A test method based on a main helium fan frequency converter fault detection platform

Abstract

This invention discloses a testing method based on a fault detection platform for a main helium blower frequency converter, belonging to the field of industrial equipment fault detection technology. The method includes acquiring the original operating parameters of the target frequency converter, processing and generating a multi-dimensional time-series operating parameter stream; constructing a configurable fault mode library, separating core feature subsets and auxiliary feature subsets according to a key feature extraction strategy, and generating a hybrid feature sequence containing simulated fault information. The hybrid sequence is coupled with the auxiliary subset to construct a fault feature fusion matrix, which is input into an adaptability verification engine. The auxiliary subset simulates nuclear power environment constraints to verify the authenticity of the fault information, and outputs test cases with environment identification. This method uses iterative training of a fault detection algorithm to generate an optimized model adapted to the target environment and deploy it on the platform. It then performs online analysis on the new standardized multi-dimensional time-series operating parameter stream, outputting fault diagnosis results and confidence levels.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment fault detection technology, specifically a test method based on a fault detection platform for a main helium blower frequency converter. Background Technology

[0002] As a critical piece of equipment in nuclear power systems, the main helium blower frequency converter relies on the analysis of operating parameters and model-based diagnosis for fault detection. Existing technologies typically involve acquiring the converter's raw operating parameters, preprocessing them to extract features, combining these with preset fault modes to generate test cases, training a fault detection model, and then deploying it. While these methods focus on the correlation between parameters and faults, they do not delve into the different roles of features in fault characterization and environmental adaptation.

[0003] Existing technical solutions have shortcomings. The feature extraction process does not separate core fault characterization features from environmental adaptation and discrimination features, often using all features together, which leads to fault signals being interfered with by environmental noise. When simulating fault injection, there is a lack of correlation verification with the actual nuclear power plant operating environment. Fault information in test cases may conflict with constraints such as temperature fluctuations and load changes in real scenarios, resulting in insufficient adaptability of the trained model to complex environments during online diagnosis, which can easily lead to misjudgments or missed judgments.

[0004] To address the aforementioned problems, this invention aims to solve how to divide features into core feature subsets and auxiliary feature subsets based on key feature extraction strategies to generate a hybrid feature sequence containing simulated fault information; and how to use the auxiliary feature subsets to simulate the constraints of the actual nuclear power plant operating environment, verify the environmental authenticity of the simulated fault information in the hybrid feature sequence, form a set of test cases adapted to the actual environment, and thereby improve the adaptability of the fault detection model to nuclear power plant scenarios. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a test method based on a fault detection platform for the main helium blower inverter, comprising: After processing the original set of operating parameters of the target frequency converter, a multi-dimensional time-series operating parameter stream is generated, a configurable fault mode library is constructed, and according to the preset key feature extraction strategy, the core feature subset for fault characterization and the auxiliary feature subset for environmental adaptability judgment are separated to generate a hybrid feature sequence containing simulated fault information. The hybrid feature sequence is coupled and associated with the auxiliary feature subset to construct a fault feature fusion matrix for testing; The fault feature fusion matrix is ​​input into a preset adaptability verification engine. The adaptability verification engine performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence based on the auxiliary feature subset simulating the constraints of the actual nuclear power plant operating environment. The adaptability verification engine outputs the fault characteristics after environment verification, and combines them with the fault tags in the configurable fault mode library to form a set of test cases with environment identifiers. Using the set of test cases with environmental identifiers, the preset fault detection algorithm is iteratively trained and its parameters are tuned to generate an optimized fault detection model adapted to the target nuclear power plant operating environment. The optimized fault detection model is deployed on the main helium blower frequency converter fault detection platform to perform online analysis on the new standardized multi-dimensional time-series operating parameter stream, and output fault diagnosis results and confidence levels.

[0006] Furthermore, a hybrid feature sequence containing simulated fault information is generated, including: The original operating parameter set of the target inverter is obtained through the main helium blower inverter fault detection platform. The original operating parameter set is then synchronized and noise-reduced to generate a standardized multi-dimensional time-series operating parameter stream. In the main helium blower frequency converter fault detection platform, a configurable fault mode library is constructed. The fault mode library contains a variety of typical fault condition modes that match the special operating environment of nuclear power equipment. From the standardized multi-dimensional time-series operating parameter stream, based on the preset key feature extraction strategy, a core feature subset for fault characterization and an auxiliary feature subset for environmental adaptability discrimination are separated. Based on the configurable fault mode library, the selected typical fault condition modes are injected into the core feature subset to generate a hybrid feature sequence containing simulated fault information.

[0007] Furthermore, the original set of operating parameters is synchronized and denoised to generate a standardized multi-dimensional timing operating parameter stream, including: the original set of operating parameters includes at least voltage waveform, current waveform, temperature sampling value and control signal timing; The voltage waveform and the current waveform are resampled to unify their timing resolution and their timestamps are aligned based on the hardware clock signal. The temperature sample values ​​are filtered using wavelet threshold denoising to eliminate high-frequency measurement noise. The timing sequence of the control signals is logically analyzed and state encoded to convert them into a continuous numerical sequence; All processed parameter sequences are spliced ​​and synchronized according to a unified time base to form a time-aligned, dimensionally consistent multidimensional parameter matrix; The multidimensional parameter matrix is ​​subjected to maximum and minimum normalization processing to calibrate parameters of different physical dimensions to the same numerical range, thereby generating the standardized multidimensional time-series running parameter stream.

