Relay protection device test system and method

Abstract

The invention relates to the technical field of power system testing, in particular to a relay protection device testing system and method. The system comprises a tested device, a tester, a device test case module, a test module, a test instrument interface module, a simulation master station module and a comprehensive diagnosis module. The output end of the device test case module is connected with the input end of the test module; the test module is respectively connected with the test instrument interface module, the simulation master station module and the comprehensive diagnosis module; the test instrument interface module is connected with the tested device through the tester, and the tested device is also connected with the test module through the simulation master station module. According to the invention, comprehensive and intelligent testing of the relay protection device is realized; the device can simulate various fault scenes, accurately detect the action response of the device, provide detailed performance evaluation and diagnosis reports through multi-path data feedback and comprehensive diagnosis, and effectively improve the test efficiency and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of power system testing, and in particular to a relay protection device testing system and method. Background Art

[0002] With the accelerated development of smart grids, relay protection devices, as the first line of defense for ensuring the safe and stable operation of power systems, have a reliability that directly determines the stability and safety of grid operations. However, traditional relay protection device testing systems generally use simulation testing methods based on a single signal source, relying on manual experience to perform offline analysis of fault recording data (such as COMTRADE format files). This model has significant limitations: insufficient test scenario coverage makes it difficult to fully simulate complex grid fault scenarios, especially the combination of multi-source and multi-type faults; weak dynamic response assessment: the lack of real-time quantitative analysis capabilities for the dynamic response characteristics of protection devices makes it difficult to accurately assess their operating speed and accuracy; low recognition rate for complex fault modes: the limited ability to identify complex modes such as mixed faults and hidden faults can easily lead to false or non-operation of protection devices; and poor multimodal data coordination. Existing systems fail to effectively integrate the temporal correlation between switch operation timing, telesignaling state changes, and transient waveform characteristics, resulting in a lack of closed-loop verification of the protection device's operating logic and difficulty in forming a complete fault deduction and assessment system.

[0003] These issues make traditional testing methods unable to meet the high-precision, full-dimensional testing requirements of new power systems for relay protection devices. Therefore, there is an urgent need to build an intelligent testing system that integrates multimodal data (such as transient waveforms, switching values, and telesignaling values) for time-series correlation analysis to accurately reconstruct fault scenarios and dynamically evaluate protection device performance. This system can overcome existing technical bottlenecks and provide more reliable and intelligent relay protection testing solutions for new power systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a relay protection device testing system and method in order to solve at least one of the above technical problems.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] A relay protection device test system includes a device under test, a tester, a device test case module, a test module, a test instrument interface module, a simulation master station module, and a comprehensive diagnosis module;

[0007] The output end of the device test case module is connected to the input end of the test module; the test module is respectively connected to the test instrument interface module, the simulation master station module, and the comprehensive diagnosis module;

[0008] The test instrument interface module is connected to the device under test through the test instrument, and the device under test is also connected to the test module through the simulated master station module;

[0009] The test module is used to generate a fault simulation command based on the test case sent by the device test case module, and perform multi-source data fusion and processing on the data fed back by the device under test based on a deep learning model;

[0010] The comprehensive diagnosis module is used to generate a comprehensive diagnosis report based on the preliminary evaluation report generated by the test module.

[0011] Furthermore, the test module includes: a file parsing unit, a signal processing unit, a feature engineering unit, and a deep learning unit;

[0012] The file parsing unit is used to parse and process file data from the device under test, extract key metadata, and convert unstructured file data into a structured array;

[0013] The signal processing unit is used to pre-process the waveform data, optimize the quality of the waveform data, and extract useful fault feature segments;

[0014] The feature engineering unit is used to encode, interpolate, dynamically scale, and extract features of device status information such as input, output, and remote signaling, and pre-process telemetry data;

[0015] The deep learning unit is used to fuse multi-source data based on a deep learning model and perform fault diagnosis and result generation.

[0016] Furthermore, the deep learning model includes, in order of execution: a 1D CNN model layer, an Embedding layer, a three-level attention layer, a Dense layer, and a Deepseek model layer; wherein,

[0017] The 1D CNN model layer is used to extract continuous waveform features and perform weighted fusion with discrete device status signals;

[0018] The three-level attention layer is used for deep fusion of multi-source data;

[0019] The Deepseek model layer uses the Deepseek-R1-Distill-QWen-32B model to perform status judgment and anomaly evaluation on the data output by the Dense layer.

[0020] Furthermore, the comprehensive diagnosis module includes: a test record database and a comprehensive diagnosis reporting unit;

[0021] The test record database is used to store all test records of the system and provide data support for subsequent analysis and query;

[0022] The comprehensive diagnosis reporting unit is used to analyze and process data based on a dynamic threshold adjustment algorithm, and then generate a detailed comprehensive diagnosis report.

[0023] Furthermore, the comprehensive diagnostic report includes: equipment status, comprehensive score, failure probability, switch consistency, remote signaling rationality, and recommended measures.

