A photovoltaic power station equipment fault diagnosis method based on multi-modal large model

Abstract

The present invention relates to the field of photovoltaic power station fault diagnosis technology, and specifically to a photovoltaic power station equipment fault diagnosis method based on a multimodal large model, comprising the following steps: multimodal data acquisition and synchronous processing to form a comprehensive multimodal data set; establishing an equipment health status model; by comparing the generated comprehensive health status vector with the health status baseline of the equipment, calculating the difference between the two, and triggering a physical intervention test when an abnormal health status of the current equipment is detected; the vibration excitation also includes adaptive frequency adjustment, automatically adjusting the vibration excitation frequency according to the material characteristics of the equipment to avoid resonance with the equipment, and identifying the fault type. The present invention enables the system to accurately monitor the health status of the equipment without adding additional computing burden, improves the speed and accuracy of equipment fault identification, and has important advantages in complex scenarios such as photovoltaic power stations with multiple devices and multiple parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power station fault diagnosis, and in particular to a photovoltaic power station equipment fault diagnosis method based on a multi-modal large model. Background Art

[0002] As an important part of modern new energy systems, the stability and reliability of photovoltaic power station equipment directly affect the power generation efficiency and economic benefits. The operating environment of photovoltaic power stations is complex, and there are many types of equipment, including photovoltaic modules, inverters, distribution systems, etc. These devices may be affected by factors such as fatigue damage, mechanical vibration, and ambient temperature changes during long-term operation, resulting in performance degradation or even failure. Traditional fault diagnosis methods usually rely on single modal data (such as electrical parameters or mechanical vibration signals) for monitoring, lack a comprehensive understanding of the overall operating status of the equipment, and are prone to problems such as inaccurate fault feature detection and untimely response.

[0003] To solve this problem, in recent years, multimodal data fusion technology has been gradually applied to equipment fault diagnosis. By integrating data from multiple sensors, the operating status of the equipment can be comprehensively evaluated from multiple angles. However, existing multimodal data fusion methods are mostly static analysis, which reacts slowly to dynamically changing equipment states, and the extraction of vibration signals is not sufficiently correlated with fault characteristics, resulting in limited effects in early fault detection and precise fault location.

[0004] In addition, vibration excitation, as an effective means to enhance fault feature extraction, has also been widely used in the field of equipment health monitoring. However, most of the current vibration excitation technologies lack targeted analysis of equipment material characteristics and failure modes, which can easily cause equipment resonance or over-stimulation, affecting the safety and operational stability of the equipment. Summary of the invention

[0005] The present invention provides a photovoltaic power station equipment fault diagnosis method based on a multi-modal large model.

[0006] A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model comprises the following steps:

[0007] S1, multimodal data acquisition and synchronous processing: deploy a variety of sensors on various devices in the photovoltaic power station, including electrical sensors (such as current and voltage), mechanical vibration sensors, environmental monitoring sensors (such as temperature, humidity, wind speed) and acoustic sensors, respectively, to collect different modal data during the operation of the equipment, and ensure the consistency of the data through timestamp synchronization to form a comprehensive multimodal data set;

[0008] S2, initialization and construction of the equipment health status model, based on the multimodal data set collected in S1, establishes the equipment health status model, including the following submodules:

[0009] S21, initial health status baseline submodule: by analyzing the historical data of the equipment in normal operation, using multimodal feature extraction technology to extract the operation characteristics of the equipment in the absence of faults, and constructing a health status baseline, the health status baseline includes an electrical feature baseline, a mechanical vibration feature baseline, and a feature baseline of the impact of environmental data on equipment operation. The health status baseline is a standard reference when the equipment is not faulty;

[0010] S22, comprehensive health status vector generation submodule: using multimodal feature extraction technology, extracting equipment operation features from multimodal data collected in real time, generating a comprehensive health status vector reflecting the current state of the current equipment, the comprehensive health status vector including the overall health information of the equipment under various modes;

[0011] S3, real-time health status comparison and fault identification: by comparing the comprehensive health status vector generated in S2 with the health status baseline of the device, the difference between the two is calculated. If the difference exceeds the preset threshold, the current health status of the current device is abnormal;

[0012] S4, physical intervention test and feature enhancement extraction: When the health status of the current device is detected to be abnormal in S3, a physical intervention test is triggered. The physical intervention test includes applying vibration excitation when the device is running, monitoring the response characteristics of the vibration signal, extracting the enhanced mechanical vibration feature vector, and then comparing it with the mechanical vibration feature baseline in the health status baseline to confirm the existence of the fault. At the same time, by analyzing the response behavior characteristics of the device under vibration excitation, it is confirmed whether there is physical damage or structural problems. The vibration excitation also includes adaptive frequency adjustment, which automatically adjusts the vibration excitation frequency according to the material characteristics of the device to avoid resonance with the device and identify the type of fault.

[0013] Optionally, the multimodal data collection and synchronization processing in S1 specifically includes:

[0014] S11, sensor deployment strategy: deploy multiple sensors on various equipment in the photovoltaic power station:

[0015] Electrical sensors: Install current and voltage sensors on inverters and distribution cabinets to monitor the electrical parameters of the equipment in real time;

[0016] Mechanical vibration sensor: Arrange mechanical vibration sensors on brackets, cooling systems, and moving parts of mechanical equipment to collect vibration status data of the equipment;

[0017] Environmental monitoring sensors: Temperature, humidity, and wind speed sensors are installed in the center of the photovoltaic array to obtain the impact of the environment on equipment operation;

[0018] S12, each sensor collects data according to the operating status and monitoring requirements of the equipment. The electrical sensor records the real-time changes of current and voltage. The mechanical vibration sensor captures the vibration characteristics of the equipment under different working conditions. The environmental monitoring sensor records the dynamic changes of the external environment.

[0019] S13, to ensure the time consistency of data of different modalities, all collected data are marked through a unified timestamp synchronization mechanism. The data collected by each sensor are timestamped, and the Network Time Protocol (NTP) is used to ensure that the time base of each sensor is consistent;

[0020] S14, the time-synchronized multimodal data is organized according to the device and time dimensions to form a comprehensive multimodal data set.

