Equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment

Abstract

The invention discloses an equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment, and relates to the technical field of equipment health management, the equipment fault intelligent early warning system comprises an equipment fault early warning platform, and the equipment fault early warning platform is in communication connection with the following modules: a data sensing fusion module, a voiceprint AI analysis module and a voiceprint AI analysis module; the data acquisition module is used for acquiring high-frequency voiceprint signals, temperature field distribution and vibration data in real time during operation of energy equipment through a distributed sensor network to form a comprehensive data set; and the voiceprint AI analysis module is used for extracting voiceprint features from the comprehensive data set, and identifying whether the equipment emits abnormal voiceprints or not by using a pre-trained voiceprint AI model. According to the method, early abnormity is identified through the high-precision AI model, sudden equipment faults are effectively prevented, equipment physical field interaction is simulated in combination with the digital twin technology, and a fault evolution path is dynamically deduced, so that operation and maintenance personnel can take intervention measures at the initial stage of the faults, the stability and reliability of equipment operation are remarkably improved, and the non-planned downtime is shortened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment health management, and in particular to an equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment. BACKGROUND

[0002] With the expansion of the scale and the improvement of the complexity of energy infrastructure, the demand for equipment reliability and operating efficiency in the modern energy industry is increasing. Once the equipment in the key energy fields such as power, oil and natural gas fails, not only will it result in high maintenance costs and downtime losses, but also may cause safety accidents and affect the stability of energy supply.

[0003] A new energy power station supervision and early warning control system (CN114583827A) processes the relevant data of the power station in the past to obtain corresponding classified data, and identifies and matches the real-time power station related data according to the classified data.

[0004] In the prior art, by identifying and integrating the related data of the power station, the correlation analysis of various data is carried out, the influence values of various influencing factors are processed, the safety analysis of the equipment in operation is carried out, and the normality of the equipment operation is determined through the correlation degree calculation and determination of the related data, thereby solving the problem that the safety analysis of the equipment cannot be carried out when the equipment is running, and the existing safety hazards are found in advance. However, since the energy equipment has a sub-healthy state when in use, there is a lack of effective characterization means, and when the temperature / vibration parameters exceed the standard, it has entered the irreversible damage stage, and there is a lack of fault evolution path modeling capability, so it is difficult for the operation and maintenance personnel to quickly locate the root cause. Therefore, how to combine the digital twin technology to embed the sound-heat-force multi-physical field coupling model, when the voiceprint AI detects an abnormality, automatically trigger virtual simulation deduction, dynamically predict the remaining life combined with the voiceprint feature change speed, and assist in fault early warning and positioning, is the problem to be solved by the present application. Therefore, the present application proposes an equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment. SUMMARY

[0005] The present application aims to provide an equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment to solve the problems raised in the background art.

[0006] To solve the above technical problems, the technical solution adopted by the present application is:

[0007] The equipment fault intelligent early warning system based on abnormal voiceprint AI analysis of energy equipment comprises an equipment fault early warning platform, and the equipment fault early warning platform is communicatively connected with the following modules, wherein:

[0008] A data perception fusion module is configured to collect high-frequency soundprint signals, temperature field distribution and vibration data of the energy equipment in real time through a distributed sensor network, and solve the time-space inconsistency problem of multi-physical field data by using timestamp synchronization and space registration technology to form a comprehensive data set;

[0009] A soundprint AI analysis module is configured to extract soundprint features from the comprehensive data set, and identify whether the equipment emits abnormal soundprint by using a pre-trained soundprint AI model;

[0010] A digital twin simulation deduction module is configured to construct a digital twin model of the equipment, deduce a fault evolution path when the soundprint AI detects an abnormality, and predict the remaining life according to the soundprint degradation trend;

[0011] A fault root cause positioning module is configured to fuse the simulation deduction result and the historical fault library, match the fault mode by using a graph neural network, analyze the soundprint features and physical field changes of different parts, determine the source of the fault, and locate the specific position of the equipment fault;

[0012] A three-dimensional visualization early warning module is configured to send early warning information to the operation and maintenance personnel, and present the internal damage evolution process of the equipment and the parameter threshold value through a visualization platform.