[0008] Furthermore, the construction of a configurable failure mode library includes: Collect historical failure cases and potential failure modes of the main helium blower frequency converter in the nuclear power boilers and supporting auxiliary equipment of megawatt-class advanced pressurized water reactors, fast neutron reactors, and high-temperature gas-cooled reactors; For each fault mode, the laws governing the changes in its electrical characteristics, thermodynamic characteristics, and abnormal control logic are abstracted. The electrical characteristic variation law, thermodynamic characteristic variation law, and control logic abnormality characteristics are quantified into programmable parameterized templates and assigned a unique fault mode identifier. All parameterized templates are categorized and stored according to the device type, severity level, and probability of occurrence of the fault, forming a structured configurable fault mode library.

[0009] Furthermore, based on the configurable fault mode library, selected typical fault condition modes are injected into the core feature subset to generate a hybrid feature sequence containing simulated fault information, including: From the configurable fault mode library, select one or more parameterized templates corresponding to fault mode identifiers according to the test target; Using the core feature subset as the basic signal carrier, the feature change pattern described by the selected parameterized template is superimposed onto the data dimension corresponding to the core feature subset according to the preset intensity and time curves. During the overlay process, the data in the non-target dimensions of the core feature subset remains unchanged, and feature tampering is only performed on the dimensions where the fault is injected. The integrity of the core feature subset after the injection fault is verified to ensure its temporal continuity and physical rationality, and finally the hybrid feature sequence containing simulated fault information is generated.

[0010] Further, the hybrid feature sequence is coupled and associated with the auxiliary feature subset to construct a fault feature fusion matrix for testing, including: Calculate the statistical correlation coefficient between the mixed feature sequence and the auxiliary feature subset at different time points; Based on the statistical correlation coefficient, a feature association graph is constructed, where nodes represent features and edge weights represent correlation strength. Based on the topological structure of the feature association graph, the hybrid feature sequence and the feature dimensions with strong correlation in the auxiliary feature subset are concatenated at the data level. The spliced ​​wide-dimensional data is arranged in chronological order to form a two-dimensional matrix in which rows represent time points and columns represent the fused feature dimensions, which is the fault feature fusion matrix.

[0011] Furthermore, the adaptability verification engine performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence based on the constraints of the actual nuclear power plant operating environment simulated by the auxiliary feature subset, including: The adaptability verification engine has built-in multiple typical nuclear power operating environment models, including steady-state operating environment model, variable load operating environment model and accident transient operating environment model; The auxiliary feature subset is input into the adaptability verification engine and matched with the built-in models of various typical nuclear power operating environments to determine the type of environment simulated by the current test case. Based on the determined environment type, the corresponding environmental physical constraint rules are loaded, which define the reasonable range and associated restrictions of parameter changes; Using the loaded environmental physical constraint rules, the feature changes caused by simulated fault information in the hybrid feature sequence are verified to determine whether they are within the reasonable evolution range of the corresponding environmental type, and abnormal data points that violate the constraints are marked.

[0012] Furthermore, the adaptability verification engine outputs the fault characteristics after environment verification, and combines them with the fault tags in the configurable fault mode library to form a set of test cases with environment identification, including: For simulated fault information that passes the environmental authenticity verification, retain its original feature data; For data points that fail the verification or are marked as abnormal, feature correction or uncertainty marking is performed according to the environmental physical constraint rules. For each verified fault feature data segment, the fault mode identifier from which it originated is appended as a fault label, and the environment type determined by the adaptability verification engine is appended as an environment identifier. All verified fault feature data fragments with fault labels and environment identifiers are archived in chronological order or by fault type to form a structured set of test cases with environment identifiers.

[0013] Furthermore, the iterative training and parameter tuning of the preset fault detection algorithm using the set of test cases with environmental identifiers includes: From the set of test cases with environment identifiers, data is extracted hierarchically according to environment identifiers and fault labels, and divided into training subsets and verification subsets; Using the training subset as input, a preset initial fault detection model is trained, wherein the initial fault detection model learns the mapping relationship from fault features to fault labels during the training process; After each training iteration, the performance of the current fault detection model is evaluated using the validation subset, and its fault identification accuracy under each environmental identifier is calculated. Based on the fault identification accuracy, the internal parameters and structure of the initial fault detection model are adaptively adjusted, and the training and verification process is repeated until the model performance meets the preset convergence condition, thus obtaining the optimized fault detection model.

[0014] Furthermore, the new standardized multi-dimensional time-series operating parameter stream is analyzed online to output fault diagnosis results and confidence levels, including: The new standardized multi-dimensional time-series operating parameter stream is input into the optimized fault detection model; The optimized fault detection model extracts features from the input data stream in real time and performs matching calculations with the learned fault modes. Output the most matching fault mode type within the current time window as the fault diagnosis result; Simultaneously, the matching probability corresponding to the fault diagnosis result is output, along with the environmental adaptability factor considered in the matching probability calculation process, which together constitute the confidence level.

[0015] Compared with the prior art, the beneficial effects of the present invention are: A configurable fault mode library is constructed. Based on a preset key feature extraction strategy, the features in the processed multi-dimensional time-series operating parameter stream are separated into a core feature subset for fault characterization and an auxiliary feature subset for environmental adaptability discrimination. Then, simulated fault information is combined with the two types of features to generate a hybrid feature sequence. This technique enables feature representation to simultaneously carry the essential attributes of the fault and environmental correlation information, accurately embedding simulated fault information into the feature structure, avoiding the masking of fault signals by environmental noise, and providing a hierarchical and structured data foundation for feature fusion and subsequent verification. This allows subsequent processing to specifically focus on both the core fault and environmental adaptability requirements.

