included in each bay within each power plant substation; Each safety measure ticket includes a safety measure header, safety measure entries, and information on the implementation and restoration of safety measures. The safety measure items include the object of the safety measure and the expected state of the object of the safety measure.

8. The system according to claim 5, characterized in that, The human-computer interaction module of the safety measure verification terminal includes: The safety measure ticket display area is used to display the content of the currently implemented safety measure ticket; The image acquisition area is used to display images of the safety measures objects captured in real time by the camera module; The verification results area is used to display the verification results of the current step in real time; The control button area is used to perform operations such as next step, previous step, confirm execution, and generate report; The status information area is used to display the current operator, supervisor, execution time, and terminal status information.

Images