[0013] The further improvement of the technical scheme of the present application is that the data perception fusion module comprises:

[0014] A distributed sensor including a soundprint sensor, a temperature sensor and a vibration sensor is pre-deployed at a key node of the energy equipment to form a distributed sensor network, and the operation state of the equipment is monitored through the distributed sensor network to collect high-frequency soundprint signals, temperature field distribution and vibration data of the equipment, wherein the soundprint sensor captures the high-frequency soundprint signals of the energy equipment, the temperature sensor obtains infrared temperature field distribution data, and the vibration sensor collects vibration spectrum data;

[0015] The timestamp synchronization technology is used to mark the accurate time for the data collected at different times and by different sensors to ensure time consistency, and the space registration technology is used to unify the spatial coordinates of different sensor data to eliminate the spatial deviation caused by the position difference of the sensors and solve the time-space inconsistency problem of multi-physical field data;

[0016] The high-frequency soundprint signals, temperature field distribution and vibration data after time-space processing are integrated to remove redundant and error data, and the effective data is subjected to format unification and standardization processing to form a comprehensive data set containing multi-physical field information.

[0017] The further improvement of the technical scheme of the present application is that the soundprint AI analysis module comprises a soundprint feature extraction unit and an abnormal soundprint identification unit;

[0018] The voiceprint feature extraction unit extracts voiceprint features from the high-frequency voiceprint signals of the comprehensive data set by using signal processing technology, and determines normal voiceprint features in combination with a standard voiceprint mode library of each component of the energy equipment.

[0019] The abnormal voiceprint recognition unit uses a voiceprint AI model pre-trained based on a support vector machine to perform real-time analysis and judgment on the high-frequency voiceprint signals in combination with the extracted voiceprint features, judges whether the equipment emits abnormal voiceprints, identifies an early sub-health state, and determines the type of voiceprint abnormality.

[0020] The voiceprint feature extraction unit includes:

[0021] The high-frequency voiceprint signals in the comprehensive data set are subjected to noise reduction and filtering processing, a band-pass filter is used to retain the equipment characteristic frequency band and suppress environmental noise interference, a short-time Fourier transform (STFT) is used to convert the time-domain signals into time-frequency domain representations, the dynamic spectral characteristics of the voiceprint are extracted, and the high-frequency voiceprint signals are subjected to frame windowing processing to ensure the balance between time continuity and spectral resolution of each frame of data.

[0022] Based on the processed time-frequency domain representations, the mel-frequency cepstral coefficients (MFCC) of the voiceprint are extracted to characterize the spectral envelope features, the linear predictive coding (LPC) coefficients are combined to capture the sound channel response characteristics, the time-domain features including spectral centroid and frequency band energy are calculated to quantify the energy distribution and frequency center of gravity of the voiceprint, and the signal is decomposed by wavelet transform to extract multi-scale time-frequency detail features, thereby forming a multi-dimensional feature vector covering the frequency domain, time domain and scale space.

[0023] According to the design standards of each component of the energy equipment, the standard voiceprint features of each component of the energy equipment are determined, a standard voiceprint mode library is integrated and constructed, the standard voiceprint modes of each component of the energy equipment are determined, and the normal voiceprint features of each component of the energy equipment are synchronously labeled.

[0024] The abnormal voiceprint recognition unit includes:

[0025] The extracted multi-dimensional feature vector covering the frequency domain, time domain and scale space of the voiceprint features is preprocessed, and the voiceprint features are input into the voiceprint AI model pre-trained based on a support vector machine, the current voiceprint features are matched with the normal voiceprint features of the standard voiceprint mode library, the similarity score is calculated, and the similarity of the current voiceprint and the standard mode is quantified.

[0026] The decision boundary trained according to historical data is combined with a support vector machine classifier to classify the matching result, a decision boundary for distinguishing normal and abnormal states is divided, for a voiceprint deviating from a standard mode, an abnormal score is calculated through Mahalanobis distance, a deviation degree is quantified, a health level is evaluated in combination with a preset abnormal score threshold, and the health level is respectively normal, sub-health and abnormal.