[0021] Optionally, the initial health status baseline submodule specifically includes:

[0022] S211, based on the comprehensive multimodal data set, extracts multimodal data under fault-free conditions from the historical operation records of the equipment, including electrical parameters (current, voltage), mechanical vibration characteristics, and environmental monitoring data (temperature, humidity, wind speed), to form a multimodal historical data set of the equipment under normal operating conditions;

[0023] S212, in the initial health status baseline submodule, an autoencoder algorithm is used to analyze the multimodal historical data set of the device, specifically:

[0024] Feature extraction: The autoencoder algorithm is used to automatically extract the main operating characteristics of the equipment in normal operation, including electrical characteristics, mechanical vibration characteristics, and the impact of environmental data on equipment operation, to generate a low-dimensional feature vector Z.

[0025] Mode clustering: The K-means clustering algorithm clusters the low-dimensional feature vector Z and groups similar features together to form a typical operation mode of the device. Each cluster represents the healthy operation status of the device under specific conditions.

[0026] S213, constructing a health status baseline: using the cluster center Cj obtained by the K-means clustering algorithm as the core part of the health status baseline, each cluster center represents a standard health status mode of a device, and constructing a health status baseline of the device through a combination of modes. The health status baseline includes a reference range of multimodal features of the device as a standard model of the normal state of the device;

[0027] S214, baseline storage: The constructed health status baseline is stored in the device health status model as a standard reference for subsequent real-time health status comparison and fault identification.

[0028] Optionally, the autoencoder algorithm extracts the main operating features from the multimodal historical data set of the device. The input of the autoencoder algorithm is the multimodal data of the device, and the output is a low-dimensional feature vector when the device operates normally. The input is expressed as: , represents the historical multimodal data from different sensors, where each Represents a multimodal data point at a moment, expressed as follows:

[0029] ;

[0030] in, Indicates voltage, Indicates the current, represents the vibration frequency, represents the vibration acceleration, Indicates temperature, Indicates humidity, Indicates wind speed;

[0031] The autoencoder algorithm maps the input multimodal data into a low-dimensional feature vector , each yes The low-dimensional feature representation at each moment indicates the operating status of the device:

[0032] ,in, ReLU represents the activation function in the autoencoder, which is used to introduce nonlinear transformation. is the weight matrix of the encoder in the autoencoder, is the bias vector of the encoder in the autoencoder, It contains comprehensive features of electrical characteristics, mechanical vibration patterns, and environmental data, which are used to describe the main operating status of the device at the current moment. The low-dimensional vector is the extracted feature representation, including the main features of the device under normal operating conditions.

[0033] Optionally, the feature extraction technology in the comprehensive health status vector generation submodule of S22 adopts an autoencoder algorithm to extract key features reflecting the current operating status of the device for data of different modalities:

[0034] Input and output real-time data:

[0035] Input: multimodal data collected in real time;

[0036] Output: The autoencoder maps the input real-time multimodal data to a low-dimensional space and generates a real-time feature vector that represents the operating status of the device at the current moment:

[0037] ,in, It is the comprehensive operation feature vector of the device at the current time t extracted by the autoencoder. It has the same format as the extracted historical baseline features mentioned above and can be used to compare with the health status baseline. Represents the data vector input by the device in real time at time t;

[0038] The electrical characteristics, mechanical vibration characteristics and environmental characteristics are integrated to generate a comprehensive health status vector that reflects the current comprehensive operating status of the equipment:

[0039] in, represents the comprehensive health state vector generated at time t, represents the electrical characteristic vector generated at time t, represents the mechanical vibration eigenvector generated at time t, Represents the environmental feature vector generated at time t, which fully reflects the operating status of the device at the current moment. It is compared with the health status baseline constructed by class analysis in real time.

[0040] Optionally, the physical intervention test and feature enhancement extraction in S4 specifically include:

[0041] S41, when the health status of the equipment is determined to be abnormal by comparing the comprehensive health status vector with the health status baseline, the physical intervention test process is automatically triggered to confirm and enhance the extraction of fault features;

[0042] S42, vibration excitation application: applying external vibration excitation to the mechanical structure or vibration-sensitive parts of the equipment during the physical intervention test;

[0043] S43, vibration signal response monitoring: through the mechanical vibration sensor installed on the equipment, the response characteristics of the equipment under vibration excitation are monitored in real time, and the vibration acceleration, frequency change and dynamic response behavior of the equipment are recorded;

[0044] S44, extracting the response characteristics of the vibration signal again under the action of the vibration excitation signal to generate an enhanced mechanical vibration feature vector;

[0045] S45, the enhanced mechanical vibration characteristic vector is compared with the healthy status baseline again to confirm the existence of a fault, and based on the response behavior characteristics of the equipment under vibration excitation, it is confirmed whether there is physical damage or structural problems.

[0046] Optionally, under the action of the vibration excitation signal, the enhanced vibration characteristics are extracted by a Fourier transform frequency domain analysis method to generate an enhanced vibration feature vector, where the enhanced vibration feature vector includes enhanced characteristics of the mechanical vibration signal, reflecting the vibration frequency, acceleration change and response mode of the equipment:

[0047] ,in, represents the enhanced vibration frequency characteristics, Represents the enhanced acceleration feature, which will enhance the vibration feature vector Vibration signature part of the equipment health baseline Compare and calculate the difference between the two to determine whether there is any abnormality or damage to the mechanical structure of the equipment:

[0048] If the difference If the preset threshold is exceeded, combined with the preliminary difference calculation of S3, it is verified again that the device is abnormal.

[0049] Optionally, the adaptive frequency adjustment specifically includes:

[0050] Material property analysis: Based on the mechanical structure material properties of the equipment, obtain the natural frequency of the equipment :

[0051] ,in: is the natural frequency of the device, is the material stiffness of the device, It is the quality of the equipment;

[0052] Fault mode identification: Identify fault modes based on the historical fault data of the equipment. Fault modes include fatigue, cracks, and structural looseness. Different fault modes correspond to mechanical vibration characteristics at specific frequencies. By classifying historical fault modes, associate faults with their corresponding vibration frequency ranges.