[0016] A fault feature fusion matrix is ​​constructed by coupling and associating a hybrid feature sequence with an auxiliary feature subset, and input into a pre-set adaptability verification engine. The engine calls the auxiliary feature subset to simulate the constraints of the actual nuclear power plant operating environment, performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence, and outputs the verified fault features, which are combined with fault labels to form a set of test cases with environmental labels. This technology, through auxiliary feature-driven environmental simulation, filters out simulated faults that conflict with the actual nuclear power plant environment, ensuring that the test cases reflect the fault behavior in real-world scenarios. The optimized fault detection model trained based on such test cases is more in line with the characteristics of the nuclear power plant operating environment, reduces misjudgments caused by environmental interference when analyzing new parameter streams online, and improves the reliability and confidence of diagnostic results. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a test method based on a fault detection platform for a main helium blower frequency converter, as described in this invention. Figure 2 A flowchart for generating a hybrid feature sequence containing simulated fault information; Figure 3 A flowchart for building a configurable failure mode library; Figure 4 Radar charts comparing the multi-dimensional performance of fault detection models under different operating environments; Figure 5 This is a performance graph of the fault detection model. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1The system acquires the original set of operating parameters for the target frequency converter. After a series of processing steps, this set generates a standardized multi-dimensional time-series operating parameter stream. Simultaneously, a configurable fault mode library is constructed. Based on a pre-defined key feature extraction strategy, a core feature subset for fault characterization and an auxiliary feature subset for environmental adaptability assessment are separated from the parameter stream. Based on the configurable fault mode library, simulated fault information is injected into the core feature subset, generating a hybrid feature sequence containing simulated fault information. This hybrid feature sequence is coupled with the auxiliary feature subset to construct a fault feature fusion matrix for testing. This fault feature fusion matrix is ​​input to a pre-defined adaptability verification engine. This engine simulates the constraints of the actual nuclear power plant operating environment based on the auxiliary feature subset and performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence. After verification, the adaptability verification engine outputs the environmentally verified fault features and, combined with fault tags from the configurable fault mode library, forms a set of test cases with environmental identifiers. Using this test case set, the preset fault detection algorithm is iteratively trained and its parameters are tuned to generate an optimized fault detection model adapted to the target nuclear power plant operating environment. This optimized fault detection model is then deployed on the main helium blower frequency converter fault detection platform to perform online analysis of the new standardized multi-dimensional time-series operating parameter streams and output fault diagnosis results and corresponding confidence levels.

[0020] See Figure 2In one embodiment of the present invention, the original operating parameter set of the target frequency converter is obtained through the main helium blower frequency converter fault detection platform. The original operating parameter set includes three-phase voltage waveforms, three-phase current waveforms, bearing temperature sampling values, winding temperature sampling values, and pulse width modulation control signal timing. The voltage and current waveforms are high-frequency data with 10,000 sampling points per second, the temperature sampling values ​​are low-frequency data with 10 sampling points per second, and the control signal timing is a discrete logic signal recording the switching state. The voltage and current waveforms are resampled, and the sampling rate of the voltage and current waveforms is uniformly adjusted to 1,000 points per second. The timestamps of each data point in the voltage and current waveforms are aligned according to the hardware clock signal synchronized from the programmable logic controller. Wavelet thresholding denoising is used to filter the bearing temperature sampling values ​​and winding temperature sampling values. A soft thresholding function is applied to the 'sym4' wavelet basis on the third-level decomposition detail coefficients to eliminate high-frequency measurement noise caused by sensor thermal noise. The timing sequence of the pulse width modulation control signal is logically parsed and state-encoded, with high-level states encoded as the value 1 and low-level states as the value 0. Discontinuous switching event sequences are converted into continuous numerical sequences sampled at fixed time intervals. All processed parameter sequences—namely, the resampled and aligned voltage and current waveforms, the filtered temperature samples, and the encoded control signal numerical sequences—are spliced ​​and synchronized according to a unified microsecond-level time base obtained from the programmable logic controller (PLC), forming a multidimensional parameter matrix with a completely aligned time axis and a fixed number of dimensions. The multidimensional parameter matrix is ​​then subjected to max-min normalization, calibrating the parameter values ​​of different physical dimensions—voltage, current, temperature, and control codes—to the [0,1] interval, generating a standardized multidimensional timing parameter stream.

[0021] In some embodiments, a configurable fault mode library is synchronously constructed in the main helium blower inverter fault detection platform. This library pre-stores various typical fault condition modes that match the special operating environment of nuclear power equipment, such as an inverter input anomaly mode simulating grid voltage dips or an output current distortion mode simulating power device aging. From a standardized multi-dimensional time-series operating parameter stream, a core feature subset for fault characterization and an auxiliary feature subset for environmental adaptability judgment are separated according to a preset key feature extraction strategy. The key feature extraction strategy can be defined based on prior knowledge; for example, the core feature subset extracts specific harmonic amplitudes from the current waveform and fundamental RMS values ​​from the voltage waveform, while the auxiliary feature subset extracts the temperature rise rate from temperature sampling values ​​and the switching frequency from control signals. Based on the configurable fault mode library, selected typical fault condition modes are injected into the core feature subset to generate a hybrid feature sequence containing simulated fault information.