[0027] In combination with a sliding window detection, matching results of three continuous fixed time periods are analyzed to determine whether an abnormality is persistent or deteriorating.

[0028] The further improvement of the technical scheme of the present application is that the digital twin simulation deduction module comprises a digital twin model construction unit and a virtual simulation deduction unit.

[0029] The digital twin model construction unit is configured to embed a digital twin model of a device of a sound-heat-force coupled finite element simulation model construction, to simulate physical field interaction under normal and abnormal states of the device.

[0030] The virtual simulation deduction unit is configured to automatically call the digital twin model of the device to deduce a fault evolution path when an abnormality is detected by the voiceprint AI, and to predict a remaining life of the device according to a voiceprint degradation trend in combination with a Weibull distribution.

[0031] The further improvement of the technical scheme of the present application is that the digital twin model construction unit comprises:

[0032] Based on geometric structures and material parameters of the energy device, a digital twin model of the device of a sound-heat-force coupled finite element simulation model construction is embedded in combination with digital twin technology, a multi-physical field partial differential equation set is constructed by integrating a sound field wave equation, a heat field heat conduction equation and a force field structural mechanics equation, the multi-physical field partial differential equation set is discretized and solved by a numerical solver, and a bottom simulation framework of physical behavior of the device is formed.

[0033] Real-time monitoring data (voiceprint, temperature and vibration) are mapped as dynamic input parameters of the simulation model, the simulation output is calibrated through data assimilation technology to reduce model error, and a state parameter database is constructed to record physical field distribution under different working conditions.

[0034] Based on the calibrated digital twin model, a physical field evolution process of the energy device under normal working conditions and preset abnormal modes is simulated to generate a simulation data set containing sound-heat-force multi-dimensional responses.

[0035] The further improvement of the technical scheme of the present application is that the virtual simulation deduction unit comprises:

[0036] When the voiceprint AI model detects that the device voiceprint feature deviates from the normal voiceprint feature and determines to be abnormal, a virtual simulation deduction is automatically triggered, a digital twin model matched with the real-time state of the energy device is called, current energy device geometric parameters, material properties and operating condition data are loaded, a high-fidelity simulation environment is provided for fault evolution deduction, and it is ensured that the deduction result is consistent with the actual physical behavior;

[0037] Based on the abnormal voiceprint feature, the corresponding fault mode is implanted in the digital twin model, the fault evolution path is dynamically deduced through multi-physical field coupling simulation (sound-heat-force interaction), the voiceprint change trend, stress distribution evolution and temperature field migration in the fault development process are output, and full-dimensional time sequence data of fault evolution are formed;

[0038] Combined with the voiceprint degradation trend data, i.e., the abnormal score time sequence, a device life degradation model is constructed using Weibull distribution, the degradation rate is quantified through parameter estimation, and based on the deduced fault terminal state, the residual life probability distribution of the device is reversely calculated, and the life expectancy and confidence interval are output.

[0039] The further improvement of the technical scheme of the present application is that the fault root cause positioning module comprises:

[0040] The simulation deduction result of the digital twin and the case features in the historical fault library are integrated, spatio-temporal alignment is performed, a graph structure is used to store the device topology relationship, nodes represent key components and edges represent physical coupling relationship, feature normalization and dynamic time warping are used to eliminate the influence of operating condition fluctuations, and it is ensured that the simulation data and the historical fault mode are comparable in the same measurement space;

[0041] A heterogeneous graph neural network is constructed, nodes are embedded with sound, heat and force multi-modal features, edge weights reflect the coupling strength of physical fields, and a message passing mechanism is used to aggregate neighborhood information, calculate the graph similarity between the current state and the historical fault mode, focus on the abnormal high-occurrence area through the attention mechanism, output the Top-3 candidate fault modes, and form a fault hypothesis set;

[0042] Based on the candidate fault mode, the multi-physical field equation is reversely solved, it is verified whether the sound-heat-force abnormal propagation path conforms to the simulation deduction trend, the physical field and spatial coordinates where the feature mutation occurs earliest are located through causal reasoning analysis, and finally a three-dimensional visual report of the fault source component, failure mechanism and influence range is output.