[0024] Furthermore, the test instrument interface module communicates with the test instrument via Ethernet to issue fault simulation commands and receive input information.

[0025] Furthermore, the simulated master station module communicates with the device under test via the 101 / 104 / 61850 communication protocol to obtain fault recording and telesignaling information of the device under test.

[0026] A method for testing a relay protection device, using any one of the relay protection device testing systems described above, the method comprising the following steps:

[0027] Connect the various modules in the test system in sequence and prepare to perform the test tasks;

[0028] The test module generates a fault simulation command based on the test case of the device test case module;

[0029] The fault simulation command is sent to the tester via the tester interface module;

[0030] The tester simulates a fault and sends a fault signal to the device under test, and at the same time detects the operation of the device under test and feeds back the result to the test module; the simulated master station module monitors the status of the device under test and feeds back data to the test module;

[0031] The testing module generates a preliminary evaluation report based on the received feedback information, and the comprehensive diagnosis module generates a comprehensive diagnosis report based on the preliminary evaluation report.

[0032] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the relay protection device testing method described above when executed by the processor.

[0033] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the relay protection device testing method described above is implemented.

[0034] The beneficial effects of the present invention are:

[0035] This paper combines the waveform feature extraction capabilities of a 1D CNN with the discrete semantic modeling of the Embedding layer and the information focusing capabilities of a hierarchical attention mechanism to achieve deep integration of multi-source data, including waveforms, telemetry, telesignaling, and switching values. In fault diagnosis, the 1D CNN accurately captures waveform mutations, the Embedding layer deeply explores potential correlations between switch state combinations, and the hierarchical attention mechanism dynamically focuses on key fault data. These three collaboratively construct an enhanced feature representation system. Specifically, the present invention utilizes the GELU activation function to enhance the 1D CNN's representation of the high-frequency characteristics of power system fault waveforms and combines it with global pooling technology to achieve feature dimensionality reduction. Furthermore, the present invention introduces the Deepseek-R1-Distill-QWen-32B large model, endowing the system with powerful knowledge reasoning capabilities. Based on training data and fused multi-source data, it can automatically generate judgment results containing key information such as device status, comprehensive score, and fault probability. This enables comprehensive and intelligent testing of relay protection devices, significantly improving the accuracy and efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a schematic diagram of the principle of a relay protection device test system according to one embodiment of the present invention;

[0037] Figure 2 A schematic diagram of the deep learning model principle of one embodiment of the present invention;

[0038] Figure 3 A schematic structural diagram of a relay protection device test system according to an embodiment of the present invention;

[0039] Figure 4 This is a flow chart of a relay protection device testing method according to one embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of a method for testing a relay protection device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention will now be discussed with reference to exemplary embodiments. It should be understood that the embodiments discussed are only intended to enable those skilled in the art to better understand and implement the present invention, rather than to imply any limitation on the scope of the present invention.

[0042] As used herein, the term "including" and variations thereof are to be interpreted as open-ended terms meaning "including, but not limited to." The term "based on" is to be interpreted as "based, at least in part, on." The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment."

[0043] Example 1

[0044] Figure 1 This is a schematic diagram of the principle of a relay protection device test system according to one embodiment of the present invention. Figure 1 As shown, according to one embodiment of the present invention, a relay protection device test system includes a device under test 10, a tester 20, a device test case module 30, a test module 40, a test instrument interface module 50, a simulation master station module 60, and a comprehensive diagnosis module 70;

[0045] The output end of the device test case module 30 is connected to the input end of the test module 40; the test module 40 is respectively connected to the test instrument interface module 50, the simulation master station module 60, and the comprehensive diagnosis module 70;

[0046] The test instrument interface module 50 is connected to the device under test 10 via the tester 20, and the device under test 10 is also connected to the test module 40 via the simulated master station module 60;

[0047] The test module 40 is used to generate fault simulation commands based on the test cases sent by the device test case module 30, and perform multi-source data fusion and processing on the data fed back by the device under test 10 based on the deep learning model;

[0048] The comprehensive diagnosis module 70 is used to generate a comprehensive diagnosis report based on the preliminary evaluation report generated by the testing module 40 .

