Artificial intelligence-based fault detection method and device for high-power pulse sources

Abstract

The application discloses a fault detection method and device of a high-power pulse source based on artificial intelligence, and relates to the technical field of fault detection. The method comprises the following steps: in a cloud server, a multi-target swarm intelligent optimization algorithm is used to construct a fault detection model, and a model parameter self-adaptive adjuster is configured; a multi-channel synchronous acquisition system is used to acquire original monitoring signals of the high-power pulse source, to perform denoising and waveform self-adaptive alignment, to obtain standard monitoring data, and to upload the standard monitoring data to the cloud server; in the cloud server, running state data is acquired, and a model parameter self-adaptive adjuster is used to adjust model parameters, to obtain an optimal fault detection model; and the standard monitoring data is input into the optimal fault detection model to perform fault detection, to obtain a fault detection result of the high-power pulse source. The application solves the problems of weak anti-interference ability, inability to balance detection accuracy and reasoning speed, and difficulty in adapting to a cloud dynamic load environment in the prior art.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a fault detection method and apparatus for a high-power pulse source based on artificial intelligence. Background Technology

[0002] High-power pulse sources (such as pulse generators and Marx generators) have wide applications in nuclear physics, fusion research, and electromagnetic launch. These devices typically operate at extremely high voltage, current, and power densities, and their operating environments are often accompanied by strong electromagnetic interference.

[0003] Traditional fault detection methods primarily rely on manually set thresholds or simple physical indicator monitoring. However, high-power pulse sources operate under complex conditions, exhibiting diverse fault modes that are often latent and coupled. Traditional single-threshold methods struggle to capture the deep nonlinear characteristics of signals and are susceptible to electromagnetic noise interference, leading to false alarms or missed alarms. Furthermore, while deep learning-based fault detection methods improve accuracy to some extent, they often suffer from complex and fixed model structures, making it difficult to balance detection accuracy with inference speed. In real-world cloud deployments, when concurrent requests surge or computing resources are strained, complex models with fixed structures can cause severe latency, failing to meet real-time requirements.

[0004] Therefore, there is an urgent need for a high-power pulse source fault detection method that can adaptively adjust model parameters, integrate physical prior knowledge and deep features, and operate stably in high-interference environments. Summary of the Invention

[0005] This invention provides a fault detection method and device for high-power pulse sources based on artificial intelligence. This invention solves the problems of weak anti-interference ability, inability to balance detection accuracy and inference speed, and difficulty in adapting to dynamic load environments in the cloud, which exist in the prior art.

[0006] In a first aspect, embodiments of the present invention provide a fault detection method for a high-power pulse source based on artificial intelligence, the method comprising:

[0007] On a cloud server, a fault detection model based on dual-stream feature fusion and attention mechanism is constructed using a multi-objective swarm intelligent optimization algorithm, and an adaptive adjuster for model parameters is configured.

[0008] A multi-channel synchronous acquisition system is used to acquire the raw monitoring signal of a high-power pulse source. The raw monitoring signal is then denoised and the waveform is adaptively aligned to obtain standard monitoring data, which is then uploaded to a cloud server.

[0009] On the cloud server, operational status data is collected, and based on the operational status data, the model parameters of the fault detection model are adjusted using a model parameter adaptive adjuster to obtain the optimal fault detection model.

[0010] Standard monitoring data is input into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

[0011] The technical solution provided in this application has at least the following beneficial effects:

[0012] By fusing dual-stream features in the fault detection model, combining the explicit physical meaning of physical parameters with the latent features of deep waveforms, and using an anti-interference attention mechanism to suppress noise, the accuracy of fault detection is significantly improved. A multi-objective swarm intelligent optimization algorithm is used to pre-optimize model parameters (high precision, balanced, and ultra-fast) for different application scenarios, and dynamically switches them based on the real-time running status of the cloud server, resolving the contradiction between model accuracy and inference speed and ensuring the model's real-time response capability under high concurrency. Variational mode decomposition and cross-correlation alignment techniques are employed to effectively remove strong electromagnetic interference and eliminate time deviations caused by trigger jitter, providing high-quality input data for the model.

[0013] In one alternative implementation, the fault detection model includes a physical parameter stream extraction module, a deep waveform stream extraction module, an anti-interference attention weighting module, and a fault metric classification module.

[0014] In one alternative implementation, a fault detection model based on dual-stream feature fusion and attention mechanism is constructed on a cloud server using a multi-objective swarm intelligent optimization algorithm, and an adaptive adjuster for model parameters is configured, including:

[0015] On a cloud server, a fault detection model is constructed based on dual-stream feature fusion and attention mechanism, and the set of parameters to be optimized for the fault detection model and the multi-objective optimization function of the multi-objective swarm intelligent optimization algorithm are determined.

[0016] Based on the training sample set in the historical database of the cloud server, and based on the multi-objective optimization function, the multi-objective swarm intelligent optimization algorithm is used to optimize the set of parameters to be optimized, and the Pareto non-dominated optimal solution set is obtained.

[0017] Based on the optimization objective value, the application scenarios of the model parameters of the fault detection model corresponding to each solution in the Pareto non-dominated optimal solution set are divided into high-precision, balanced, and ultra-fast types.

[0018] The model parameters of the fault detection model corresponding to each application scenario are packaged into Docker containers, all Docker containers are deployed to the model parameter adaptive adjuster, and the model parameter adaptive adjuster is connected to the fault detection model.

[0019] In one alternative implementation, the multi-objective optimization function includes maximizing the fault classification accuracy function, minimizing the model computational complexity function, and minimizing the model runtime function.