[0043] The further improvement of the technical scheme of the present application is that the three-dimensional visual warning module comprises:

[0044] The integrated real-time monitoring data and digital twin simulation results are mapped to the device three-dimensional geometric model through spatial registration technology, the voxel modeling method is used to convert the sound field energy distribution, temperature gradient and stress cloud into grid data with physical properties, and at the same time, the parameter threshold is displayed in the form of a semi-transparent isosurface, forming an interactive mixed reality scene;

[0045] Based on the time series simulation data, the device internal damage evolution animation is generated frame by frame, the heat flow / stress transmission path is displayed in combination with the stream chart, and the spatial coordinates and deterioration rate of the abnormal area are marked, so that the operation and maintenance personnel can backtrack the whole process of fault development through the time axis control, automatically focus on the highest risk area, and at the same time, the deviation degree of each parameter relative to the threshold is displayed in the form of a broken line chart, for comprehensive warning in time and space dimensions;

[0046] An alarm signal is generated and a three-dimensional diagnostic report is pushed, the three-dimensional diagnostic report includes fault source positioning mark, influence range heat map and treatment suggestion, all operation logs and diagnostic conclusions are automatically archived, and after-the-fact review and model optimization are supported.

[0047] Due to the adoption of the above technical solutions, the technical progress achieved by the present application relative to the prior art is:

[0048] 1. The present application provides a device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis, which identifies early abnormalities through a high-precision AI model, effectively prevents equipment sudden failure, simulates device physical field interaction through digital twin technology, dynamically deduces the fault evolution path, so that the operation and maintenance personnel can take intervention measures at the early stage of failure, significantly improve the stability and reliability of device operation, and reduce unplanned downtime.

[0049] 2. The present application provides a device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis, which realizes early detection of device sub-health state through voiceprint AI analysis, avoids irreversible damage caused by temperature / vibration parameter exceeding in traditional methods, reduces high maintenance cost and downtime loss, and at the same time, the digital twin simulation deduction function helps the operation and maintenance personnel to accurately locate the fault source, reduces the on-site troubleshooting time, and reduces the work intensity and safety risk of the operation and maintenance personnel. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0051] Figure 1 The workflow schematic diagram of the present application;

[0052] Figure 2 The workflow schematic diagram of the virtual simulation deduction unit of the present application. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0054] Embodiment 1, as shown in Figure 1 , Figure 2 The present application provides an equipment fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis, including an equipment fault early warning platform, and the equipment fault early warning platform is communicatively connected with the following modules, wherein:

[0055] The data perception fusion module is used for collecting high-frequency voiceprint signals, temperature field distribution and vibration data of the energy equipment in real time through a distributed sensor network, and solving the time-space inconsistency problem of multi-physical field data by using timestamp synchronization and space registration technology, forming a comprehensive data set. The distributed sensor including a voiceprint sensor, a temperature sensor and a vibration sensor is deployed in advance at the key nodes of the energy equipment to form a distributed sensor network, and the running state of the equipment is monitored through the distributed sensor network to collect the high-frequency voiceprint signals, temperature field distribution and vibration data of the equipment. The voiceprint sensor captures the high-frequency voiceprint signals of the energy equipment, the temperature sensor acquires infrared temperature field distribution data, and the vibration sensor collects vibration spectrum data. The timestamp synchronization technology is used to mark the accurate time for the data collected at different times and by different sensors, to ensure time consistency. At the same time, the space registration technology is used to unify the space coordinates of different sensor data, eliminate the space deviation caused by the difference in sensor positions, solve the time-space inconsistency problem of multi-physical field data, integrate the high-frequency voiceprint signals, temperature field distribution and vibration data after time-space processing, remove redundant and error data, format unify and standardize the effective data, and form a comprehensive data set containing multi-physical field information;