[0049] This embodiment provides a relay protection device testing system, comprising a device under test (DUT) 10, a tester 20, a device test case module 30, a test module 40, a test instrument interface module 50, a simulated master station module 60, and a comprehensive diagnostic module 70. In the relay protection device testing system, an operator selects a corresponding test case from the device test case module 30 based on test requirements. The test case defines specific fault types, such as line short circuit faults and line overcurrent faults. The test module 40 generates a fault simulation command based on the selected test case. The test instrument interface module 50 sends the fault simulation command to the tester 20 via Ethernet. Upon receiving the command, the tester 20 utilizes its state output, binary input and output, and waveform playback functions to simulate the fault scenario and generate corresponding fault current, voltage, and other signals. The tester 20 outputs the simulated fault signal to the input port of the DUT 10 via a dedicated test line. Upon receiving the simulated fault signal, the DUT 10 processes it according to its internal protection logic, determines whether a fault has occurred, and performs corresponding actions, such as tripping or alarming. Feedback of the device under test (DUT) data occurs through two pathways: 1. The tester 20 detects the tripping signals of the DUT 10, determines whether the protection system is operating correctly, and transmits the results to the test module 40. 2. The simulated master station module 60 communicates with the DUT 10 to obtain fault waveforms and telesignaling information. The test module 40 receives feedback from the tester 20 and the simulated master station module 60, records the test results, and analyzes and evaluates them. The comprehensive diagnostic module 70 further analyzes the test results, comprehensively evaluates the performance of the DUT 10, and provides a diagnostic report and recommended measures.

[0050] The relay protection device testing system of the present invention realizes comprehensive and intelligent testing of relay protection devices by integrating multiple functional modules; the system can simulate various fault scenarios, accurately detect the action response of the device, and provide detailed performance evaluation and diagnostic reports through multi-path data feedback and comprehensive diagnosis, effectively improving test efficiency and reliability.

[0051] According to one embodiment of the present invention, the testing module 40 includes: a file parsing unit 41, a signal processing unit 42, a feature engineering unit 43, and a deep learning unit 44;

[0052] The file parsing unit 41 is used to parse and process file data from the device under test 10, extract key metadata, and convert unstructured file data into a structured array;

[0053] The signal processing unit 42 is used to pre-process the waveform data, optimize the quality of the waveform data, and extract useful fault feature segments;

[0054] The feature engineering unit 43 is used to encode, interpolate, dynamically scale, and extract features of device status information such as input, output, and telemetering, and pre-process telemetry data;

[0055] The deep learning unit 44 is used to fuse multi-source data based on a deep learning model and perform fault diagnosis and result generation.

[0056] Figure 2 This is a schematic diagram of the deep learning model principle of an embodiment of the present invention. Figure 2 As shown, preferably, the deep learning model includes, in order of execution: a 1D CNN model layer, an Embedding layer, a three-level attention layer, a Dense layer, and a Deepseek model layer; wherein,

[0057] The 1D CNN model layer is used to extract continuous waveform features and perform weighted fusion with discrete device status signals;

[0058] The three-level attention layer is used for deep fusion of multi-source data;

[0059] The Deepseek model layer uses the Deepseek-R1-Distill-QWen-32B model to perform status judgment and anomaly evaluation on the data output by the Dense layer.

[0060] In this embodiment, the test module 40 includes a file parsing unit 41, a signal processing unit 42, a feature engineering unit 43, and a deep learning unit 44. The file parsing unit 41 is used to parse configuration files such as CFG and DAT waveform data, extract key metadata, restructure unstructured file data into a structured array, and store data alternately according to timestamp and channel data. The signal processing unit 42 is used to eliminate baseband interference using a power frequency notch filter, standardize channels, and extract fault feature segments. The feature engineering unit 43 is used to process switching quantities, including preprocessing output data through time alignment and sliding window statistical methods; record and process telesignaling data, including preprocessing telesignaling data through interpolation methods; and process telemetry data, including linear interpolation and statistical feature extraction. The deep learning unit 44 is used to fuse multi-source data based on a deep learning model and perform fault diagnosis and result generation.

[0061] In the deep learning model constructed by this invention, the 1D CNN model layer uses a 1D CNN model and a global pooling operation in the pooling layer to extract the key time-domain features of the fault waveform. The 1D CNN model can use the GELU (Gaussian Error Linear Unit) activation function to enhance the expression of high-frequency information.

[0062] GELU(X))=X·P(X≤X)

[0063] Where X represents a random variable that obeys a standard normal distribution, P(X≤x) represents the probability that X is less than or equal to x; x represents the input variable. The GELU activation function can be approximately expressed as:

[0064] GELU(x)≈0.5x(1+tanh(x+0.044715x 3 ))

[0065] The GELU activation function is more sensitive to subtle changes in high-frequency fault signals, extracting and expressing high-frequency features more accurately. For high-frequency oscillations or spikes that appear at the moment of a fault, the GELU can better convey these signal characteristics to subsequent network layers, helping to improve the accuracy and timeliness of fault diagnosis.

[0066] The global pooling operation of the pooling layer refers to the use of global maximum pooling (GMP) in the 1D CNN pooling layer, which can extract the maximum value in the fault waveform, that is, the most significant feature.

[0067] To achieve better results, an embedding layer is added to automatically learn the semantic relationships between state combinations. The output data of the 1D CNN model layer is input into the embedding layer for training. During training, the deep learning model automatically adjusts the parameters of the embedding layer based on feedback from the loss function, ensuring that data such as the switch vector can best represent the semantic relationships between state combinations. The embedding layer uses mean squared error as the loss function, and the output of the embedding layer serves as the input to the three-level attention layer.