[0053] Adaptive Frequency Based on an optimization algorithm, the vibration excitation frequency is automatically selected to induce a specific failure mode.

[0054] Optionally, the optimization algorithm specifically includes:

[0055] Initialize Frequency Population: Generates a set of initial frequency values ​​within the frequency range of material properties and failure modes: ,in, is the initial frequency group;

[0056] Fitness function evaluation: For each frequency value, use the fitness function to evaluate its effectiveness for fault feature extraction. The fitness function can be defined based on the response characteristics of the material properties and the fault mode, and the goal is to maximize the response signal strength of the fault feature extraction;

[0057] ,in, represents the fitness function, Indicates frequency The extraction effect of fault features, ResonanceRisk represents the resonance risk, Indicates frequency Risk of causing resonance in the equipment (based on the natural frequency of the material), and is the weight factor, ;

[0058] Crossover and mutation: Through the crossover and mutation operations in the genetic algorithm, the frequency population is updated to find a better frequency combination. The crossover operation combines frequency features to generate new frequencies, and the mutation operation introduces random disturbances to explore more frequency values.

[0059] Frequency selection: After multiple rounds of iterations, the optimal vibration excitation frequency is selected , that is, the frequency that can most effectively induce the fault characteristics and avoid resonance based on the material characteristics and failure mode, and use the selected optimal frequency , vibration excitation is applied by an electromagnetic vibrator or a piezoelectric vibrator.

[0060] Optionally, the fault mode identification includes analyzing historical fault data of the device to identify different fault modes;

[0061] A failure mode library is established, which includes the characteristics of each failure mode and the vibration frequency range associated with it.

[0062] Beneficial effects of the present invention:

[0063] The present invention uses multimodal feature extraction technology to quickly extract the operating characteristics of the current device from various sensor data of the device (such as electrical parameters, mechanical vibration characteristics, environmental data, etc.), generate a comprehensive health status vector, and combine it with the autoencoder feature extraction technology to achieve efficient data compression and feature expression, so that the system can accurately monitor the health status of the device without adding additional computing burden, improve the speed and accuracy of equipment fault identification, and have important advantages in complex scenarios such as photovoltaic power stations with multiple devices and multiple parameters.

[0064] The present invention proposes an adaptive vibration excitation frequency adjustment method based on equipment material characteristics and fault modes. By analyzing the mechanical material characteristics (natural frequency) and historical fault data of the equipment and combining the optimization algorithm, the optimal vibration excitation frequency is automatically selected, thereby avoiding the risk of resonance between the equipment and the vibration excitation frequency. This not only improves the accuracy of fault feature extraction, but also greatly reduces unnecessary frequency adjustment time, thereby ensuring the safety of vibration excitation for the equipment.

[0065] The present invention, by introducing a classification model, automatically classifies common failure modes of equipment (such as fatigue damage, structural looseness, etc.) based on historical data, and can accurately select the most suitable vibration excitation frequency range according to the characteristics of different failure modes, thereby achieving a deep understanding of the failure mode and targeted correction, and providing strong technical support for the rapid repair of equipment and the formulation of maintenance strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0068] Figure 2 Schematic diagram of a device health status model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0070] It should be noted that the references to "one embodiment", "embodiment", "exemplary embodiments", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0071] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0072] like Figure 1-Figure 2 As shown, a photovoltaic power station equipment fault diagnosis method based on a multi-modal large model includes the following steps:

[0073] S1, multimodal data acquisition and synchronous processing: deploy a variety of sensors on various devices in the photovoltaic power station, including electrical sensors (such as current and voltage), mechanical vibration sensors, environmental monitoring sensors (such as temperature, humidity, wind speed) and acoustic sensors, respectively, to collect different modal data during the operation of the equipment, and ensure the consistency of the data through timestamp synchronization to form a comprehensive multimodal data set;

[0074] S2, initialization and construction of the equipment health status model, based on the multimodal data set collected in S1, establishes the equipment health status model, including the following submodules:

[0075] S21, initial health status baseline submodule: by analyzing the historical data of the equipment in normal operation, using multimodal feature extraction technology to extract the operating characteristics of the equipment in the absence of faults, and constructing a health status baseline. The health status baseline includes an electrical feature baseline, a mechanical vibration feature baseline, and a feature baseline of the impact of environmental data on equipment operation. The health status baseline is a standard reference when the equipment is not faulty;

[0076] S22, comprehensive health status vector generation submodule: using multimodal feature extraction technology, extracting equipment operation features from multimodal data collected in real time, generating a comprehensive health status vector reflecting the current state of the current equipment, the comprehensive health status vector including the overall health information of the equipment under various modes;

[0077] S3, real-time health status comparison and fault identification: by comparing the comprehensive health status vector generated in S2 with the health status baseline of the device, the difference between the two is calculated. If the difference exceeds the preset threshold, the current health status of the current device is abnormal;

[0078] S4, physical intervention test and feature enhancement extraction: When an abnormal health status of the current device is detected in S3, a physical intervention test is triggered. The physical intervention test includes applying vibration excitation when the device is running, monitoring the response characteristics of the vibration signal, extracting the enhanced mechanical vibration feature vector, and then comparing it with the mechanical vibration feature baseline in the healthy status baseline to confirm the existence of the fault. At the same time, by analyzing the response behavior characteristics of the device under vibration excitation, it is confirmed whether there is physical damage or structural problems. The fault characteristics of the device are further extracted and enhanced through physical intervention means, and these enhanced characteristics are used for comparative analysis with the device health status baseline to improve the accuracy and reliability of fault identification. Vibration excitation also includes adaptive frequency adjustment, which automatically adjusts the vibration excitation frequency according to the material characteristics of the device to avoid resonance with the device and identify the type of fault.