[0020] In one alternative implementation, based on the training sample set in the historical database of the cloud server, and using a multi-objective swarm intelligent optimization algorithm based on a multi-objective optimization function, the set of parameters to be optimized is optimized to obtain the Pareto non-dominated optimal solution set, including:

[0021] The set of parameters to be optimized is encoded into individual vectors of the multi-objective Osprey optimization algorithm, and the multi-objective Osprey optimization algorithm is initialized.

[0022] The chaotic sequence is generated using the Logistic mapping, and then mapped to the solution space of individual ospreys to obtain the initial osprey population.

[0023] Based on the training sample set in the historical database of the cloud server, according to the set of candidate parameters to be optimized for each individual osprey in the initial osprey population, the corresponding candidate fault detection model is quickly trained, and the corresponding optimization target value is calculated using a multi-objective optimization function.

[0024] Based on the optimization objective value, the initial osprey population is subjected to fast non-dominated sorting to obtain different frontier levels, and an external archive set is maintained to store all non-dominated solutions of the first frontier level.

[0025] By introducing the Levy flight mechanism and the gray wolf cooperation mechanism, and using the hovering search behavior and diving capture behavior of the multi-target osprey optimization algorithm, the initial osprey population is updated to obtain an updated osprey population.

[0026] Using a multi-objective optimization function, the optimization objective value corresponding to each updated osprey in the updated osprey population is calculated, and the frontier level and external archive set are updated according to the optimization objective value.

[0027] The iteration continues until the iteration count reaches the threshold, at which point the iterative update of the osprey population stops, and all non-dominated solutions in the external archive set are output as the Pareto non-dominated optimal solution set.

[0028] In one alternative implementation, the Levy flight mechanism and the gray wolf cooperative mechanism are introduced. The initial osprey population is updated using the hovering search behavior and dive-capture behavior of the multi-target osprey optimization algorithm, resulting in an updated osprey population, including:

[0029] Based on the gray wolf collaboration mechanism and crowding, the top three local optima from the external archives were identified and defined as Alpha individuals, Beta individuals, and Delta individuals, respectively. The optimal Osprey individual was then expanded to include Alpha individuals, Beta individuals, and Delta individuals.

[0030] Alpha, Beta, and Delta individuals are introduced. Based on the circling search behavior of the Osprey algorithm, the initial Osprey individuals in the initial Osprey population are updated to obtain the updated first Osprey individual.

[0031] The formula is:

[0032]

[0033] In the formula, Number of iterations t+ The first result obtained from the hovering search behavior is the first i The first updated Osprey individual; The convergence factor; Number of iterations t The osprey individual, the initial osprey individual during the first iteration update; Number of iterations t Alpha, Beta, and Delta individuals; Individual ospreys were selected randomly. Flight radius; To explore weights; Prioritize development; t This represents the current iteration number;

[0034]

[0035] In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. , To adjust the parameters; It is the hyperbolic tangent function;

[0036] The Levy flight mechanism is introduced, and the initial Osprey individuals in the initial Osprey population are updated based on the swooping capture behavior of the Osprey algorithm to obtain the updated second Osprey individuals.

[0037] The formula is:

[0038]

[0039] In the formula, Number of iterationst+ The first dive-and-capture maneuver yielded the [number missing]th [missing information]. i A newer second Osprey individual; For the Levy flight mechanism Levy Distributed random numbers; b for Levy Step length, and b ∈[1,2]; As the second development weight;

[0040] By integrating several updated first-generation osprey individuals and several updated second-generation osprey individuals, an updated osprey population is obtained.

[0041] In one alternative implementation, a multi-channel synchronous acquisition system is used to acquire the raw monitoring signal from the high-power pulse source. The raw monitoring signal is then denoised and adaptively aligned to obtain standard monitoring data, which is then uploaded to a cloud server, including:

[0042] A multi-channel synchronous acquisition system is used to acquire the raw monitoring signals of the high-power pulse source. The raw monitoring signals include high-frequency voltage waveforms, current waveforms, drain current of switching devices, and temperature data.

[0043] The variational mode decomposition algorithm is used to perform strong electromagnetic interference denoising on the high-frequency voltage waveform, current waveform and drain current of switching device in the original monitoring signal, and to obtain the denoised high-frequency voltage waveform, denoised current waveform and denoised drain current of switching device.

[0044] The moving average filtering algorithm is used to remove random jitter interference from the temperature data in the original monitoring signal to obtain the denoised temperature data.

[0045] A standard high-frequency voltage waveform is selected from the historical normal data in the historical database of the cloud server as a time reference template, and the cross-correlation function between the denoised high-frequency voltage waveform and the time reference template is calculated to find the time offset corresponding to the peak value of the cross-correlation function.

[0046] Based on the time offset, the denoised high-frequency voltage waveform, the denoised current waveform, and the denoised drain current of the switching device are adaptively aligned to obtain the aligned high-frequency voltage waveform, the aligned current waveform, and the aligned drain current of the switching device.

[0047] The aligned high-frequency voltage waveform, aligned current waveform, aligned drain current of switching devices, and denoised temperature data are normalized to obtain standard monitoring data, which are then uploaded to the cloud server.

[0048] In one alternative implementation, runtime status data is collected on a cloud server, and based on this data, an adaptive model parameter adjuster is used to adjust the model parameters of the fault detection model to obtain the optimal fault detection model, including:

[0049] On the cloud server, operational status data is collected, including the cloud server's CPU utilization. and concurrent request queue length ;

[0050] If CPU utilization CPU utilization threshold and / or concurrent request queue length Length threshold Then, the model parameters corresponding to the high-speed type are retrieved from the model parameter adaptive adjuster, the fault detection model is adjusted, and the optimal fault detection model is obtained.

[0051] If CPU utilization CPU utilization threshold and concurrent request queue length Length threshold Then, the model parameters corresponding to the high-precision type are retrieved from the model parameter adaptive adjuster, the fault detection model is adjusted, and the optimal fault detection model is obtained.