[0056] The data-aware fusion module specifically works as follows: a distributed sensor network is pre-deployed at key nodes of the energy equipment, covering voiceprint sensors, temperature sensors, and vibration sensors. The key nodes refer to areas with high failure rates, such as mechanical transmission components such as bearings, gearboxes, and couplings, electrical system vulnerable parts such as motor windings and insulation layers, and pressure-bearing components such as high-temperature and high-pressure pipelines and valves. The voiceprint sensor focuses on capturing high-frequency voiceprint signals generated during the operation of the energy equipment. The temperature sensor is responsible for obtaining infrared temperature field distribution data, reflecting the thermal operating state of the equipment. The vibration sensor collects vibration spectrum data to analyze the mechanical vibration characteristics of the equipment. The sensors work together to comprehensively collect multi-physical field data during the operation of the equipment. Time stamp synchronization technology is used to mark the precise time of data collected by different sensors at different times, ensuring that all data remains consistent in the time dimension. This allows the dynamic changes in the equipment operating state to accurately correspond to the time sequence. At the same time, spatial registration technology is used to unify the spatial coordinates of different sensor data, allowing different sensor data to be accurately expressed in the same spatial reference system and eliminating spatial deviations caused by different installation locations. The high-frequency voiceprint signals, temperature field distribution, and vibration data that have been time and space calibrated are cleaned to remove outliers and redundant information. The high-frequency voiceprint signals are filtered for noise based on spectral characteristics, the temperature field distribution data are identified for outliers using a sliding window statistical method, and the vibration data are decomposed for signals using wavelet transform to remove high-frequency noise, ensuring data effectiveness. The multi-source data are also standardized in format to conform to a unified data structure. The voiceprint, temperature, and vibration data are then integrated, and the data are associated by time stamp to generate equipment state snapshots, forming a comprehensive data set containing multi-physical field information.

[0057] The voiceprint AI analysis module is used to extract voiceprint features from the comprehensive data set and identify whether the equipment emits abnormal voiceprints using a pre-trained voiceprint AI model. The voiceprint AI analysis module includes a voiceprint feature extraction unit and an abnormal voiceprint identification unit.

[0058] The voiceprint feature extraction unit extracts voiceprint features from the high-frequency voiceprint signals of the comprehensive data set using signal processing techniques, determines normal voiceprint features in combination with the standard voiceprint mode library of each component of the energy equipment, performs noise reduction and filtering processing on the high-frequency voiceprint signals in the comprehensive data set, uses a band-pass filter to retain the equipment characteristic frequency band and suppress environmental noise interference, converts the time-domain signal into a time-frequency domain representation through short-time Fourier transform (STFT), extracts the dynamic spectral characteristics of the voiceprint, at the same time, performs frame windowing processing on the high-frequency voiceprint signals to ensure the balance between time continuity and spectral resolution of each frame of data, extracts the mel-frequency cepstral coefficient (MFCC) of the voiceprint based on the processed time-frequency domain representation to characterize the spectral envelope features, captures the sound channel response characteristics in combination with the linear predictive coding (LPC) coefficient, calculates the time-domain features including spectral centroid and frequency band energy, quantifies the energy distribution and frequency center of gravity of the voiceprint, and extracts multi-scale time-frequency detail features by wavelet transform to form a multi-dimensional feature vector covering the frequency domain, time domain and scale space, determines the standard voiceprint features of each component of the energy equipment according to the design standards of the energy equipment, integrates and constructs the standard voiceprint mode library, determines the standard voiceprint mode of each component of the energy equipment, and synchronously labels the normal voiceprint features of each component of the energy equipment.