[0079] The multimodal data acquisition and synchronous processing in S1 specifically includes:

[0080] S11, sensor deployment strategy: deploy multiple sensors on various equipment in the photovoltaic power station:

[0081] Electrical sensors: Install current and voltage sensors on inverters and distribution cabinets to monitor the electrical parameters of the equipment in real time;

[0082] Mechanical vibration sensor: Arrange mechanical vibration sensors on brackets, cooling systems, and moving parts of mechanical equipment to collect vibration status data of the equipment;

[0083] Environmental monitoring sensors: Temperature, humidity, and wind speed sensors are installed in the center of the photovoltaic array to obtain the impact of the environment on equipment operation;

[0084] S12, each sensor collects data according to the operating status and monitoring requirements of the equipment. The electrical sensor records the real-time changes of current and voltage. The mechanical vibration sensor captures the vibration characteristics of the equipment under different working conditions. The environmental monitoring sensor records the dynamic changes of the external environment.

[0085] S13, to ensure the time consistency of data of different modalities, all collected data are marked through a unified timestamp synchronization mechanism. The data collected by each sensor are timestamped, and the Network Time Protocol (NTP) is used to ensure that the time base of each sensor is consistent to avoid data deviation caused by time asynchrony;

[0086] S14, the multimodal data after time synchronization is organized according to the device and time dimensions to form a comprehensive multimodal data set, which also contains the electrical characteristics, mechanical vibration characteristics, and environmental conditions of the equipment, providing high-quality input for the construction of the subsequent equipment health status model.

[0087] 1. Equipment dimension: There may be multiple devices in a photovoltaic power station, such as photovoltaic modules, inverters, power distribution cabinets, cooling systems, etc. Each device will correspond to a set of sensors (electrical sensors, vibration sensors, environmental monitoring sensors, etc.), and the multimodal data generated by these sensors will be associated with the devices to which they belong. Therefore, the collected data needs to be organized based on the device to ensure that it is clear which device each set of data comes from.

[0088] 2. Time dimension: The data of each device is collected continuously over time, so it is necessary to sort the data by timestamp to ensure that the status changes of the device at a specific time can be tracked. The organization of the time dimension helps the system to perform time series analysis and analyze the performance changes and fault precursors of the device in a specific time period.

[0089] The initial health status baseline submodule specifically includes:

[0090] S211, based on the comprehensive multimodal data set, extracts multimodal data under fault-free conditions from the historical operation records of the equipment, including electrical parameters (current, voltage), mechanical vibration characteristics, and environmental monitoring data (temperature, humidity, wind speed), to form a multimodal historical data set of the equipment under normal operating conditions;

[0091] S212, in the initial health status baseline submodule, an autoencoder algorithm is used to analyze the multimodal historical data set of the device, specifically:

[0092] Feature extraction: The autoencoder algorithm is used to automatically extract the main operating characteristics of the device under normal operating conditions, including electrical characteristics (normal fluctuation range of voltage and current), mechanical vibration characteristics (variation range of vibration frequency and acceleration), and the impact of environmental data on device operation, and generate a low-dimensional feature vector Z.

[0093] Mode clustering: The K-means clustering algorithm clusters the low-dimensional feature vector Z and groups similar features together to form a typical operation mode of the device. Each cluster represents the healthy operation status of the device under specific conditions.

[0094] S213, health status baseline construction: The cluster center Cj obtained by the K-means clustering algorithm is used as the core part of the health status baseline. Each cluster center represents a standard health status mode of a device. The cluster center can be used as a reference point for subsequent fault detection. Any features beyond these reference ranges will trigger abnormal detection. The health status baseline of the device is constructed through the combination of patterns. The health status baseline includes the reference range of the multi-modal features of the device. As a standard model of the normal state of the device, the baseline not only records the normal operating values ​​under each mode, but also includes the correlation features between different modes (such as the coupling relationship between electrical data and mechanical vibration);

[0095] S214, baseline storage: The constructed health status baseline is stored in the equipment health status model as a standard reference for subsequent real-time health status comparison and fault identification. When the real-time data is compared with the baseline, if the characteristics that deviate from the health status baseline exceed the threshold, it can be determined that the equipment is abnormal or faulty.

[0096] An autoencoder is a neural network structure used for unsupervised feature extraction and is often used for dimensionality reduction and reconstruction of data. In this system, an autoencoder can be used to extract features from multimodal data and extract the operating characteristics of the device under normal conditions.

[0097] The autoencoder consists of two parts:

[0098] Encoder: Maps input data into a low-dimensional latent space, generating a compressed feature representation.

[0099] Decoder: reconstructs the original data from the low-dimensional feature representation.

[0100] The input data is a multimodal data set of the device. The encoder maps the input data to a low-dimensional space through a nonlinear function. The decoder maps the low-dimensional feature vector back to the original space in an attempt to reconstruct the input data. The training goal of the autoencoder is to minimize the error between the input data and the reconstructed data.

[0101] The autoencoder algorithm extracts the main operating features from the multimodal historical data set of the device. The input of the autoencoder algorithm is the multimodal data of the device, and the output is the low-dimensional feature vector when the device is operating normally. The input is expressed as: , represents the historical multimodal data from different sensors, where each Represents a multimodal data point at a moment, expressed as follows:

[0102] ;

[0103] in, Indicates voltage, Indicates the current, represents the vibration frequency, represents the vibration acceleration, Indicates temperature, Indicates humidity, Indicates wind speed. The input value comes from electrical sensors, mechanical vibration sensors and environmental monitoring sensors, reflecting the comprehensive status of equipment operation.

[0104] The autoencoder algorithm maps the input multimodal data into a low-dimensional feature vector , each yes The low-dimensional feature representation at each moment indicates the operating status of the device:

[0105] ,in, ReLU represents the activation function in the autoencoder, which is used to introduce nonlinear transformation. is the weight matrix of the encoder in the autoencoder, is the bias vector of the encoder in the autoencoder, It contains comprehensive features of electrical characteristics, mechanical vibration patterns, and environmental data, which are used to describe the main operating status of the device at the current moment. The low-dimensional vector is the extracted feature representation, including the main features of the device under normal operating conditions.