[0052] If the above rules are not met, the model parameters corresponding to the balanced type are retrieved from the model parameter adaptive adjuster to adjust the fault detection model and obtain the optimal fault detection model.

[0053] In one alternative implementation, standard monitoring data is input into the optimal fault detection model for fault detection, and the fault detection results of the high-power pulse source include:

[0054] Input standard monitoring data into the optimal fault detection model;

[0055] The physical parameter flow extraction module of the optimal fault detection model calculates the physical parameter flow feature vector of the standard monitoring data based on the preset physical sensitivity index.

[0056] The deep waveform stream extraction module of the optimal fault detection model extracts deep spatiotemporal features, namely deep waveform stream feature vectors, from standard monitoring data based on convolutional neural networks.

[0057] Using the anti-interference attention weighting module of the optimal fault detection model, feature weights of the physical parameter flow feature vector are generated based on the time position index. Based on the feature weights, the physical parameter flow feature vector is differentially weighted to obtain the weighted physical parameter flow feature vector.

[0058] The fault metric classification module of the optimal fault detection model concatenates and fuses the weighted physical parameter flow feature vector and the depth waveform flow feature vector to obtain fused features. Based on metric learning, the distance between the fused features and the center of various fault prototypes is calculated, and the fault detection results are generated based on the distance and the set confidence threshold.

[0059] Secondly, embodiments of the present invention provide a fault detection device for a high-power pulse source based on artificial intelligence, used to implement a fault detection method for a high-power pulse source, the device comprising:

[0060] The model building unit is used on a cloud server to build a fault detection model based on dual-stream feature fusion and attention mechanism using a multi-objective swarm intelligent optimization algorithm, and to configure an adaptive adjuster for model parameters.

[0061] The data acquisition unit is used to acquire the raw monitoring signal of the high-power pulse source using a multi-channel synchronous acquisition system, perform noise reduction and waveform adaptive alignment on the raw monitoring signal to obtain standard monitoring data, and upload it to the cloud server.

[0062] The model adjustment unit is used to collect operational status data on the cloud server and, based on the operational status data, use the model parameter adaptive adjuster to adjust the model parameters of the fault detection model to obtain the optimal fault detection model.

[0063] The fault detection unit is used to input standard monitoring data into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

[0064] A third aspect of this invention provides an electronic device, which includes:

[0065] At least one processor; and a memory communicatively connected to the at least one processor; wherein,

[0066] The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0067] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention;

[0069] Figure 2 This is a flowchart of the steps of a fault detection method for a high-power pulse source based on artificial intelligence provided in an embodiment of the present invention;

[0070] Figure 3 This is a functional unit diagram of a fault detection device for a high-power pulse source based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0071] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0072] The present invention will be further described below with reference to the accompanying drawings.

[0073] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0074] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0075] Those skilled in the art will understand that Figure 1The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0076] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and an electronic program for a fault detection device based on an artificial intelligence high-power pulse source.

[0077] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the electronic program of the fault detection device of the high power pulse source based on artificial intelligence stored in the memory 1005 through the processor 1001, and executes the fault detection method of the high power pulse source based on artificial intelligence provided in the embodiment of the present invention.

[0078] Reference Figure 2 The present invention provides a fault detection method for a high-power pulse source based on artificial intelligence, the method comprising:

[0079] S201: On a cloud server, a fault detection model based on dual-stream feature fusion and attention mechanism is constructed using a multi-objective swarm intelligent optimization algorithm, and an adaptive adjuster for model parameters is configured.

[0080] S202: Using a multi-channel synchronous acquisition system, the raw monitoring signal of the high-power pulse source is acquired, the raw monitoring signal is denoised and the waveform is adaptively aligned to obtain standard monitoring data, and then uploaded to the cloud server;

[0081] S203: On the cloud server, collect operational status data, and based on the operational status data, use the model parameter adaptive adjuster to adjust the model parameters of the fault detection model to obtain the optimal fault detection model;

[0082] S204: Input the standard monitoring data into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

[0083] The technical solution provided in this application has at least the following beneficial effects:

[0084] By fusing dual-stream features in the fault detection model, combining the explicit physical meaning of physical parameters with the latent features of deep waveforms, and using an anti-interference attention mechanism to suppress noise, the accuracy of fault detection is significantly improved. A multi-objective swarm intelligent optimization algorithm is used to pre-optimize model parameters (high precision, balanced, and ultra-fast) for different application scenarios, and dynamically switches them based on the real-time running status of the cloud server, resolving the contradiction between model accuracy and inference speed and ensuring the model's real-time response capability under high concurrency. Variational mode decomposition and cross-correlation alignment techniques are employed to effectively remove strong electromagnetic interference and eliminate time deviations caused by trigger jitter, providing high-quality input data for the model.

[0085] In one optional implementation, the fault detection model includes a physical parameter stream extraction module, a deep waveform stream extraction module, an anti-interference attention weighting module, and a fault metric classification module.

[0086] The physical parameter stream extraction module is used to calculate the physical parameter stream feature vector of standard monitoring data based on preset physical sensitivity indicators;

[0087] The deep waveform stream extraction module is used to extract deep spatiotemporal features, i.e., deep waveform stream feature vectors, from standard monitoring data based on convolutional neural networks;

[0088] The anti-interference attention weighting module is used to generate feature weights for the physical parameter flow feature vector based on the time position index, and to perform differential weighting on the physical parameter flow feature vector based on the feature weights to obtain the weighted physical parameter flow feature vector.

[0089] The fault measurement and classification module is used to concatenate and fuse the weighted physical parameter flow feature vector and the depth waveform flow feature vector to obtain fused features. Based on metric learning, the distance between the fused features and the center of various fault prototypes is calculated, and fault detection results are generated based on the distance and the set confidence threshold.