[0022] In practice, fault injection involves direct modification of the core feature subset data. A parameterized template corresponding to the "output current distortion mode caused by power device aging" is selected from a configurable fault mode library. This template defines the curves of the third and fifth harmonic components growing over time. Using the core feature subset as the basic signal carrier, the harmonic growth curves described by the parameterized template are superimposed onto the corresponding current harmonic amplitude data dimension within the core feature subset, according to a preset intensity ratio and a set fault start time. During the superposition process, the data dimensions related to the effective value of the fundamental voltage wave within the core feature subset remain unchanged; feature tampering is only performed on the current harmonic amplitude dimension of the injected fault. The core feature subset after fault injection undergoes integrity verification, checking for interruptions in the time series and verifying whether the tampered harmonic amplitudes conform to basic power electronics principles. Finally, a hybrid feature sequence containing simulated fault information is generated.

[0023] Optionally, different parameters can be used for the synchronization and noise reduction processing stages. When resampling voltage and current waveforms, the uniform time resolution can be set to 500 points per second or 2,000 points per second according to actual analysis needs. When using wavelet thresholding for noise reduction, the 'db6' wavelet basis can be selected and thresholding can be performed on the fourth level of decomposition. It is understood that the range of the original set of operating parameters is not limited to the example described above; the original set of operating parameters can also include vibration acceleration signals and coolant pressure signals. The typical fault condition modes pre-stored in the fault mode library can also include modes simulating communication interruption of the control board or modes simulating cooling fan failure. Feature extraction strategies are not limited to methods based on prior knowledge; unsupervised feature learning and separation can also be performed using autoencoders. In specific implementations, other standardization methods can be used for the normalization of the multidimensional parameter matrix, but max-min normalization can clearly constrain the data of each dimension to the same interval. The generation of a hybrid feature sequence containing simulated fault information is the starting point for subsequent testing procedures. The hybrid feature sequence retains the time synchronization characteristics and standardized format of the original data and incorporates controllable, known-labeled fault information.

[0024] See Figure 3In one embodiment of this invention, historical fault cases and potential fault modes of the main helium blower frequency converter in the boilers and supporting auxiliary equipment of megawatt-class advanced pressurized water reactors, fast neutron reactors, and high-temperature gas-cooled reactors are collected. Historical fault cases can be obtained from power plant operation event reports, and potential fault modes can be obtained based on failure mode and effects analysis. For each fault mode, its electrical characteristic change law, thermodynamic characteristic change law, and control logic abnormality characteristic are abstracted. For example, for the "insulated gate bipolar transistor open circuit fault" mode, the abstracted electrical characteristic change law is that the output current is missing in a specific phase; the thermodynamic characteristic change law is that the junction temperature rise rate of the faulty phase module is accelerated; and the control logic abnormality characteristic is that the drive alarm signal flips. The electrical characteristic change law, thermodynamic characteristic change law, and control logic abnormality characteristic are quantified into programmable parameterized templates. The parameterized templates can be represented by mathematical functions, data sequences, or rule sets, and are assigned a unique fault mode identifier, such as "F_ID_101". All parameterized templates are categorized and stored according to the equipment type, severity level, and probability of occurrence of the fault, forming a structured configurable fault mode library. The fault mode library can use "power module fault" and "control unit fault" as equipment type directories in the database, and "minor" and "serious" as severity level labels.

[0025] In some embodiments, when generating a hybrid feature sequence containing simulated fault information based on a configurable fault mode library, the operation involves precise modification of the core feature subset data. From the configurable fault mode library, one or more parameterized templates corresponding to fault mode identifiers are selected according to the test target. For example, the parameterized template for the "DC bus capacitance attenuation" fault corresponding to the fault mode identifier "F_ID_201" is selected. The core feature subset is used as the basic signal carrier. The core feature subset, for example, includes the bus voltage ripple coefficient and the total harmonic distortion (THD) feature sequence of the input current. The feature variation law described by the selected parameterized template is superimposed onto the data dimension corresponding to the core feature subset according to a preset intensity and time curve. The parameterized template may describe the exponential growth of the bus voltage ripple amplitude over time. During the superposition process, the data in non-target dimensions of the core feature subset remains unchanged, for example, the THD data of the input current remains unchanged; feature modification is only performed on the dimension of the bus voltage ripple coefficient where the fault is injected.

[0026] In practical implementation, the mathematical expression for feature tampering can be shown in the following formula:

[0027] in: This represents the bus voltage ripple coefficient at time t after the alteration. This represents the original bus voltage ripple factor. It is a preset fault intensity coefficient. It is a fault evolution time function defined by a parameterized template. It is a unit step function. This is the fault injection start time. The integrity of the core feature subset after the fault injection is verified, the continuity of the time series is checked, and based on circuit principles, it is verified whether a reasonable physical correlation still exists between the tampered bus voltage ripple coefficient and the total harmonic distortion rate of the input current. Finally, a hybrid feature sequence containing simulated fault information is generated.

[0028] Optionally, the configurable fault mode library can be constructed using different quantization methods. When quantifying the changes in electrical characteristics, thermodynamic characteristics, and control logic anomalies into programmable parameterized templates, the parameterized templates can use lookup table interpolation rather than a single mathematical function. The encoding system for fault mode identifiers can include equipment type codes and serial numbers.