[0068] The three-level attention layer is the waveform feature temporal attention layer, the cross-modal attention layer The first layer uses temporal attention to capture the key time points of waveform features; the second layer establishes a correlation matrix between state signals and waveform features; and the third layer uses spatial attention to highlight important parameters of telemetry data.

[0069] The data is unified to the model's required 512 dimensions through a Dense layer and projected into the large model input space. The Dense layer has 512 neurons and uses the ReLU activation function. The data after layered attention fusion is typically shaped like (batch_size, feature_dim), where feature_dim is the feature dimension of the fused data. This data is fed into a Dense layer to produce 512-dimensional data, which serves as the input for the Deepseek model layer.

[0070] The Deepseek model layer calls the trained Deepseek-R1-Distill-QWen-32B and obtains the model output. Deepseek-R1-Distill-QWen-32B is a lightweight, large language model optimized for specialized domains. It natively supports 512-dimensional structured feature vector input (such as waveform features and device status codes), eliminating the need for text conversion.

[0071] The relay protection device testing system of the present invention achieves comprehensive testing and intelligent diagnosis of relay protection devices through the file parsing, signal processing, feature engineering, and deep learning units in the test module. The file parsing unit converts unstructured data into structured arrays, the signal processing unit optimizes waveform data quality, the feature engineering unit encodes and extracts features from device status information, and the deep learning unit utilizes technologies such as 1D CNN, embedding layers, a three-level attention layer, and a DeepSeek model to fuse multi-source data, improving the accuracy and timeliness of fault diagnosis. This invention effectively enhances the intelligence level and diagnostic accuracy of relay protection device testing.

[0072] According to one embodiment of the present invention, the comprehensive diagnosis module 70 includes: a test record database 71 and a comprehensive diagnosis reporting unit 72;

[0073] The test record database 71 is used to store all test records of the system and provide data support for subsequent analysis and query;

[0074] The comprehensive diagnosis reporting unit 72 is used to analyze and process the data based on the dynamic threshold adjustment algorithm, and then generate a detailed comprehensive diagnosis report.

[0075] Preferably, the comprehensive diagnostic report includes: equipment status, comprehensive score, failure probability, switch consistency, remote signaling rationality, and recommended measures.

[0076] In this embodiment, the comprehensive diagnosis module 70 is responsible for interpretable output, that is, converting the output of the deep learning model into a comprehensive score breakdown and quantitatively displaying the fault probability, providing specific treatment suggestions so that users can clearly understand the operating status of the equipment and take appropriate measures. The comprehensive diagnosis module 70 includes a test record database 71 and a comprehensive diagnosis report unit 72. The test record database 71 is responsible for storing all test records and providing data support for subsequent analysis and query; while the comprehensive diagnosis report unit 72 analyzes and processes the data based on a dynamic threshold adjustment algorithm to generate a detailed comprehensive diagnosis report. The content of the comprehensive diagnosis report includes equipment status, comprehensive score, fault probability, switch consistency, remote signal rationality and recommended measures.

[0077] The comprehensive diagnostic module of the present invention stores test records and generates detailed diagnostic reports, clearly presenting the device status and providing specific suggestions, effectively improving the user's understanding and management capabilities of the device's operating status.

[0078] According to one embodiment of the present invention, the test instrument interface module 50 communicates with the test instrument 20 via Ethernet to issue fault simulation commands and receive input information.

[0079] The simulated master station module 60 communicates with the device under test 10 via the 101 / 104 / 61850 communication protocol to obtain the fault recording and telesignaling information of the device under test 10.

[0080] In this embodiment, the test instrument interface module 50 establishes a communication connection with the tester 20 via Ethernet, and is responsible for sending fault simulation commands to the tester 20 and receiving the input information fed back by the tester 20. This Ethernet communication method can achieve high-speed and reliable data transmission, ensuring the accurate issuance of commands and the timely receipt of feedback information. At the same time, the simulation master station module 60 uses communication protocols such as 101, 104, 61850 to communicate with the device under test 10 to obtain the fault recording data and telesignaling information of the device under test 10. These communication protocols are standard protocols widely used in power systems and can ensure compatibility and interoperability between different devices. The device under test 10 communicates with the test module 40 through the communication protocol engine to ensure that the test module 40 can receive and process data from the device under test 10.

[0081] The present invention realizes efficient communication and data interaction between components of the test system by combining Ethernet and standard power communication protocols, thereby improving test efficiency and accuracy.