[0090] In one alternative implementation, a fault detection model based on dual-stream feature fusion and attention mechanism is constructed on a cloud server using a multi-objective swarm intelligent optimization algorithm, and an adaptive adjuster for model parameters is configured, including:

[0091] S2011: On a cloud server, based on dual-stream feature fusion and attention mechanism, a fault detection model is constructed, and the set of parameters to be optimized for the fault detection model and the multi-objective optimization function of the multi-objective swarm intelligent optimization algorithm are determined.

[0092] In this embodiment, the set of parameters to be optimized includes the number of convolutional layers, the size of the convolutional kernel, the number of neurons in the fully connected layer, and the initial learning rate;

[0093] The search space for the number of convolutional layers is [2, 3, 4, 5, 6];

[0094] The search space for the kernel size is [3, 5, 7, 9];

[0095] The search space for the number of neurons in the fully connected layer is [64, 128, 256, 512].

[0096] The search space for the initial learning rate is

[10] . −4 10 −3 5×10 −3 ];

[0097] S2012: Based on the training sample set in the historical database of the cloud server, and based on the multi-objective optimization function, the multi-objective swarm intelligent optimization algorithm is used to optimize the set of parameters to be optimized, and the Pareto non-dominated optimal solution set is obtained.

[0098] S2013: Based on the optimization objective value, the application scenarios of the model parameters of the fault detection model corresponding to each solution in the Pareto non-dominated optimal solution set are divided into high-precision type, balanced type and ultra-fast type.

[0099] In this embodiment, the model types are classified as follows: Based on the objective function value of each solution in the Pareto optimal solution set, they are divided into three categories:

[0100] High-precision type: highest accuracy, but with higher computational complexity and time consumption (suitable for low-load periods).

[0101] Balanced type: All indicators are in the middle (suitable for normal loads);

[0102] Ultra-fast: Shortest processing time, lowest complexity, slight sacrifice in accuracy (suitable for high concurrency or high load periods).

[0103] S2014: Package the model parameters of the fault detection model corresponding to each application scenario into a Docker container, deploy all Docker containers to the model parameter adaptive adjuster, and connect the model parameter adaptive adjuster to the fault detection model.

[0104] In this embodiment, a Docker image is built for the complete model file (weight file + structure configuration) corresponding to each type of representative solution (or all solutions), and the runtime environment is pre-configured in the container;

[0105] Deploy an adaptive model parameter adjuster (based on Kubernetes or a microservice architecture) on a cloud server. This adjuster loads all the aforementioned Docker containers and listens for status signals from the cloud server, ready to switch models at any time.

[0106] In one alternative implementation, the multi-objective optimization function includes a function to maximize fault classification accuracy, a function to minimize model computational complexity, and a function to minimize model runtime.

[0107] The formula for maximizing fault classification accuracy:

[0108]

[0109] In the formula, For individual ospreys The corresponding fault detection model's fault classification accuracy on the validation set; The total number of validation sets; For the first j Weights of class samples; j For sample indication; For indicator functions; For the first j Predicted labels and true labels for class samples;

[0110] The formula for minimizing the computational complexity function of the model:

[0111]

[0112] In the formula, For individual ospreys The computational complexity of the corresponding fault detection model; For the first l The kernel size, height, and width of the feature maps in each convolutional layer; For the first l The number of feature maps output by each convolutional layer, i.e., the number of output channels; For the first l The number of feature maps input to each convolutional layer, i.e., the number of input channels; l This is an indicator for the convolutional layer; This represents the total number of convolutional layers. This represents the total number of fully connected layers. m This is an indicator for the fully connected layer; For the first m The number of output neurons in a fully connected layer; For the first m The number of input neurons in a fully connected layer;

[0113] The formula for minimizing the model's runtime function:

[0114]

[0115] In the formula, For individual ospreys The running time of the corresponding fault detection model; The average time it takes for the model to perform one inference operation on a cloud server; To find the minimum value, to prevent the denominator from being zero;

[0116] The optimization objective value obtained by the multi-objective optimization function is .

[0117] In one alternative implementation, based on the training sample set in the historical database of the cloud server, and using a multi-objective swarm intelligent optimization algorithm based on a multi-objective optimization function, the set of parameters to be optimized is optimized to obtain the Pareto non-dominated optimal solution set, including:

[0118] S20121: Encode the set of parameters to be optimized into individual vectors of the multi-objective Osprey optimization algorithm, and initialize the multi-objective Osprey optimization algorithm;

[0119] S20122: Use Logistic mapping to generate chaotic sequences and map the chaotic sequences to the solution space of individual ospreys to obtain the initial osprey population;

[0120] S20123: Based on the training sample set in the historical database of the cloud server, according to the set of candidate parameters to be optimized for each individual osprey in the initial osprey population, quickly train the corresponding candidate fault detection model, and use a multi-objective optimization function to calculate the corresponding optimization target value.

[0121] S20124: Based on the optimization objective value, perform fast non-dominated sorting on the initial osprey population to obtain different frontier levels, and maintain an external archive set to store all non-dominated solutions of the first frontier level;

[0122] S20125: Introducing the Levy flight mechanism and the gray wolf cooperation mechanism, using the hovering search behavior and diving capture behavior of the multi-target osprey optimization algorithm to update the initial osprey population and obtain the updated osprey population.

[0123] S20126: Using a multi-objective optimization function, calculate the optimization objective value corresponding to each updated osprey in the updated osprey population, and update the frontier level and external archive set according to the optimization objective value;

[0124] S20127: Stop iterative updates of the osprey population until the number of iterations reaches the iteration threshold, and output all non-dominated solutions in the external archive set as the Pareto non-dominated optimal solution set.