[0029] It is understood that the collected historical failure cases and potential failure modes are not limited to the aforementioned reactor type and may also include other types of nuclear energy devices. The feature change patterns described by the parameterized template can be multi-dimensional joint changes, such as simultaneously defining the coupling relationship between voltage and temperature changes. The integrity verification includes logical verification, such as verifying whether the control signal state and power output are consistent after a fault is injected. In specific implementations, the intensity and time curve of the fault injection can be set through configuration files, allowing testers to simulate different stages of fault development. Keeping the non-target dimension data in the core feature subset unchanged is to ensure the relevance of the simulated fault and avoid introducing uncontrollable interference features. The mixed feature sequence containing simulated fault information is synthetic data with clear fault labels, used for subsequent test case construction.

[0030] In one embodiment of the present invention, the statistical correlation coefficient between the mixed feature sequence and the auxiliary feature subset at different time points is calculated. The mixed feature sequence contains feature data with M dimensions, and the auxiliary feature subset contains feature data with N dimensions. For each pair of dimensions of the mixed feature sequence and the auxiliary feature subset, the Pearson correlation coefficient is calculated over the entire time series. Based on the calculated statistical correlation coefficient, a feature association graph is constructed. Nodes in the feature association graph represent M+N features, and the weight of the edge connecting any two feature nodes in the feature association graph represents the absolute value of the correlation coefficient between the two features. According to the topology of the feature association graph, the feature dimensions with strong correlation between the mixed feature sequence and the auxiliary feature subset are concatenated at the data level. For example, a correlation threshold is set, and feature dimensions connected by edges with weights exceeding the threshold are concatenated adjacently at the data level. The concatenated wide-dimensional data is arranged in chronological order to form a two-dimensional matrix, i.e., a fault feature fusion matrix, where rows represent time points and columns represent the fused feature dimensions. The number of columns in the fault feature fusion matrix is ​​the total number of feature dimensions participating in the concatenation after filtering.

[0031] In some embodiments, the adaptability verification engine performs environmental authenticity verification on simulated fault information in a mixed feature sequence based on the constraints of the actual nuclear power plant operating environment, using an auxiliary feature subset. The adaptability verification engine has built-in models of various typical nuclear power plant operating environments, including steady-state operating environment models, variable load operating environment models, and accident transient operating environment models. Each operating environment model is defined by a set of feature distribution parameters and association rules. The auxiliary feature subset is input into the adaptability verification engine and matched against the built-in models of various typical nuclear power plant operating environments to determine the environment type simulated by the current test case. The matching process can calculate the KL divergence between the feature distribution of the auxiliary feature subset and the feature distribution parameters of each environment model. Based on the determined environment type, the corresponding environmental physical constraint rules are loaded. These rules define the reasonable range and association limits for parameter changes. For example, the rules for the steady-state operating environment model limit the rate of temperature change to a very small range, while the rules for the accident transient operating environment model allow temperature and pressure to rise rapidly within defined limits.

[0032] In practical implementation, environmental physical constraint rules are used to verify the feature changes caused by simulated fault information in the mixed feature sequence. These rules include physical relationship constraints between parameters, such as the power-to-current ratio and the balance between temperature rise and heat dissipation rate. The verification process determines whether the feature changes caused by simulated fault information in the mixed feature sequence are within the reasonable evolution range of the corresponding environmental type and marks abnormal data points that violate the constraints. One verification logic can be formally expressed as checking whether the following inequalities hold:

[0033] in: The physical relationship function representing the definition of environmental physical constraint rules. This represents the mixed feature sequence vector at time t. The vector representing the auxiliary feature subset at time t. This represents the allowable deviation threshold for the rule. When the inequality does not hold, the data point corresponding to the current time point t is marked as an outlier.

[0034] Optionally, different methods can be used for constructing and fusing the feature correlation graph. When calculating the statistical correlation coefficient, the Spearman rank correlation coefficient can be used instead of the Pearson correlation coefficient. The weight of the edges in the feature correlation graph can also be defined as the square of the correlation coefficient. When constructing the fault feature fusion matrix, threshold screening can be omitted, and all features can be arranged in descending order of correlation before being concatenated. It is understood that the built-in typical nuclear power operating environment model is not limited to the three mentioned above, and can be further refined into sub-environment models such as "start-up," "shutdown," and "load shedding." The definition of environmental physical constraint rules can be derived from nuclear power system thermal-hydraulic design manuals, equipment operation procedures, or simulation models. The marking of abnormal data points can include the violated rule number and the degree of deviation.

[0035] In one embodiment of the invention, for simulated fault information that passes environmental realism verification, its original feature data is retained. This data consists of the portion of the mixed feature sequence that has been verified to conform to environmental physical constraints. For data points that fail verification or are marked as abnormal, feature correction or uncertainty marking is performed according to environmental physical constraints. Feature correction, for example, truncates temperature values ​​exceeding a reasonable range to the allowable upper limit. Adding uncertainty marking involves attaching a label indicating low confidence to the data point. For each verified fault feature data segment, a fault mode identifier from its source is attached as a fault label, and an environmental type determined by the adaptability verification engine is attached as an environmental identifier. A data segment may correspond to the fault mode identifier "F_ID_101" and the environmental identifier "steady-state operation". All verified fault feature data segments with fault labels and environmental identifiers are archived in chronological order or by fault type to form a structured test case set with environmental identifiers. The test case set with environmental identifiers can be stored in the form of a database table.