[0106] The purpose of the autoencoder is to automatically extract the main operating characteristics of the device from the input high-dimensional data. The following is the specific extraction process:

[0107] 1. Electrical feature extraction: Input electrical parameters (voltage and current ) is mapped to a low-dimensional feature vector through an encoder. The autoencoder extracts the characteristic pattern of the device under stable operation by learning the voltage and current fluctuation range of the device under normal conditions. For example, under normal operating conditions, the fluctuation range of voltage and current is usually within a certain range. When their values ​​deviate from this range, it indicates that the device is abnormal, and the output low-dimensional feature vector is In the diagram, the electrical characteristics are expressed as: ,This part of the eigenvector is used to reflect the normal fluctuation range of voltage and current.

[0108] 2. Mechanical vibration feature extraction: vibration frequency and acceleration As the key parameter of mechanical vibration, it is input into the autoencoder. By analyzing the vibration data under different operating conditions, the autoencoder can learn the vibration characteristics of the equipment during normal operation (such as the variation range of vibration frequency and acceleration). Among the output low-dimensional features, the vibration feature part is expressed as: ,This part of the eigenvector describes the vibration frequency and acceleration range of the ,machine's mechanical structure during normal operation. Any abnormal fluctuation ,will reflect the failure of the mechanical components.

[0109] 3. Environmental feature extraction: environmental parameters (temperature ,humidity , wind speed ) directly affects the operating status of the device. The autoencoder automatically extracts the impact characteristics of the device operation by inputting these environmental parameters. The output of the environmental characteristics is expressed as:

[0110] ,This part of the features reflects the impact of environmental changes on ,the operation of the equipment.,For example, excessive temperature may cause the equipment to overheat, ,while changes in wind speed may affect the efficiency of the ,cooling system.

[0111] By inputting the multimodal data of the equipment in normal operation into the autoencoder, the autoencoder will learn the main operating characteristics of the equipment when it is fault-free and generate a low-dimensional feature vector ,These low-dimensional features represent the stable states of the equipment’s electrical, mechanical vibrations and environmental conditions under normal circumstances.

[0112] By performing cluster analysis on these low-dimensional features, we can form a health status baseline for the equipment, that is, the typical operating mode of the equipment when there is no fault. These baseline features are used for subsequent real-time comparisons. When new data deviates from the baseline features, it can be determined whether the equipment has abnormalities or potential faults.

[0113] K-means clustering is used to identify the normal operation mode of the device in different states. By clustering the low-dimensional feature vector Z, a baseline reference for different operating states is formed. The algorithm flow of K-means clustering is as follows:

[0114] 1. Initialize cluster center: randomly select Cluster Center ,Each cluster center represents an operation mode;

[0115] 2. Assign data points: Assign each data point Assign to the nearest cluster center:

[0116] ,in Represents data points With cluster center The Euclidean distance between .

[0117] 3. Update cluster centers: Recalculate the center position of each cluster, that is, the mean of all data points in each cluster: ,in, Indicates The set of all member data points of a cluster.

[0118] 4. Repeat steps 2 and 3 until the cluster center no longer changes or the change is less than the set threshold.

[0119] Objective function of K-means: The goal is to minimize the intra-cluster squared error, which is the sum of the squared distances between data points and their cluster centers: ,After clustering, each cluster center represents the characteristic value of the ,device in different operating modes, and the health status baseline can ,be constructed based on the mean of these cluster centers.

[0120] The feature extraction technology in the S22 comprehensive health status vector generation submodule uses an autoencoder algorithm to extract key features that reflect the current operating status of the device for data of different modalities:

[0121] Electrical feature extraction: By analyzing the real-time fluctuations of voltage and current, the stability and fluctuation range of electrical parameters at the current moment can be extracted.

[0122] Mechanical vibration feature extraction: Based on the real-time vibration frequency and acceleration data collected by the vibration sensor, the current mechanical state characteristics of the equipment are extracted to identify the stability of the mechanical structure and potential abnormal vibrations.

[0123] Environmental feature extraction: Combine real-time temperature, humidity, and wind speed data to extract the impact of current environmental conditions on device operation and identify whether the device is in a critical state under environmental influence.

[0124] Input and output real-time data:

[0125] Input: Multimodal data (voltage, current, vibration, temperature) collected in real time, which is consistent with the historical data type. The input of the autoencoder is all multimodal data of the device at the current moment, which is similar to the processing of historical data.

[0126] Output: The autoencoder maps the input real-time multimodal data to a low-dimensional space and generates a real-time feature vector that represents the operating status of the device at the current moment:

[0127] ,in, is the comprehensive operation feature vector of the device at the current time t extracted by the autoencoder, represents the data vector input by the device in real time at time t, which has the same format as the extracted historical baseline features and can be used to compare with the health status baseline;

[0128] The electrical characteristics, mechanical vibration characteristics and environmental characteristics are integrated to generate a comprehensive health status vector that reflects the current comprehensive operating status of the equipment:

[0129] in, represents the comprehensive health state vector generated at time t, represents the electrical characteristic vector generated at time t, represents the mechanical vibration eigenvector generated at time t, Represents the environmental feature vector generated at time t, which fully reflects the operating status of the device at the current moment. It is compared with the health status baseline constructed by class analysis in real time.

[0130] The comprehensive health state vector generated by real-time multimodal feature extraction Compared with the health status baseline constructed by autoencoder and cluster analysis Perform one-by-one matching, baseline Represents the typical health state of the device in a fault-free situation. The difference between the real-time comprehensive health state vector and the health state baseline is measured by calculating the Euclidean distance between the two:

[0131] ,in, Represents the first Features, The corresponding features representing the baseline of health status;

[0132] Difference threshold determination: The calculated difference value With the preset threshold For comparison:

[0133] like , it indicates that the current health status of the device is normal and the comprehensive health status vector matches the health status baseline.

[0134] like , it indicates that the current health status of the device is abnormal, indicating that there may be a fault or potential risk.