[0082] Example 2

[0083] Figure 3 FIG1 is a schematic diagram of the structure of a relay protection device test system according to an embodiment of the present invention. Figure 3 As shown, according to one embodiment of the present invention, a relay protection device test system includes: ① an interface layer, ② an automatic test layer, and ③ a secondary development platform layer. This layered architecture makes the system modular and scalable. ① The interface layer ensures compatibility with different devices, ② the automatic test layer realizes automated and intelligent testing, and ③ the secondary development platform layer provides customization possibilities for professional users. Specifically, it includes:

[0084] The devices under test 10 include various relay protection devices, including high-voltage protection devices, low-voltage protection devices, and distribution network devices. The tester 20 features status output, binary input and output, and waveform playback. It connects to the device under test 10 via a dedicated test line to simulate its operation and fault conditions in the circuit. The device test case module 30 stores a large number of automated and semi-automated test cases for relay protection devices, providing test case support for the system.

[0085] ① The interface layer serves as a bridge for interaction between the system and external devices, and is responsible for data input and output, including the test instrument interface module 50 and the simulation master station module 60.

[0086] The test instrument interface module 50 communicates with the tester 20 via Ethernet and is responsible for sending fault simulation commands and receiving input signals. The test instrument interface module 50 connects the test module 40 and the tester 20, enabling communication between the two via Ethernet. Using the test instrument interface module 50, the test module 40 can control the tester 20 via Ethernet to perform fault simulations, including real-time simulations of faults such as line short circuits and overcurrents, as well as ground faults using real-world waveform playback.

[0087] The simulated master station module 60 communicates with the device under test 10 via the 101 / 104 / 61850 communication protocol to obtain fault recording and telesignaling information.

[0088] The second layer, the automated test layer, is the core of the system, responsible for automated test execution and data processing. It includes the device test system main program (test module 40) and the communication protocol engine. Test module 40 generates fault simulation commands, analyzes feedback information, and provides a test case execution environment. It also includes a file parsing unit 41, a signal processing unit 42, a feature engineering unit 43, and a deep learning unit 44. The communication protocol engine processes data from the device under test (DUT) 10, ensuring effective communication with the test module 40 and parsing and generating data frames that conform to specific communication protocols.

[0089] The test module 40 generates fault simulation commands based on the test cases of the device test case module 30, analyzes the feedback information and generates a preliminary evaluation report. The file parsing unit 41 parses the file data of the device under test 10, extracts key metadata and converts the unstructured file data into a structured array. The signal processing unit 42 preprocesses the waveform data, optimizes the quality and extracts the fault feature segment. The file parsing unit 41 and the signal processing unit 42 are used to collect and preprocess data; including: eliminating 50 / 60Hz baseband interference through the power frequency notch filter, performing channel-level normalization to eliminate sensor range differences, intercepting the waveform (first 200ms + last 800ms) based on the trigger point dynamic window, and using mirror padding to process non-integer cycle data. The feature engineering unit 43 encodes, interpolates, dynamically scales the range and extracts features of the device status information such as input, output and telemetering, and preprocesses the telemetry data. Specifically, it includes: using different encoding methods to process binary switch quantities such as input and output signals and telemetering information; using the interpolation method to interpolate and align the waveform data of discrete telemetering state quantities (such as switch opening / closing, pressure plate insertion and withdrawal) with the state signal based on the fault trigger time, and performing relative timestamp encoding; implementing time window filtering on the switch input / output state (0 / 1) sequence to achieve anti-shake, and generating time difference encoding relative to the fault moment; implementing linear interpolation on the telemetry data to handle missing values, performing dynamic range scaling based on the equipment rating, and extracting statistical features at the same time.

[0090] Deep Learning Unit 44 builds a deep learning model, using a 1D CNN model and global pooling operations in the pooling layer to extract key time-domain features of the fault waveform. A three-level attention layer is added for hierarchical attention fusion. The deep learning model is trained in stages. Stage 1 freezes the large model to optimize feature extraction capabilities and trains the 1D CNN and three-level attention layers. Stage 2 performs joint fine-tuning, unfreezing the Deepseek adaptation layer and projection layer for joint training.

[0091] The 1D CNN model uses the GELU activation function to enhance the representation of high-frequency information. Global pooling uses global max pooling to extract the most significant features. These technologies collectively improve the accuracy and timeliness of fault diagnosis. Specifically, the 1D CNN model uses the GELU (Gaussian Error Linear Unit) activation function to enhance the representation of high-frequency information.

[0092] GELU(X)=X·P(X≤X)

[0093] Where X is a random variable that follows a standard normal distribution, and P(X≤x) represents the probability that X is less than or equal to x. It can be approximately expressed as:

[0094] GELU(x)≈0.5x(1+tanh(x+0.044715x3 ))

[0095] The GELU activation function is more sensitive to subtle changes in high-frequency fault signals, extracting and expressing high-frequency features more accurately. For high-frequency oscillations or spikes that appear at the moment of a fault, GELU can better transmit these signal features to subsequent network layers, helping to improve the accuracy and timeliness of fault diagnosis.

[0096] The global pooling operation refers to the use of global maximum pooling (GMP) in the 1D CNN pooling layer, which can extract the maximum value in the fault waveform, that is, the most significant feature.