[0125] In one alternative implementation, the Levy flight mechanism and the gray wolf cooperative mechanism are introduced. The initial osprey population is updated using the hovering search behavior and dive-capture behavior of the multi-target osprey optimization algorithm, resulting in an updated osprey population, including:

[0126] S201251: From the external archive set, based on the gray wolf cooperation mechanism, the top three local optimal solutions are identified according to the crowding degree and defined as Alpha individuals, Beta individuals and Delta individuals respectively. The optimal Osprey individual is then expanded to Alpha individuals, Beta individuals and Delta individuals.

[0127] S201252: Introduce Alpha, Beta, and Delta individuals. Based on the circling search behavior of the Osprey algorithm, update the initial Osprey individuals in the initial Osprey population to obtain the updated first Osprey individual.

[0128] The formula is:

[0129]

[0130] In the formula, Number of iterations t+ The first result obtained from the hovering search behavior is the first i The first updated Osprey individual; The convergence factor; Number of iterations t The osprey individual, the initial osprey individual during the first iteration update; Number of iterations t Alpha, Beta, and Delta individuals; Individual ospreys were selected randomly. Flight radius; To explore weights; Prioritize development; t This represents the current iteration number;

[0131]

[0132] In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. , To adjust the parameters; It is the hyperbolic tangent function;

[0133] S201253: Introducing the Levy flight mechanism, based on the osprey algorithm's dive-and-capture behavior, the initial osprey individuals in the initial osprey population are updated to obtain the updated second osprey individuals;

[0134] The formula is:

[0135]

[0136] In the formula, Number of iterations t+ The first dive-and-capture maneuver yielded the [number missing]th [missing information]. i A newer second Osprey individual; For the Levy flight mechanism Levy Distribute random numbers; b for Levy Step length, and b ∈[1,2]; As the second development weight;

[0137] S201254: Integrate several updated first osprey individuals and several updated second osprey individuals to obtain an updated osprey population.

[0138] In one alternative implementation, a multi-channel synchronous acquisition system is used to acquire the raw monitoring signal from the high-power pulse source. The raw monitoring signal is then denoised and adaptively aligned to obtain standard monitoring data, which is then uploaded to a cloud server, including:

[0139] S2021: Use a multi-channel synchronous acquisition system to acquire the raw monitoring signals of the high-power pulse source. The raw monitoring signals include high-frequency voltage waveforms, current waveforms, drain current of switching devices, and temperature data.

[0140] In this embodiment, the multi-channel synchronous acquisition system uses a high-voltage differential probe (for measuring voltage) with a bandwidth of not less than 200MHz and a sampling rate of not less than 100MS / s and a high-frequency Rogowski coil (for measuring current); at the same time, high-precision thermocouples or fiber optic temperature sensors are used to acquire the case temperature and heat sink temperature of key switching devices (such as IGBTs and MOSFETs).

[0141] FPGA is used to control a multi-channel analog-to-digital converter to ensure that the sampling clocks of all channels are strictly synchronized and the clock jitter is controlled in the nanosecond level (e.g., <5ns) to eliminate time deviation between channels.

[0142] S2022: Using variational mode decomposition algorithm, strong electromagnetic interference denoising is performed on the high-frequency voltage waveform, current waveform and drain current of switching device in the original monitoring signal to obtain the denoised high-frequency voltage waveform, denoised current waveform and denoised drain current of switching device.

[0143] In this embodiment, for high-frequency voltage waveforms, current waveforms, and drain currents of switching devices, since they contain strong electromagnetic interference, variational mode decomposition algorithm is used to filter out high-frequency white noise and electromagnetic interference in specific frequency bands, while retaining the transient edge features of the waveforms.

[0144] Specific steps: Set the number of modes K =4, decompose the noisy signal into KEach intrinsic mode function (IMF) is used to calculate the center frequency of each IMF, identify and remove high-frequency noise components whose center frequency is higher than the effective signal bandwidth, and reconstruct the denoised waveform from the remaining components.

[0145] S2023: Use the moving average filtering algorithm to remove random jitter interference from the temperature data in the original monitoring signal to obtain the denoised temperature data.

[0146] In this embodiment, random jitter noise during the sensor acquisition process is filtered out, while the true temperature change trend is preserved;

[0147] S2024: Select a standard high-frequency voltage waveform as a time reference template from the historical normal data in the historical database of the cloud server, calculate the cross-correlation function between the denoised high-frequency voltage waveform and the time reference template, and find the time offset corresponding to the peak value of the cross-correlation function.

[0148] In this embodiment, due to the slight time jitter in the triggering system, the electrical signals acquired each time are relatively offset on the time axis, so it is necessary to align the fast-changing electrical signals (high-frequency voltage waveform, load current waveform, and drain current of switching devices).

[0149] S2025: Based on the time offset, the denoised high-frequency voltage waveform, the denoised current waveform, and the denoised drain current of the switching device are adaptively aligned to obtain the aligned high-frequency voltage waveform, the aligned current waveform, and the aligned drain current of the switching device.

[0150] In this embodiment, based on the calculated offset, the denoised high-frequency voltage waveform, the denoised current waveform, and the denoised drain current of the switching device are shifted (leading or lagging) on ​​the time axis so that the triggering start point of the waveform is strictly aligned with the reference template.

[0151] The denoised temperature data has a large thermal inertia and an extremely low frequency of change. Its characteristics are unrelated to the microsecond-level jitter at the trigger moment. Therefore, this alignment step is skipped, and its original timestamp remains unchanged.

[0152] S2026: Normalize the aligned high-frequency voltage waveform, aligned current waveform, aligned drain current of switching devices, and denoised temperature data to obtain standard monitoring data, and upload them to the cloud server.