[0135] When the difference value exceeds the threshold When the error occurs, the system will trigger the fault warning mechanism, analyze the abnormal situation of the equipment according to the size of the difference value and the characteristic distribution, and generate a fault warning report for the operator to conduct further inspection and maintenance.

[0136] The physical intervention test and feature enhancement extraction in S4 specifically include:

[0137] S41, when the health status of the equipment is determined to be abnormal by comparing the comprehensive health status vector with the health status baseline, the physical intervention test process is automatically triggered to confirm and enhance the extraction of fault features;

[0138] S42, vibration excitation application: During the physical intervention test, external vibration excitation is applied to the mechanical structure or vibration-sensitive parts of the equipment; an electromagnetic vibrator or a piezoelectric vibrator is used as an external vibration excitation device, and it is fixed to the key mechanical parts or structures of the equipment (bracket, housing, moving parts) to apply external vibration excitation. The electromagnetic vibrator is driven by electric current to generate mechanical vibration, and the piezoelectric vibrator is driven by voltage, which can provide vibrations of different frequencies and intensities according to the specific needs of the equipment.

[0139] The frequency and intensity of the vibration excitation device are automatically controlled by the system. Appropriate vibration excitation parameters are selected according to the current operating status, material properties and fault mode of the equipment to avoid resonance with the natural frequency of the equipment, ensuring that the vibration excitation can effectively induce fault characteristics without causing damage to the equipment.

[0140] S43, vibration signal response monitoring: through the mechanical vibration sensor installed on the equipment, the response characteristics of the equipment under vibration excitation are monitored in real time, and the vibration acceleration, frequency change and dynamic response behavior of the equipment are recorded;

[0141] S44, extracting the response characteristics of the vibration signal again under the action of the vibration excitation signal to generate an enhanced mechanical vibration feature vector, wherein the extracted enhanced characteristics further improve the extraction accuracy of the equipment fault characteristics by deeply analyzing the fault signals such as the vibration response and the mechanical structure stress;

[0142] S45, the enhanced mechanical vibration characteristic vector is compared with the healthy status baseline again to confirm the existence of a fault, and based on the response behavior characteristics of the equipment under vibration excitation, it is confirmed whether there is physical damage or structural problems.

[0143] Under the action of vibration excitation signal, the enhanced vibration characteristics are extracted through Fourier transform frequency domain analysis method to generate enhanced vibration feature vector. The enhanced vibration feature vector includes the enhanced characteristics of mechanical vibration signal, reflecting the vibration frequency, acceleration change and response mode of the equipment:

[0144] ,in, represents the enhanced vibration frequency characteristics, Represents the enhanced acceleration feature, which will enhance the vibration feature vector Vibration signature part of the equipment health baseline Compare and calculate the difference between the two to determine whether there is any abnormality or damage to the mechanical structure of the equipment:

[0145] If the difference If the preset threshold is exceeded, combined with the preliminary difference calculation of S3, it is verified again that the device is abnormal.

[0146] Adaptive frequency adjustment specifically includes:

[0147] Material property analysis: Based on the mechanical structure material properties of the equipment, such as metal, alloy or composite material, obtain the natural frequency of the equipment :

[0148] ,in: is the natural frequency of the device, is the material stiffness of the device, It is the quality of the equipment;

[0149] Fault mode identification: Identify fault modes based on the historical fault data of the equipment. Fault modes include fatigue, cracks, and structural looseness. Different fault modes correspond to mechanical vibration characteristics at specific frequencies. By classifying historical fault modes, associate faults with their corresponding vibration frequency ranges.

[0150] Adaptive frequency automatically selects the vibration excitation frequency based on an optimization algorithm to induce a specific failure mode. The goal of the algorithm is to select the best vibration frequency that will not induce resonance and can effectively extract the fault characteristics based on the material characteristics of the equipment.

[0151] The optimization algorithm specifically includes:

[0152] Initialize Frequency Population: Generates a set of initial frequency values ​​within the frequency range of material properties and failure modes: ,in, is the initial frequency group;

[0153] Fitness function evaluation: For each frequency value, use the fitness function to evaluate its effectiveness for fault feature extraction. The fitness function can be defined based on the response characteristics of the material properties and the fault mode, and the goal is to maximize the response signal strength of the fault feature extraction;

[0154] ,in, represents the fitness function, Indicates frequency The extraction effect of fault features, ResonanceRisk represents the resonance risk, Indicates frequency Risk of causing resonance in the equipment (based on the natural frequency of the material), and is the weight factor, ;

[0155] Crossover and mutation: Through the crossover and mutation operations in the genetic algorithm, the frequency population is updated to find a better frequency combination. The crossover operation combines frequency features to generate new frequencies, and the mutation operation introduces random disturbances to explore more frequency values.

[0156] Frequency selection: After multiple rounds of iterations, the optimal vibration excitation frequency is selected , that is, the frequency that can most effectively induce the fault characteristics and avoid resonance based on the material characteristics and failure mode, and use the selected optimal frequency , vibration excitation is applied by an electromagnetic vibrator or a piezoelectric vibrator.

[0157] Failure mode identification involves analyzing the historical failure data of the equipment to identify different failure modes. Typical failure modes include:

[0158] Fatigue damage: Tiny cracks or fatigue stress caused by long-term operation often show obvious vibration characteristics at higher frequencies.

[0159] Loose structure: Looseness of mechanical parts may cause large vibrations in the medium and low frequency range.

[0160] Material cracks: Tiny cracks may induce local resonance of the device within a specific frequency range, exhibiting a specific frequency response.

[0161] A failure mode library is established, which includes the characteristics of each failure mode and the vibration frequency range associated with it.

[0162] The specific steps are as follows:

[0163] Data collection: Extract the vibration frequency and amplitude of the fault event from the historical operation data of the equipment, and use labeled data to identify different failure modes;

[0164] When establishing a fault mode library, fault mode modeling is included: historical fault data is classified through the support vector machine (SVM) classification model, and a feature vector is assigned to each fault mode. The feature vector includes the frequency component and amplitude change of the vibration signal; for example, the feature of the fatigue damage mode may be the frequency range high amplitude vibrations within the structure; loose structures may induce lower frequencies (such as ), then the fault mode characteristic vector ,in, is the vibration frequency, For the amplitude, assign each failure mode its typical frequency range and establish a failure mode-frequency comparison table.