[0097] By adding an Embedding layer, the semantic relationships between state combinations are automatically learned to achieve better results. The output data of feature engineering unit 43 is fed into the Embedding layer for training. During training, the model automatically adjusts the parameters of the Embedding layer based on feedback from the loss function, ensuring that data such as the switch vector can best represent the semantic relationships between state combinations. This layer uses mean squared error as the loss function. The output of this layer can serve as the input to the third-level attention layer.

[0098] The three-level attention layer is the waveform feature temporal attention layer, the cross-modal attention layer Layer, spatial attention (feature importance weighting) layer. The first layer uses temporal attention to capture the key time points of waveform features; the second layer establishes the correlation matrix between state signals and waveform features; the third layer uses spatial attention to highlight important parameters of telemetry data;

[0099] The 1D CNN model and the global pooling operation of the pooling layer in the deep learning model extract the key time-domain features of the fault waveform. A three-level attention layer is added after the 1D CNN output to perform hierarchical attention fusion on the waveform features, telesignaling codes, switch code, and telemetry data. After fusion, the Dense layer unifies the data to the required 512 dimensions and projects it into the large model input space. The trained Deepseek-R1-Distill-QWen-32B is called to obtain the model output.

[0100] ③ The secondary development platform layer provides a flexible development environment to support users to carry out customized development according to specific needs; including: comprehensive diagnosis module 70; the comprehensive diagnosis module 70 includes a system test record library (test record database 71) and a comprehensive diagnosis report unit 72, which is responsible for storing test records and generating comprehensive diagnosis reports.

[0101] The test record database 71 is used to store all test records of the system and provide data support for subsequent analysis and query;

[0102] The Comprehensive Diagnostic Report Unit 72 uses a dynamic threshold adjustment algorithm to conduct in-depth data analysis and output a detailed comprehensive diagnostic report in formats such as Word and WPS. It transforms complex model output into a clear and understandable comprehensive score breakdown, quantifying the probability of failure and providing specific action recommendations. The report covers key information such as device status, comprehensive score, failure probability, switch consistency, remote signaling rationality, and recommended actions, providing users with a comprehensive and intuitive tool for evaluating device performance.

[0103] The relay protection device testing system of this invention achieves efficient automated testing and intelligent diagnosis through a modular, layered architecture. The interface layer ensures compatibility with different devices, while the automated testing layer integrates deep learning and feature engineering to improve the accuracy and timeliness of fault diagnosis. The secondary development platform layer provides a flexible and customized development environment, and the comprehensive diagnostic module generates detailed diagnostic reports, comprehensively improving the testing efficiency and reliability of relay protection devices.

[0104] Example 3

[0105] Figure 4 This is a flow chart of a relay protection device testing method according to an embodiment of the present invention. Figure 4 As shown, a relay protection device testing method adopts any relay protection device testing system of the present invention, and the method includes the following steps:

[0106] Step S102: Connect the modules in the test system in sequence to prepare for executing the test task.

[0107] Step S104, the test module 40 generates a fault simulation command based on the test case of the device test case module 30;

[0108] Step S106 , the fault simulation command is sent to the tester 20 via the tester interface module 50 ;

[0109] In step S108, the tester 20 simulates a fault and sends a fault signal to the device under test 10, while detecting the operation of the device under test 10 and feeding back the result to the test module 40; the simulated master station module 60 monitors the status of the device under test 10 and feeds back the data to the test module 40;

[0110] In step S110 , the testing module 40 generates a preliminary evaluation report based on the received feedback information, and the comprehensive diagnosis module 70 generates a comprehensive diagnosis report based on the preliminary evaluation report.

[0111] This embodiment proposes a relay protection device testing method using a relay protection device testing system based on a large model and multi-source data fusion. Comprehensive testing and intelligent diagnosis of relay protection devices are achieved through the following steps: First, the various modules in the test system are connected sequentially to prepare for the test task. Next, the test module 40 sets the device type under test based on the selected test case, generates a fault simulation command, and sends it to the tester 20 via the test instrument interface module 50. The tester 20 simulates the fault and sends a fault signal to the device under test 10, while simultaneously detecting the operation of the device under test 10 and feeding back the results to the test module 40. The test module 40 acquires fault recordings, telesignaling, hard-node input, and telemetry data from the device under test 10. The test module 40 processes the hard-node input and output signals, telesignaling, and telemetry data through methods such as interpolation, various encoding methods, dynamic range scaling, time window filtering, and channel-level normalization, achieving time alignment, standardization, filtering, and valid segment extraction, thereby forming a data set suitable for multimodal fusion. Specifically, the following steps are performed: using the parse_cfg and load_dat functions contained in the file parsing unit 41 to process the fault waveform .cfg and .dat files and extract the waveform data. Using the signal processing unit 42, the power frequency notch filter is used to eliminate 50 / 60Hz baseband interference, channel-level normalization is performed to eliminate sensor range differences, the fault point waveform (first 200ms + last 800ms) is intercepted based on the trigger point dynamic window, and non-integer cycle data is processed using mirror padding. The fault point time can be obtained from the output feedback of the tester 20. Using the feature engineering unit 43, the acquired hard node input and output signals, telesignaling, telemetry, and other information are processed through different encoding, interpolation, and valid segment extraction methods. The 1D CNN convolutional neural network, GELU activation function, and global pooling operation in the deep learning unit 44 are used to extract the key time-domain features of the fault waveform. The time-domain key features of the fault waveform, along with the processed and encoded hard-node input and output signals, telesignaling, and telemetry datasets, are input into the trained Embedding layer. The output of the Embedding layer is input into the three-level attention layer to achieve deep fusion of multi-source data. The Dense layer unifies the data to the required 512 dimensions and projects it into the large model input space. The trained Deepseek-R1-Distill-QWen-32B is then called to obtain the model output. The comprehensive diagnostic module 70 analyzes the output of the deep learning model, displaying device status, comprehensive score, fault probability, switch consistency, telesignaling rationality, and recommended measures.