[0036] In some embodiments, a pre-defined fault detection algorithm is iteratively trained and its parameters are tuned using a set of test cases with environment labels. The training process emphasizes coverage of multiple environmental scenarios. Data is extracted hierarchically from the set of test cases with environment labels, based on environment label and fault label, and divided into training subsets and validation subsets. For example, it is ensured that each environment label ("steady-state operation"), "variable load operation," and "accident transient" and each fault label have proportionally representative samples in both the training and validation subsets. Using the training subset as input, a pre-defined initial fault detection model is trained. During training, the initial fault detection model learns the mapping relationship from fault features to fault labels. The initial fault detection model can be a convolutional neural network model. After each training iteration, the performance of the current fault detection model is evaluated using the validation subset, and its fault identification accuracy under each environment label is calculated. The fault identification accuracy is the ratio of the number of correctly classified test cases to the total number of test cases under that environment label.

[0037] In practical implementation, the weighted accuracy calculation method for the overall performance of an evaluation model under various environmental labels can be expressed as follows:

[0038] in: Represents weighted accuracy. Represents the collection of all environmental labels. Representative of environmental labeling The weighting coefficients, Representative models in environmental labeling Fault identification accuracy under [specific conditions]. Weighting coefficients. The model can be preset based on the importance or frequency of occurrence of each environment in actual operation. Based on the fault identification accuracy, the internal parameters and structure of the initial fault detection model are adaptively adjusted, and the training and verification process is repeated until the model performance meets the preset convergence conditions, resulting in an optimized fault detection model. The adaptive adjustment can be based on the gradient descent algorithm to update the model's internal parameters, and the model structure adjustment can include increasing or decreasing the number of channels in the convolutional layers.

[0039] Optionally, the structure and partitioning of the test case set with environment labels can be varied. When adding uncertainty markers, probability distributions can be used instead of binary labels to represent data confidence. When stratifying data extraction to divide the training and validation subsets, a five-fold cross-validation strategy can be used. It can be understood that the fault detection algorithm is not limited to convolutional neural networks; it can also be a long short-term memory network or a gradient boosting tree model. The mapping learned during training is a mapping from a high-dimensional fault feature space to a discrete fault label space. The convergence condition can be that the weighted accuracy no longer improves over multiple consecutive rounds, or that the preset maximum number of training rounds is reached. Refer to Table 1, which shows a simplified example of a test case set fragment with environment labels.

[0040] Table 1: Example fragments of test case sets with environment identifiers Time segment number Fault Label Environmental labeling Feature data vector (example) Uncertainty Marking TS_001 F_ID_101 steady state operation [0.12,0.85,...,0.33] none TS_002 F_ID_203 Variable load operation [0.45,0.22,...,0.78] none TS_003 F_ID_101 accident transient [0.10,0.92,...,0.30] have TS_004 F_ID_305 steady state operation [0.33,0.15,...,0.65] none In practice, the set of test cases with environmental labels provides labeled data for model training. This environmental labeling information enables the model to learn the differences in fault characteristics under different operating environments. The iterative training and parameter tuning process aims to ensure that the fault detection model not only identifies fault modes but also understands the impact of environmental context on fault characteristics. The optimized fault detection model is a diagnostic model adaptable to various nuclear power plant operating environments.

[0041] See Figure 4 This is a radar chart comparing the multi-dimensional performance of fault detection models under different operating environments. It showcases the multi-dimensional performance of the main helium blower frequency converter fault detection model under three different nuclear power plant operating environments, and is a core visualization result of the model's comprehensive evaluation phase. In steady-state operation, all indicators are significantly superior, especially fault identification accuracy and stability. The model has the strongest diagnostic capability and the most stable performance under normal operating conditions. In variable load operation, the performance is at a moderate level, with various indicators ranging from approximately 0.81 to 0.83. The model can adapt well to load change scenarios, but its performance shows a significant decline. In accident transient conditions, all indicators are the lowest, ranging from approximately 0.75 to 0.78. Under extreme operating conditions, the fault characteristics are complex, making model diagnosis the most difficult and limiting performance. In actual deployment, the model thresholds or confidence requirements can be dynamically adjusted according to the current operating environment to balance the false alarm rate and the false negative rate.

[0042] In one embodiment of the present invention, a new standardized multi-dimensional time-series operating parameter stream is input into the optimized fault detection model. This new stream originates from the real-time monitoring system of the main helium blower inverter and is generated through the same preprocessing procedure. The optimized fault detection model extracts features from the input data stream in real time. It automatically extracts deep spatiotemporal features of the input data through its internal convolutional layers or attention mechanisms and performs matching calculations with learned fault modes. The matching calculations are completed by comparing the similarity between the real-time extracted features and the various fault mode feature prototypes memorized by the model during the training phase.

[0043] In some embodiments, the most matching fault mode type within the current time window is output as the fault diagnosis result. The output layer of the optimized fault detection model is a Softmax classifier, and the category index corresponding to the maximum probability value in its output vector is mapped to the specific fault mode type, such as "F_ID_101" or "normal" category. Simultaneously, the matching probability corresponding to the output fault diagnosis result, along with the environmental adaptability factor considered in the matching probability calculation, together constitute the confidence score. The matching probability is the maximum probability value output by the Softmax classifier, and the environmental adaptability factor is a scalar calculated by the intermediate layer of the model based on the input data features, representing the consistency of the distribution of a certain type of environmental feature between the current data and the training data.