[0165] Use the normalized historical fault data to train the support vector machine model:

[0166] Input: Fault characteristics (vibration frequency, vibration acceleration, etc.)

[0167] Output: Failure mode label (fatigue damage, structural loosening, etc.)

[0168] Through training, the SVM classification model will find the optimal classification boundary that can distinguish different failure modes.

[0169] In the SVM classification model, the goal of the training process is to find a hyperplane that maximizes the classification interval. Specifically, the SVM classification model performs classification by optimizing the following objective function:

[0170] ; The constraints are as follows: ,in, is the normal vector of the hyperplane, is the offset, is the input sample feature vector, is the class label of the sample, that is, +1 or -1, It means "for all ”, which means that for all , that is, for each sample in the data set , the inequality holds, and here Represents the index of the data sample. The inequality in the support vector machine optimization problem requires that each sample The classification boundary condition is met on one side of the hyperplane;

[0171] Select Gaussian kernel RBF; set the regularization parameter C (control the penalty for misclassification) and kernel parameter (control the width of the Gaussian kernel) of the SVM classification model, and find the optimal parameter combination through cross-validation; use the training set data, input features and corresponding fault mode labels for model training. The SVM classification model will find an optimal hyperplane to distinguish different fault modes. After the model training is completed, the SVM model can be used to classify the test set or new fault data to determine which mode the new equipment fault belongs to.

[0172] Classification decision: For new input data , SVM calculates the category of the data based on the position of the hyperplane: ,in, is the distance from the input data to the hyperplane, Returns the classification results.

[0173] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0174] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model, characterized in that: The following steps are involved: S1, multimodal data acquisition and synchronous processing: deploy a variety of sensors on each device in the photovoltaic power station, including electrical sensors, mechanical vibration sensors, environmental monitoring sensors, and acoustic sensors, to collect different modal data during the operation of the equipment, and ensure data consistency through timestamp synchronization to form a comprehensive multimodal data set; S2, initialization and construction of the equipment health status model, based on the multimodal data set collected in S1, establishes the equipment health status model, including the following submodules: S21, initial health status baseline submodule: by analyzing the historical data of the equipment in normal operation, using multimodal feature extraction technology to extract the operating characteristics of the equipment in a fault-free state, and constructing a health status baseline. The health status baseline includes an electrical feature baseline, a mechanical vibration feature baseline, and a feature baseline of the impact of environmental data on equipment operation; S22, comprehensive health status vector generation submodule: using multimodal feature extraction technology to extract equipment operation features from multimodal data collected in real time, and generate a comprehensive health status vector reflecting the current state of the current equipment; S3, real-time health status comparison and fault identification: by comparing the comprehensive health status vector generated in S2 with the health status baseline of the device, the difference between the two is calculated. If the difference exceeds the preset threshold, the current health status of the current device is abnormal; S4, physical intervention test and feature enhancement extraction: When the health status of the current device is detected to be abnormal in S3, a physical intervention test is triggered. The physical intervention test includes applying vibration excitation when the device is running, monitoring the response characteristics of the vibration signal, extracting the enhanced mechanical vibration feature vector, and then comparing it with the mechanical vibration feature baseline in the health status baseline to confirm the existence of the fault. At the same time, by analyzing the response behavior characteristics of the device under vibration excitation, it is confirmed whether there is physical damage or structural problems. The vibration excitation also includes adaptive frequency adjustment, which automatically adjusts the vibration excitation frequency according to the material characteristics of the device to avoid resonance with the device and identify the type of fault.

2. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 1, characterized in that: The multimodal data acquisition and synchronization processing in S1 specifically includes: S11, sensor deployment strategy: deploy multiple sensors on various equipment in the photovoltaic power station: Electrical sensors: Install current and voltage sensors on inverters and distribution cabinets to monitor the electrical parameters of the equipment in real time; Mechanical vibration sensor: Arrange mechanical vibration sensors on brackets, cooling systems, and moving parts of mechanical equipment to collect vibration status data of the equipment; Environmental monitoring sensors: Temperature, humidity, and wind speed sensors are installed in the center of the photovoltaic array to obtain the impact of the environment on equipment operation; S12, each sensor collects data according to the operating status and monitoring requirements of the equipment. The electrical sensor records the real-time changes of current and voltage. The mechanical vibration sensor captures the vibration characteristics of the equipment under different working conditions. The environmental monitoring sensor records the dynamic changes of the external environment. S13, all collected data are marked through a unified timestamp synchronization mechanism, and the data collected by each sensor are timestamped, using the network time protocol to ensure that the time base of each sensor is consistent; S14, the time-synchronized multimodal data is organized according to the device and time dimensions to form a comprehensive multimodal data set.

3. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 1, characterized in that: The initial health status baseline submodule specifically includes: S211, based on the comprehensive multimodal data set, extracting multimodal data under fault-free conditions from the historical operation records of the equipment, including electrical parameters, mechanical vibration characteristics, and environmental monitoring data, to form a multimodal historical data set of the equipment under normal operating conditions; S212, in the initial health status baseline submodule, an autoencoder algorithm is used to analyze the multimodal historical data set of the device, specifically: Feature extraction: The autoencoder algorithm is used to automatically extract the main operating characteristics of the equipment in normal operation, including electrical characteristics, mechanical vibration characteristics, and the impact of environmental data on equipment operation, to generate a low-dimensional feature vector Z. Mode clustering: The K-means clustering algorithm clusters the low-dimensional feature vector Z and groups similar features together to form a typical operation mode of the device. Each cluster represents the healthy operation status of the device under specific conditions. S213, constructing a health status baseline: using the cluster center Cj obtained by the K-means clustering algorithm as the core part of the health status baseline, each cluster center represents a standard health status mode of a device, and constructing a health status baseline of the device through a combination of modes. The health status baseline includes a reference range of multimodal features of the device as a standard model of the normal state of the device; S214, baseline storage: The constructed health status baseline is stored in the device health status model as a standard reference for subsequent real-time health status comparison and fault identification.

4. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 3, characterized in that: The autoencoder algorithm extracts the main operating features from the multimodal historical data set of the device. The input of the autoencoder algorithm is the multimodal data of the device, and the output is the low-dimensional feature vector when the device is operating normally. The input is expressed as: , represents the historical multimodal data from different sensors, where each Represents a multimodal data point at a moment, expressed as follows: ； in, Indicates voltage, Indicates the current, represents the vibration frequency, represents the vibration acceleration, Indicates temperature, Indicates humidity, Indicates wind speed; The autoencoder algorithm maps the input multimodal data into a low-dimensional feature vector , each yes The low-dimensional feature representation at each moment indicates the operating status of the device: ,in, ReLU represents the activation function in the autoencoder, which is used to introduce nonlinear transformation. is the weight matrix of the encoder in the autoencoder, is the bias vector of the encoder in the autoencoder, It contains comprehensive features of electrical characteristics, mechanical vibration patterns, and environmental data, which are used to describe the main operating status of the device at the current moment. The low-dimensional vector is the extracted feature representation, including the main features of the device under normal operating conditions.

5. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 4, characterized in that: The feature extraction technology in the comprehensive health status vector generation submodule of S22 adopts an autoencoder algorithm to extract key features reflecting the current operating status of the device for data of different modalities; Input and output real-time data, where the input is multimodal data collected in real time; the output is the autoencoder mapping the input real-time multimodal data to a low-dimensional space to generate a real-time feature vector, representing the operating status of the device at the current moment: ,in, is the comprehensive operation feature vector of the device at the current time t extracted by the autoencoder, Represents the data vector input by the device in real time at time t; The electrical characteristics, mechanical vibration characteristics and environmental characteristics are integrated to generate a comprehensive health status vector that reflects the current comprehensive operating status of the equipment: ,in, represents the comprehensive health state vector generated at time t, represents the electrical characteristic vector generated at time t, represents the mechanical vibration eigenvector generated at time t, Represents the environmental feature vector generated at time t.

6. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 1, characterized in that: The physical intervention test and feature enhancement extraction in S4 specifically include: S41, when the health status of the equipment is determined to be abnormal by comparing the comprehensive health status vector with the health status baseline, the physical intervention test process is automatically triggered to confirm and enhance the extraction of fault features; S42, vibration excitation application: applying external vibration excitation to the mechanical structure or vibration-sensitive parts of the equipment during the physical intervention test; S43, vibration signal response monitoring: through the mechanical vibration sensor installed on the equipment, the response characteristics of the equipment under vibration excitation are monitored in real time, and the vibration acceleration, frequency change and dynamic response behavior of the equipment are recorded; S44, extracting the response characteristics of the vibration signal again under the action of the vibration excitation signal to generate an enhanced mechanical vibration feature vector; S45, the enhanced mechanical vibration characteristic vector is compared with the healthy status baseline again to confirm the existence of a fault, and based on the response behavior characteristics of the equipment under vibration excitation, it is confirmed whether there is physical damage or structural problems.

7. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 6, characterized in that: Under the action of the vibration excitation signal, the enhanced vibration characteristics are extracted by Fourier transform frequency domain analysis method to generate an enhanced vibration feature vector, which includes the enhanced characteristics of the mechanical vibration signal and reflects the vibration frequency, acceleration change and response mode of the equipment: ,in, represents the enhanced vibration frequency characteristics, Represents the enhanced acceleration feature, which will enhance the vibration feature vector Vibration signature part of the equipment health baseline Compare and calculate the difference between the two to determine whether there is any abnormality or damage to the mechanical structure of the equipment: If the difference If the preset threshold is exceeded, combined with the preliminary difference calculation of S3, it is verified again that the device is abnormal.

8. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 7, characterized in that: The adaptive frequency adjustment specifically includes: Material property analysis: Based on the mechanical structure material properties of the equipment, obtain the natural frequency of the equipment : ,in: is the natural frequency of the device, is the material stiffness of the device, It is the quality of the equipment; Fault mode identification: Identify fault modes based on the historical fault data of the equipment. Fault modes include fatigue, cracks, and structural looseness. Different fault modes correspond to mechanical vibration characteristics at specific frequencies. By classifying historical fault modes, associate faults with their corresponding vibration frequency ranges. Adaptive Frequency Based on an optimization algorithm, the vibration excitation frequency is automatically selected to induce a specific failure mode.

9. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 8, characterized in that: The optimization algorithm specifically includes: Initialize Frequency Population: Generates a set of initial frequency values ​​within the frequency range of material properties and failure modes: ,in, is the initial frequency group; Fitness function evaluation: For each frequency value, the fitness function is used to evaluate its effectiveness in fault feature extraction. The fitness function can be defined based on the response characteristics of the material properties and the fault mode. The goal is to maximize the response signal strength of the fault feature extraction. ,in, represents the fitness function, Indicates frequency The extraction effect of fault features, ResonanceRisk represents the resonance risk, Indicates frequency The risk of causing equipment resonance, and is the weight factor, ; Crossover and mutation: Through the crossover and mutation operations in the genetic algorithm, the frequency population is updated to find a better frequency combination. The crossover operation combines frequency features to generate new frequencies, and the mutation operation introduces random disturbances to explore frequency values. Frequency selection: After multiple rounds of iterations, the optimal vibration excitation frequency is selected , that is, the frequency that can induce the fault characteristics and avoid resonance based on the material characteristics and failure mode, and use the selected optimal frequency , vibration excitation is applied by an electromagnetic vibrator or a piezoelectric vibrator.

10. A photovoltaic power station equipment fault diagnosis method based on a multi-modal large model according to claim 8, characterized in that: The failure mode identification includes analyzing historical failure data of the equipment to identify different failure modes; A failure mode library is established, which includes the characteristics of each failure mode and the vibration frequency range associated with it.

Images