[0112] Compared to conventional 1D CNNs, the deep learning model constructed in Deep Learning Unit 44 breaks through the limitations of traditional single-series feature extraction by introducing a multimodal hierarchical attention mechanism and deep state encoding. Compared to the native Deepseek model, through front-end lightweight feature engineering and knowledge distillation adaptation, it overcomes the computational redundancy of large models directly processing power signals and compensates for the deviation in the understanding of power equipment control logic by pure text pre-trained models. The deep learning model specifically includes:

[0113] (1) 1D CNN model layers including convolutional layers, pooling layers, and fully connected layers;

[0114] To accelerate model convergence, batch normalization can be used. Before the feature maps output by the convolutional layer are passed to the activation function, they are normalized by the batch normalization layer to ensure a stable distribution of the input data. To improve the model's feature extraction capabilities, the number of convolution kernels can also be increased from 32 to 64.

[0115] (2) The Embedding layer can convert discrete data such as telesignals and switch quantities into low-dimensional vectors, explore the semantic relationships between these data, and provide more meaningful feature representations for subsequent analysis. For example, different switch state combinations can reflect their potential correlation in power system operation after Embedding.

[0116] (3) The hierarchical attention fusion module realizes the intelligent fusion of waveform features, telesignaling status, switch action, and telemetry data through three levels of attention (temporal focus, cross-modal interaction, and spatial gating).

[0117] (4) Dense layer: The number of neurons in the Dense layer is set to 512, and the activation function is ReLU. The shape of the data after layered attention fusion is generally (batch_size, feature_dim), where feature_dim is the feature dimension of the fused data. By inputting this data into the Dense layer, 512-dimensional data can be obtained as the input required by the Deepseek-R1-Distill-QWen-32B model.

[0118] (5) Deepseek model layer: using Deepseek-R1-Distill-QWen-32B model

[0119] Deepseek-R1-Distill-QWen-32B is a lightweight large language model optimized for professional fields. It natively supports 512-dimensional structured feature vector input (such as waveform features and device status codes) without the need for text conversion.

[0120] The deep learning model of the present invention breaks through the limitations of traditional single temporal feature extraction by introducing a multimodal hierarchical attention mechanism and deep state coding. Conventional 1D CNN can only process single waveform data, relies on artificial feature engineering and is difficult to capture the spatiotemporal correlation between the switch action logic and the transient waveform. This solution realizes the intelligent fusion of waveform features, telesignaling status, and switch action through a three-level attention layer (temporal focus, cross-modal interaction, and spatial gating). At the same time, it enhances the nonlinear expression ability through the GELU activation function, and solves the problem of missed detection of weak faults caused by the vanishing gradient of the conventional ReLU in the zero zone. The present invention overcomes the defects of large models in directly processing power signals through front-end lightweight feature engineering and knowledge distillation adaptation, supports complex fault diagnosis, and can simultaneously analyze multi-dimensional features such as waveform mutations, switch action timing deviations, and telemetry anomalies, greatly improving the test sensitivity for hidden faults such as AC / DC hybrid faults and high-resistance grounding.

[0121] This invention combines the waveform feature extraction capabilities of 1D CNN, the discrete semantic modeling capabilities of the Embedding layer, and the hierarchical information focusing capabilities of hierarchical attention to achieve deep fusion of multi-source data such as waveforms, telemetry, telesignaling, and switching quantities. In fault diagnosis scenarios, 1D CNN captures waveform mutation characteristics, the Embedding layer explores potential correlations between switch state combinations, and hierarchical attention dynamically focuses on key data during the fault period. The three form a complementary and enhanced feature representation system. In response to the high-frequency characteristics of power system fault waveforms, this invention uses the GELU activation function to enhance the 1D CNN convolutional neural network's ability to express high-frequency signals, and combines it with global pooling technology to achieve feature dimensionality reduction. This invention introduces the Deepseek-R1-Distill-QWen-32B large model, which gives the system knowledge reasoning capabilities. By combining training data with deeply fused multi-source data, it can automatically generate judgment results such as device status, comprehensive score, and fault probability, thus achieving full-dimensional, intelligent testing of relay protection devices.