[0044] In practice, the confidence level can be calculated by combining the matching probability and the environmental adaptability factor. One calculation method is shown in the following formula:

[0045] in: This represents the overall confidence level of the final output. This represents the matching probability output by the optimized fault detection model. The environmental adaptability factor is designed with a value close to 1 to indicate high environmental adaptability. This factor, considered in the matching probability calculation, can be obtained through a dedicated environmental feature branch network within the model. This branch network is trained to predict the environmental identifier of the input data. Even when the matching probability is low, High overall confidence level It will also be adjusted accordingly.

[0046] Optionally, the time window for online analysis can be adjusted. The time window can be a fixed-length sliding window or a variable-length window based on event triggering. The optimized fault detection model extracts features from the input data stream in real time. Specifically, this can be achieved by performing a forward propagation calculation on the latest time window data at fixed time intervals. It is understood that the output format of the fault diagnosis results is not limited to a single type; the optimized fault detection model can output the top K most probable fault mode types and their probabilities. The calculation of the environmental adaptability factor can be derived from the intermediate features of the model's backbone network through statistical properties, without relying on individual branch networks.

[0047] See Figure 5 This is a performance graph of a fault detection model, showing the performance changes of the main helium blower inverter fault detection model during 50 training rounds. It is a core visualization result of the model's iterative training phase. The curve clearly shows the process of continuously increasing accuracy and continuously decreasing and stabilizing loss value, proving the effectiveness and convergence of the model training and eliminating the risk of insufficient training or divergence. Accuracy represents the proportion of fault modes correctly identified by the model on the validation set, reflecting diagnostic capability. Loss value represents the error between the model's prediction and the true label, reflecting the learning effect. Both accuracy and loss value tend to stabilize in the later stages of training, indicating that the model has effectively converged. The accuracy did not suddenly drop in the later stages, indicating that the model has good generalization ability. There are still slight fluctuations in accuracy in the later stages. An early stopping strategy can be used to stop training at about 35-40 rounds to avoid unnecessary computational overhead.

[0048] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A test method based on a fault detection platform for a main helium blower frequency converter, characterized in that, include: After processing the original set of operating parameters of the target frequency converter, a multi-dimensional time-series operating parameter stream is generated, a configurable fault mode library is constructed, and according to the preset key feature extraction strategy, the core feature subset for fault characterization and the auxiliary feature subset for environmental adaptability judgment are separated to generate a hybrid feature sequence containing simulated fault information. The hybrid feature sequence is coupled and associated with the auxiliary feature subset to construct a fault feature fusion matrix for testing; The fault feature fusion matrix is ​​input into a preset adaptability verification engine. The adaptability verification engine performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence based on the auxiliary feature subset simulating the constraints of the actual nuclear power plant operating environment. The adaptability verification engine outputs the fault characteristics after environment verification, and combines them with the fault tags in the configurable fault mode library to form a set of test cases with environment identifiers. Using the set of test cases with environmental identifiers, the preset fault detection algorithm is iteratively trained and its parameters are tuned to generate an optimized fault detection model adapted to the target nuclear power plant operating environment. The optimized fault detection model is deployed on the main helium blower frequency converter fault detection platform to perform online analysis on the new standardized multi-dimensional time-series operating parameter stream, and output fault diagnosis results and confidence levels.

2. The test method based on the main helium blower frequency converter fault detection platform as described in claim 1, characterized in that, Generate a mixed feature sequence containing simulated fault information, including: The original operating parameter set of the target inverter is obtained through the main helium blower inverter fault detection platform. The original operating parameter set is then synchronized and noise-reduced to generate a standardized multi-dimensional time-series operating parameter stream. In the main helium blower frequency converter fault detection platform, a configurable fault mode library is constructed. The fault mode library contains a variety of typical fault condition modes that match the special operating environment of nuclear power equipment. From the standardized multi-dimensional time-series operating parameter stream, based on the preset key feature extraction strategy, a core feature subset for fault characterization and an auxiliary feature subset for environmental adaptability discrimination are separated. Based on the configurable fault mode library, the selected typical fault condition modes are injected into the core feature subset to generate a hybrid feature sequence containing simulated fault information.

3. The test method based on the main helium blower frequency converter fault detection platform as described in claim 2, characterized in that, The original set of operating parameters is synchronized and denoised to generate a standardized multi-dimensional timing operating parameter stream, including: the original set of operating parameters includes at least voltage waveform, current waveform, temperature sampling value and control signal timing; The voltage waveform and the current waveform are resampled to unify their timing resolution and their timestamps are aligned based on the hardware clock signal. The temperature sample values ​​are filtered using wavelet threshold denoising to eliminate high-frequency measurement noise. The timing sequence of the control signals is logically analyzed and state encoded to convert them into a continuous numerical sequence; All processed parameter sequences are spliced ​​and synchronized according to a unified time base to form a time-aligned, dimensionally consistent multidimensional parameter matrix; The multidimensional parameter matrix is ​​subjected to maximum and minimum normalization processing to calibrate parameters of different physical dimensions to the same numerical range, thereby generating the standardized multidimensional time-series running parameter stream.