[0153] In this embodiment, the aligned electrical signal data and the filtered temperature data are uniformly mapped to the [0,1] interval or Z-score standardized to obtain standard monitoring data with uniform length and amplitude, so as to be input into the subsequent fault diagnosis model.

[0154] In one alternative implementation, runtime status data is collected on a cloud server, and based on this data, an adaptive model parameter adjuster is used to adjust the model parameters of the fault detection model to obtain the optimal fault detection model, including:

[0155] S2031: On the cloud server, collect operational status data, including the CPU utilization of the cloud server. and concurrent request queue length ;

[0156] In this embodiment, CPU utilization Reflects the workload of computing resources; length of the concurrent request queue. Reflects the number of pending fault detection tasks currently in the pipeline;

[0157] S2032: If CPU utilization... CPU utilization threshold (e.g., 80%) and / or the length of the concurrent request queue Length threshold (For example, for 100 tasks), the corresponding model parameters for the high-speed type are retrieved from the model parameter adaptive adjuster, the fault detection model is adjusted, and the optimal fault detection model is obtained.

[0158] In this embodiment, the risk of processing delay is high. At this time, the model parameter adaptive adjuster sends an instruction to the fault detection service to switch the currently running Docker container to an "ultra-fast" model container. The ultra-fast model has fewer parameters and a simpler structure, which can quickly digest the backlog of tasks and ensure response speed.

[0159] S2033: If CPU utilization... CPU utilization threshold and concurrent request queue length Length threshold If this state persists for a period of time (e.g., more than 5 minutes), then the high-precision model parameters corresponding to the model parameter adaptive adjuster are retrieved from the model parameter adaptive adjuster to adjust the fault detection model and obtain the optimal fault detection model.

[0160] In this embodiment, the ultimate accuracy of fault detection is pursued. The model parameter adaptive adjuster sends instructions to smoothly switch to the "high-precision" model container. Although the model has a large amount of computation, it will not cause congestion under low load and can provide the most reliable diagnostic results.

[0161] S2034: If the above rules are not met, retrieve the corresponding model parameters of the balanced type from the model parameter adaptive adjuster, adjust the fault detection model, and obtain the optimal fault detection model.

[0162] If neither of the above two extreme conditions is met (i.e., the load is in a state of moderate fluctuation, or during the transition period between high and low loads), then maintain or switch to the "balanced" model container, which strikes a trade-off between speed and accuracy and can adapt to most normal operating conditions.

[0163] In one alternative implementation, standard monitoring data is input into the optimal fault detection model for fault detection, and the fault detection results of the high-power pulse source include:

[0164] S2041: Input standard monitoring data into the optimal fault detection model;

[0165] S2042: The physical parameter flow extraction module using the optimal fault detection model calculates the physical parameter flow feature vector of the standard monitoring data based on the preset physical sensitivity index.

[0166] In this embodiment, the physical sensitivity indicators include:

[0167] For the normalized high-frequency voltage waveform, the following indicators are extracted:

[0168] steepness of rising edge Calculate the time difference required for the pulse to rise from 10% peak value to 90% peak value. The formula for steepness is:

[0169]

[0170] In the formula, The steepness of the rising edge; The time difference it takes for the pulse to rise from 10% peak value to 90% peak value; The peak value of the normalized high-frequency voltage waveform;

[0171] Flat-top ripple coefficient Extract the peak region of the pulse (e.g., amplitude greater than 90%). The ratio of the root mean square (RMS) value of the voltage in the specified range to its DC average component is calculated to reflect the charging and discharging stability of the energy storage capacitor.

[0172] Reverse overshoot peak : Detects the extreme value of the negative voltage after the falling edge of the pulse, used to evaluate whether the inductive load backlash protection of the circuit is effective;

[0173] Based on the normalized load current waveform and the normalized drain current of the switching device, the following parameters are extracted:

[0174] Peak current Extract the maximum amplitude of the load current and drain current respectively, and the difference between the two (drain current - load current) is used to determine whether there is an internal short circuit or bypass fault.

[0175] Current rise rate Calculating the slope of the initial segment of the current waveform is crucial. An excessively steep slope can damage switching devices due to excessive stress, making it an important health indicator. For time indication;

[0176] Pulse width consistency : Calculate the duration of the current waveform at half maximum width and height and compare it with a standard value to detect faults such as core saturation;

[0177] For the temperature data, the following indicators were extracted:

[0178] Temperature rise rate Calculate the difference between the current temperature value and the previous temperature value to reflect sudden increases in load or abnormal heat dissipation;

[0179] absolute temperature value The current actual temperature value;

[0180] Feature vector generation:

[0181] By combining all the physical parameter flow indicators extracted from the standard monitoring data, a high-dimensional physical feature vector is generated: [ , , , , , , , ];

[0182] S2043: The deep waveform stream extraction module using the optimal fault detection model, based on a convolutional neural network, extracts deep spatiotemporal features of standard monitoring data, namely deep waveform stream feature vectors.

[0183] In this embodiment, the standard monitoring data passes through multiple convolutional layers, activation layers (ReLU), and pooling layers in sequence. The convolutional layers use convolutional kernels of different scales to slide on the waveform and automatically extract features from low-level edge features to high-level fault mode features (such as specific oscillation patterns and distortion morphologies).

[0184] S2044: The anti-interference attention weighting module of the optimal fault detection model generates feature weights for the physical parameter flow feature vector based on the time position index, and performs differential weighting on the physical parameter flow feature vector based on the feature weights to obtain the weighted physical parameter flow feature vector.

[0185] In this embodiment, the anti-interference attention weighting module receives the physical parameter flow feature vector, combines it with the time position index, and generates a set of weight vectors through a fully connected layer and a Sigmoid activation function. This mechanism is used to identify which features are most important for fault judgment at the current moment. The generated weights are multiplied element-wise with the physical parameter flow feature vector to obtain the weighted physical parameter flow feature vector. This step suppresses noise features and amplifies sensitive features.