[0059] The specific working content of the voiceprint feature extraction unit is: pre-processing the high-frequency voiceprint signal to eliminate environmental noise and retain the device characteristic frequency band, wherein a band-pass filter is used to filter out irrelevant frequency bands according to the acoustic characteristics of the target component, suppress low-frequency mechanical noise and high-frequency electromagnetic interference, and through frame windowing processing (frame length 20-40 ms, overlap rate 50%-75%, Hamming window), the continuous signal is divided into short-time stationary segments, ensuring the balance between time continuity and spectral resolution, based on the frame data, using short-time Fourier transform (STFT) to convert the time-domain signal into time-frequency domain representation, generating a dynamic spectrum diagram reflecting the time-varying characteristics of voiceprint energy; based on the time-frequency domain representation, three types of core features including spectral envelope features, time domain features and multi-scale time-frequency detail features are extracted, wherein the spectral envelope features are the calculation of mel frequency cepstral coefficients, the human ear hearing characteristics are simulated through the mel filter bank, the macroscopic envelope morphology of the spectrum is described, the linear predictive coding coefficients are combined to model the formant structure of the sound channel, and the physical state of the component is reflected; the time domain features include spectral centroid and frequency band energy, which quantize the energy distribution and frequency center of the voiceprint, and reflect the voiceprint shift caused by device load change or local overheating; the multi-scale time-frequency detail feature is to decompose the signal through wavelet transform to extract time-frequency details at different scales, capture transient impact or modulation phenomena, and finally form a multi-dimensional feature vector that integrates frequency domain, time domain and scale space; according to the design standards of each component of the energy equipment, the high-frequency voiceprint signals of each component of the energy equipment in the healthy state are collected, the standard feature vector is extracted, the standard voiceprint mode library is constructed for each component, and the component identification, working condition parameter, feature statistical quantity and version management are marked, wherein the component identification is the device position and type corresponding to the specific feature, the working condition parameter is the recording of the rotating speed, temperature, load and other operating conditions at the time of collection, ensuring that the mode library covers typical working conditions, the feature statistical quantity is the calculation of the mean and covariance matrix of the feature in the normal state, the version management is the regular update of the mode library to adapt to the aging of the device or the change of the working condition, and the historical version is retained for comparison and analysis, forming a structured and expandable standard voiceprint library to determine the normal voiceprint features of each component;

[0060] The abnormal voiceprint recognition unit, in combination with the extracted voiceprint features, uses a pre-trained voiceprint AI model based on a support vector machine to perform real-time analysis and judgment on the high-frequency voiceprint signal, judges whether the device emits an abnormal voiceprint, identifies an early sub-health state, determines the type of voiceprint anomaly, pre-processes the multi-dimensional feature vector of the extracted voiceprint features covering the frequency domain, time domain and scale space, and inputs the voiceprint features into a pre-trained voiceprint AI model based on a support vector machine, matches the current voiceprint features with the normal voiceprint features of the standard voiceprint pattern library, calculates the similarity score, quantifies the similarity of the current voiceprint with the standard pattern, classifies the matching results according to the decision boundary trained by the historical data using a support vector machine classifier, divides the decision boundary between normal and abnormal states, and for the voiceprint deviating from the standard pattern, calculates the abnormal score by Mahalanobis distance, quantifies the deviation degree, combines the pre-set abnormal score threshold, evaluates the health level, and divides it into normal health level, sub-health level and abnormal health level, and combines the sliding window detection to analyze the matching results of the continuous three fixed time periods to judge whether the abnormality is continuous or deteriorating.