4. The test method based on the main helium blower frequency converter fault detection platform as described in claim 3, characterized in that, The construction of a configurable failure mode library includes: Collect historical failure cases and potential failure modes of the main helium blower frequency converter in the nuclear power boilers and supporting auxiliary equipment of megawatt-class advanced pressurized water reactors, fast neutron reactors, and high-temperature gas-cooled reactors; For each fault mode, the laws governing the changes in its electrical characteristics, thermodynamic characteristics, and abnormal control logic are abstracted. The electrical characteristic variation law, thermodynamic characteristic variation law, and control logic abnormality characteristics are quantified into programmable parameterized templates and assigned a unique fault mode identifier. All parameterized templates are categorized and stored according to the device type, severity level, and probability of occurrence of the fault, forming a structured configurable fault mode library.

5. The test method based on the main helium blower frequency converter fault detection platform as described in claim 4, characterized in that, Based on the configurable fault mode library, selected typical fault condition modes are injected into the core feature subset to generate a hybrid feature sequence containing simulated fault information, including: From the configurable fault mode library, select one or more parameterized templates corresponding to fault mode identifiers according to the test target; Using the core feature subset as the basic signal carrier, the feature change pattern described by the selected parameterized template is superimposed onto the data dimension corresponding to the core feature subset according to the preset intensity and time curves. During the overlay process, the data in the non-target dimensions of the core feature subset remains unchanged, and feature tampering is only performed on the dimensions where the fault is injected. The integrity of the core feature subset after the injection fault is verified to ensure its temporal continuity and physical rationality, and finally the hybrid feature sequence containing simulated fault information is generated.

6. The test method based on the main helium blower frequency converter fault detection platform as described in claim 5, characterized in that, The hybrid feature sequence is coupled and associated with the auxiliary feature subset to construct a fault feature fusion matrix for testing, including: Calculate the statistical correlation coefficient between the mixed feature sequence and the auxiliary feature subset at different time points; Based on the statistical correlation coefficient, a feature association graph is constructed, where nodes represent features and edge weights represent correlation strength. Based on the topological structure of the feature association graph, the hybrid feature sequence and the feature dimensions with strong correlation in the auxiliary feature subset are concatenated at the data level. The spliced ​​wide-dimensional data is arranged in chronological order to form a two-dimensional matrix in which rows represent time points and columns represent the fused feature dimensions, which is the fault feature fusion matrix.

7. The test method based on the main helium blower frequency converter fault detection platform as described in claim 6, characterized in that, The adaptability verification engine performs environmental authenticity verification on the simulated fault information in the hybrid feature sequence based on the constraints of the actual nuclear power plant operating environment simulated by the auxiliary feature subset, including: The adaptability verification engine has built-in multiple typical nuclear power operating environment models, including steady-state operating environment model, variable load operating environment model and accident transient operating environment model; The auxiliary feature subset is input into the adaptability verification engine and matched with the built-in models of various typical nuclear power operating environments to determine the type of environment simulated by the current test case. Based on the determined environment type, the corresponding environmental physical constraint rules are loaded, which define the reasonable range and associated restrictions of parameter changes; Using the loaded environmental physical constraint rules, the feature changes caused by simulated fault information in the hybrid feature sequence are verified to determine whether they are within the reasonable evolution range of the corresponding environmental type, and abnormal data points that violate the constraints are marked.

8. The test method based on the main helium blower frequency converter fault detection platform as described in claim 7, characterized in that, The adaptability verification engine outputs the fault characteristics after environment verification, and combines them with the fault tags in the configurable fault mode library to form a set of test cases with environment identification, including: For simulated fault information that passes the environmental authenticity verification, retain its original feature data; For data points that fail the verification or are marked as abnormal, feature correction or uncertainty marking is performed according to the environmental physical constraint rules. For each verified fault feature data segment, the fault mode identifier from which it originated is appended as a fault label, and the environment type determined by the adaptability verification engine is appended as an environment identifier. All verified fault feature data fragments with fault labels and environment identifiers are archived in chronological order or by fault type to form a structured set of test cases with environment identifiers.

9. The test method based on the main helium blower frequency converter fault detection platform as described in claim 8, characterized in that, The step of iteratively training and optimizing the preset fault detection algorithm using the set of test cases with environmental identifiers includes: From the set of test cases with environment identifiers, data is extracted hierarchically according to environment identifiers and fault labels, and divided into training subsets and verification subsets; Using the training subset as input, a preset initial fault detection model is trained, wherein the initial fault detection model learns the mapping relationship from fault features to fault labels during the training process; After each training iteration, the performance of the current fault detection model is evaluated using the validation subset, and its fault identification accuracy under each environmental identifier is calculated. Based on the fault identification accuracy, the internal parameters and structure of the initial fault detection model are adaptively adjusted, and the training and verification process is repeated until the model performance meets the preset convergence condition, thus obtaining the optimized fault detection model.

10. The test method based on the main helium blower frequency converter fault detection platform as described in claim 9, characterized in that, Online analysis is performed on the new standardized multi-dimensional time-series operating parameter stream to output fault diagnosis results and confidence levels, including: The new standardized multi-dimensional time-series operating parameter stream is input into the optimized fault detection model; The optimized fault detection model extracts features from the input data stream in real time and performs matching calculations with the learned fault modes. Output the most matching fault mode type within the current time window as the fault diagnosis result; Simultaneously, the matching probability corresponding to the fault diagnosis result is output, along with the environmental adaptability factor considered in the matching probability calculation process, which together constitute the confidence level.

Images