[0122] Example 4

[0123] According to one embodiment of the present invention, an electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, any relay protection device testing method of the present invention is implemented.

[0124] According to one embodiment of the present invention, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, any relay protection device testing method of the present invention is implemented.

[0125] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and media can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0126] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention herein is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in this application.

[0127] It should be understood that the size of the serial numbers of each step in the content of the invention and the embodiments of the present invention does not absolutely mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

Claims

1. A relay protection device test system, comprising a device under test (10) and a tester (20); characterized in that: The test system further comprises: a device test case module (30), a test module (40), a test instrument interface module (50), a simulation master station module (60), and a comprehensive diagnosis module (70); The output end of the device test case module (30) is connected to the input end of the test module (40); the test module (40) is respectively connected to the test instrument interface module (50), the simulation master station module (60), and the comprehensive diagnosis module (70); The test instrument interface module (50) is connected to the device under test (10) via the test instrument (20), and the device under test (10) is also connected to the test module (40) via the simulated master station module (60); The test module (40) is used to generate a fault simulation command based on the test case sent by the device test case module (30), and to perform multi-source data fusion and processing on the data fed back by the device under test (10) based on a deep learning model; The comprehensive diagnosis module (70) is used to generate a comprehensive diagnosis report based on the preliminary evaluation report generated by the test module (40).

2. The relay protection device test system according to claim 1, characterized in that: The test module (40) includes: a file parsing unit (41), a signal processing unit (42), a feature engineering unit (43), and a deep learning unit (44); The file parsing unit (41) is used to parse and process file data from the device under test (10), extract key metadata, and convert unstructured file data into a structured array; The signal processing unit (42) is used to pre-process the waveform data, optimize the quality of the waveform data, and extract useful fault feature segments; The feature engineering unit (43) is used to encode, interpolate, dynamically scale, and extract features of device status information such as input, output, and remote signaling, and pre-process telemetry data; The deep learning unit (44) is used to fuse multi-source data based on a deep learning model and perform fault diagnosis and result generation.

3. The relay protection device test system according to claim 2, characterized in that: The deep learning model includes, in order of execution: a 1D CNN model layer, an Embedding layer, a three-level attention layer, a Dense layer, and a Deepseek model layer; wherein, The 1D CNN model layer is used to extract continuous waveform features and perform weighted fusion with discrete device status signals; The three-level attention layer is used for deep fusion of multi-source data; The Deepseek model layer uses the Deepseek-R1-Distill-QWen-32B model to perform status judgment and anomaly evaluation on the data output by the Dense layer.

4. The relay protection device testing system according to claim 1, characterized in that: The comprehensive diagnosis module (70) includes: a test record database (71) and a comprehensive diagnosis report unit (72); The test record database (71) is used to store all test records of the system and provide data support for subsequent analysis and query; The comprehensive diagnosis reporting unit (72) is used to analyze and process data based on a dynamic threshold adjustment algorithm, and then generate a detailed comprehensive diagnosis report.

5. The relay protection device test system according to claim 4, characterized in that: The comprehensive diagnostic report includes: equipment status, comprehensive score, failure probability, switch consistency, remote signal rationality, and recommended measures.

6. The relay protection device testing system according to claim 1, characterized in that: The test instrument interface module (50) communicates with the test instrument (20) via Ethernet to issue fault simulation commands and receive input information.

7. The relay protection device testing system according to claim 1, wherein: The simulated master station module (60) communicates with the device under test (10) via the 101 / 104 / 61850 communication protocol to obtain fault recording and telesignaling information of the device under test (10).

8. A method for testing a relay protection device, using the relay protection device testing system according to any one of claims 1 to 7, characterized in that: The method comprises the following steps: Connect the various modules in the test system in sequence and prepare to perform the test tasks; The test module (40) generates a fault simulation command based on the test case of the device test case module (30); The fault simulation command is sent to the tester (20) via the tester interface module (50); The tester (20) simulates a fault and sends a fault signal to the device under test (10), and at the same time detects the operation of the device under test (10) and feeds back the result to the test module (40); the simulated master station module (60) monitors the status of the device under test (10) and feeds back data to the test module (40); The testing module (40) generates a preliminary evaluation report based on the received feedback information, and the comprehensive diagnosis module (70) generates a comprehensive diagnosis report based on the preliminary evaluation report.

9. An electronic device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the computer program implements the relay protection device testing method according to claim 8 when executed by the processor.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the relay protection device testing method according to claim 8 is implemented.

Images