[0186] S2045: The fault metric classification module using the optimal fault detection model concatenates and fuses the weighted physical parameter flow feature vector and the depth waveform flow feature vector to obtain fused features. Based on metric learning, it calculates the distance between the fused features and the centers of various fault prototypes, and generates fault detection results based on the distance and the set confidence threshold.

[0187] This invention also provides a fault detection device for a high-power pulse source based on artificial intelligence, referring to... Figure 3 The diagram shows a functional unit diagram of a fault detection device 300 for a high-power pulse source based on artificial intelligence according to the present invention. The device may include the following units:

[0188] Model building unit 301 is used to build a fault detection model based on dual-stream feature fusion and attention mechanism on a cloud server using a multi-objective swarm intelligent optimization algorithm, and to configure an adaptive adjuster for model parameters.

[0189] The data acquisition unit 302 is used to acquire the raw monitoring signal of the high-power pulse source using a multi-channel synchronous acquisition system, perform noise reduction and waveform adaptive alignment on the raw monitoring signal to obtain standard monitoring data, and upload it to the cloud server.

[0190] The model adjustment unit 303 is used to collect operating status data on the cloud server and, based on the operating status data, use the model parameter adaptive adjuster to adjust the model parameters of the fault detection model to obtain the optimal fault detection model.

[0191] The fault detection unit 304 is used to input standard monitoring data into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

[0192] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0193] Memory, used to store computer programs;

[0194] When a processor executes a program stored in memory, it implements the fault detection method for a high-power pulse source based on artificial intelligence according to the present invention.

[0195] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0196] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0197] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the fault detection method for a high-power pulse source based on artificial intelligence according to embodiments of the present invention.

[0198] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0199] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0200] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0201] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0202] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0203] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fault detection method for a high-power pulse source based on artificial intelligence, characterized in that, The method includes: On a cloud server, a fault detection model based on dual-stream feature fusion and attention mechanism is constructed using a multi-objective swarm intelligent optimization algorithm, and an adaptive model parameter adjuster is configured, including: On a cloud server, a fault detection model is constructed based on dual-stream feature fusion and attention mechanism, and the set of parameters to be optimized for the fault detection model and the multi-objective optimization function of the multi-objective swarm intelligent optimization algorithm are determined. Based on the training sample set in the historical database of the cloud server, and based on the multi-objective optimization function, the multi-objective swarm intelligent optimization algorithm is used to optimize the set of parameters to be optimized, and the Pareto non-dominated optimal solution set is obtained. Based on the optimization objective value, the application scenarios of the model parameters of the fault detection model corresponding to each solution in the Pareto non-dominated optimal solution set are divided into high-precision, balanced, and ultra-fast types. Package the model parameters of the fault detection model corresponding to each application scenario into a Docker container, deploy all Docker containers to the model parameter adaptive adjuster, and connect the model parameter adaptive adjuster to the fault detection model. A multi-channel synchronous acquisition system is used to acquire the raw monitoring signal of a high-power pulse source. The raw monitoring signal is then denoised and the waveform is adaptively aligned to obtain standard monitoring data, which is then uploaded to a cloud server. On a cloud server, operational status data is collected, and based on this data, an adaptive model parameter adjuster is used to adjust the model parameters of the fault detection model to obtain the optimal fault detection model, including: On the cloud server, operational status data is collected, including the CPU utilization rate of the cloud server. and concurrent request queue length ; If CPU utilization CPU utilization threshold and / or concurrent request queue length Length threshold Then, the model parameters corresponding to the high-speed type are retrieved from the model parameter adaptive adjuster, the fault detection model is adjusted, and the optimal fault detection model is obtained. If CPU utilization CPU utilization threshold and concurrent request queue length Length threshold Then, the model parameters corresponding to the high-precision type are retrieved from the model parameter adaptive adjuster, the fault detection model is adjusted, and the optimal fault detection model is obtained. If the above rules are not met, the model parameters corresponding to the balanced type are retrieved from the model parameter adaptive adjuster to adjust the fault detection model and obtain the optimal fault detection model. Standard monitoring data is input into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

2. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 1, characterized in that, The fault detection model includes a physical parameter stream extraction module, a depth waveform stream extraction module, an anti-interference attention weighting module, and a fault measurement and classification module.

3. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 2, characterized in that, The multi-objective optimization function includes a function to maximize fault classification accuracy, a function to minimize model computational complexity, and a function to minimize model runtime.

4. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 3, characterized in that, Based on the training sample set in the historical database of the cloud server, and using a multi-objective swarm intelligent optimization algorithm based on a multi-objective optimization function, the set of parameters to be optimized is optimized to obtain the Pareto non-dominated optimal solution set, including: The set of parameters to be optimized is encoded into individual vectors of the multi-objective Osprey optimization algorithm, and the multi-objective Osprey optimization algorithm is initialized. The chaotic sequence is generated using the Logistic mapping, and then mapped to the solution space of individual ospreys to obtain the initial osprey population. Based on the training sample set in the historical database of the cloud server, according to the set of candidate parameters to be optimized for each individual osprey in the initial osprey population, the corresponding candidate fault detection model is quickly trained, and the corresponding optimization target value is calculated using a multi-objective optimization function. Based on the optimization objective value, the initial osprey population is subjected to fast non-dominated sorting to obtain different frontier levels, and an external archive set is maintained to store all non-dominated solutions of the first frontier level. By introducing the Levy flight mechanism and the gray wolf cooperation mechanism, and using the hovering search behavior and diving capture behavior of the multi-target osprey optimization algorithm, the initial osprey population is updated to obtain an updated osprey population. Using a multi-objective optimization function, the optimization objective value corresponding to each updated osprey in the updated osprey population is calculated, and the frontier level and external archive set are updated according to the optimization objective value. The iteration continues until the iteration count reaches the threshold, at which point the iterative update of the osprey population stops, and all non-dominated solutions in the external archive set are output as the Pareto non-dominated optimal solution set.

5. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 4, characterized in that, By introducing the Levy flight mechanism and the gray wolf cooperative mechanism, and using the hovering search behavior and diving capture behavior of the multi-target Osprey optimization algorithm, the initial Osprey population is updated to obtain an updated Osprey population, including: Based on the gray wolf collaboration mechanism and crowding, the top three local optima from the external archives were identified and defined as Alpha individuals, Beta individuals, and Delta individuals, respectively. The optimal Osprey individual was then expanded to include Alpha individuals, Beta individuals, and Delta individuals. Alpha, Beta, and Delta individuals are introduced. Based on the circling search behavior of the Osprey algorithm, the initial Osprey individuals in the initial Osprey population are updated to obtain the updated first Osprey individual. The formula is: In the formula, Number of iterations t+ The first result obtained from the hovering search behavior is the first i The first updated Osprey individual; The convergence factor; Number of iterations t The osprey individual, the initial osprey individual during the first iteration update; Number of iterations t Alpha, Beta, and Delta individuals; Individual ospreys were selected randomly. Flight radius; To explore weights; Prioritize development; t This represents the current iteration number; In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. , To adjust the parameters; It is the hyperbolic tangent function; The Levy flight mechanism is introduced, and the initial Osprey individuals in the initial Osprey population are updated based on the swooping capture behavior of the Osprey algorithm to obtain the updated second Osprey individuals. The formula is: In the formula, Number of iterations t+ The first dive-and-capture maneuver yielded the [number missing]th [missing information]. i A newer second Osprey individual; For the Levy flight mechanism Levy Distribute random numbers; b for Levy Step length, and b ∈[1,2]; As the second development weight; By integrating several updated first-generation osprey individuals and several updated second-generation osprey individuals, an updated osprey population is obtained.

6. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 5, characterized in that, A multi-channel synchronous acquisition system is used to acquire the raw monitoring signal from a high-power pulse source. The raw monitoring signal is then denoised and adaptively aligned to obtain standard monitoring data, which is uploaded to a cloud server, including: A multi-channel synchronous acquisition system is used to acquire the raw monitoring signals of a high-power pulse source. The raw monitoring signals include high-frequency voltage waveforms, current waveforms, drain current of switching devices, and temperature data. The variational mode decomposition algorithm is used to perform strong electromagnetic interference denoising on the high-frequency voltage waveform, current waveform and drain current of switching device in the original monitoring signal, and to obtain the denoised high-frequency voltage waveform, denoised current waveform and denoised drain current of switching device. The moving average filtering algorithm is used to remove random jitter interference from the temperature data in the original monitoring signal to obtain the denoised temperature data. A standard high-frequency voltage waveform is selected from the historical normal data in the historical database of the cloud server as a time reference template, and the cross-correlation function between the denoised high-frequency voltage waveform and the time reference template is calculated to find the time offset corresponding to the peak value of the cross-correlation function. Based on the time offset, the denoised high-frequency voltage waveform, the denoised current waveform, and the denoised drain current of the switching device are adaptively aligned to obtain the aligned high-frequency voltage waveform, the aligned current waveform, and the aligned drain current of the switching device. The aligned high-frequency voltage waveform, aligned current waveform, aligned drain current of switching devices, and denoised temperature data are normalized to obtain standard monitoring data, which are then uploaded to the cloud server.

7. The fault detection method for a high-power pulse source based on artificial intelligence according to claim 6, characterized in that, Standard monitoring data is input into the optimal fault detection model for fault detection. The fault detection results for the high-power pulse source include: Input standard monitoring data into the optimal fault detection model; The physical parameter flow extraction module of the optimal fault detection model calculates the physical parameter flow feature vector of the standard monitoring data based on the preset physical sensitivity index. The deep waveform stream extraction module of the optimal fault detection model extracts deep spatiotemporal features, namely deep waveform stream feature vectors, from standard monitoring data based on convolutional neural networks. Using the anti-interference attention weighting module of the optimal fault detection model, feature weights of the physical parameter flow feature vector are generated based on the time position index. Based on the feature weights, the physical parameter flow feature vector is differentially weighted to obtain the weighted physical parameter flow feature vector. The fault metric classification module of the optimal fault detection model concatenates and fuses the weighted physical parameter flow feature vector and the depth waveform flow feature vector to obtain fused features. Based on metric learning, the distance between the fused features and the center of various fault prototypes is calculated, and the fault detection result is generated based on the distance and the set confidence threshold.

8. A fault detection device for a high-power pulse source based on artificial intelligence, used to implement the fault detection method for a high-power pulse source as described in any one of claims 1-7, characterized in that, The device includes: The model building unit is used on a cloud server to build a fault detection model based on dual-stream feature fusion and attention mechanism using a multi-objective swarm intelligent optimization algorithm, and to configure an adaptive adjuster for model parameters. The data acquisition unit is used to acquire the raw monitoring signal of the high-power pulse source using a multi-channel synchronous acquisition system, perform noise reduction and waveform adaptive alignment on the raw monitoring signal to obtain standard monitoring data, and upload it to the cloud server. The model adjustment unit is used to collect operational status data on the cloud server and, based on the operational status data, use the model parameter adaptive adjuster to adjust the model parameters of the fault detection model to obtain the optimal fault detection model. The fault detection unit is used to input standard monitoring data into the optimal fault detection model to perform fault detection and obtain the fault detection results of the high-power pulse source.

Images