[0061] The abnormal voiceprint recognition unit specifically performs the following operations: the extracted multi-dimensional feature vector (covering the frequency domain, time domain and scale space) is preprocessed, including normalization to eliminate dimensional differences, and dimension reduction by principal component analysis to remove redundant features and improve computational efficiency. The preprocessed feature vector is input into the voiceprint AI model pre-trained based on the support vector machine, the similarity between the current voiceprint feature and the normal voiceprint feature in the standard voiceprint pattern library is calculated, the matching degree of the current voiceprint and the standard pattern is quantified, and a similarity score is generated. The voiceprint signals of each component of the energy equipment in the normal operating state are collected, covering different working conditions and labeled normal voiceprint samples and abnormal voiceprint samples. If there is a lack of real fault data, abnormal samples can be generated by a synthesis method (adding Gaussian noise or spectral modulation). The over-sampling technique is used to handle the class imbalance problem to ensure that the positive and negative sample ratio is close to 1:1, forming a multi-condition voiceprint dataset, which is divided into a training set and a test set. The voiceprint features covering the frequency domain, time domain and scale space are spliced into a high-dimensional vector. The voiceprint AI model is trained using the training set and the basic architecture of the support vector machine. The radial basis function (RBF) is selected as the SVM kernel function. The hyperparameters, including the penalty coefficient and the kernel parameter, are optimized by grid search combined with 5-fold cross-validation to maximize the classification accuracy of the validation set. The class weight adjustment strategy is introduced to give higher weight to the minority fault class samples to improve the model's ability to recognize abnormal states. During the training process, the one-versus-all multi-classification strategy is used to decompose the multi-class problem into multiple binary classification sub-problems. The class label is determined by the voting mechanism. After training, the model performance is evaluated on the test set, the recall rate and F1 score of the fault class are analyzed to ensure that the missed detection rate is less than 5%. The trained voiceprint AI model is deployed for real-time voiceprint monitoring. The decision boundary trained using historical data is used to classify the similarity matching results based on the support vector machine classifier, and the threshold for distinguishing between normal and abnormal states is determined. For voiceprint features with a similarity score below the threshold, it is determined that they deviate from the standard pattern, and the abnormal score is calculated by the Mahalanobis distance to quantify the deviation from the normal voiceprint feature distribution center. Combined with the pre-set abnormal score threshold, different health levels are divided, including normal health level, sub-health level and abnormal health level. The sliding window mechanism is introduced to analyze the time sequence of the matching results of the continuous three fixed time periods. The time sequence change of the abnormal score is calculated by the sliding window to determine whether the abnormality is persistent or the trend is deteriorating. If the abnormal score exceeds the abnormal score threshold in the continuous three periods, or the abnormal score shows a continuous upward trend, it is determined as persistent abnormality and triggers an early warning. If the abnormal score fluctuates or continuously decreases, it is considered as transient interference, reducing the risk of false positives.

[0062] The calculation expression of the similarity score is as follows:

[0063] SimilarityScore=exp(-γ·D 2 (x,N));

[0064]

[0065] where SimilarityScore is the similarity score, the closer to 1 means the more similar to normal pattern, γ is the scaling factor, controlling the decay rate of the score with the increase of distance, D 2 (x, N) is the average kernel difference, the overall deviation of the current sample from the standard sample set, x is the feature vector of the current test sample, the real-time extracted voiceprint feature, N is the standard sample set, i.e., the normal pattern set, K is the number of standard samples, the total number of normal voiceprint feature vectors in the pattern library, n k is the d-dimensional feature vector of the kth standard sample, the kth normal sample in the standard pattern library, κ(x, n k ) is the kernel function value, measuring the local similarity between the current sample and the kth standard sample, ‖x-n k ‖ 2 is the squared Euclidean distance, the straight-line distance between two samples in the feature space, δ is the kernel bandwidth, controlling the neighborhood range of similarity judgment, 2δ 2 is the normalized denominator, scaling the distance square to a reasonable range, ‖n i -n j ‖ 2 is the distance square between standard samples, reflecting the sparsity of the normal sample itself, n i , n j are two different normal samples in the standard sample set N, median(·) is the median operation, eliminating the influence of extreme values, representing the distance between typical samples;

[0066] The calculation expression of the anomaly score is as follows:

[0067]

[0068] where D M (x) is the anomaly score, the larger the value, the farther the deviation from the normal state, represents the transpose of the vector, (x-μ) is a column vector, which is transposed into a row vector by μ is the mean value vector of the normal state, the statistical mean of the normal samples in the standard voiceprint library, S is the covariance matrix of the normal state, the correlation between the feature dimensions of the historical normal data, S -1 is the inverse of the covariance matrix, the weight matrix that eliminates the coupling effect between features, U is the number of normal samples, x u is the feature vector of the uth normal sample, 0<D M (x)<Y Z is the normal health level, within the historical normal fluctuation range, Y Z ≤D M(x) < Y Y is a sub-health level, early abnormal signs, D M (x) >= Y Y is an abnormal health level, significantly deviating from the normal pattern, wherein Y Z is the upper threshold of the normal health level and the lower threshold of the sub-health level, Y Y is the upper threshold of the sub-health level and the lower threshold of the abnormal health level;

[0069] A digital twin simulation deduction module is configured to construct a digital twin model of the device, deduce a fault evolution path when the voiceprint AI detects an anomaly, and then predict the remaining life according to the voiceprint degradation trend;

[0070] A fault root cause positioning module is configured to fuse the simulation deduction result and the historical fault library, match the fault mode through a graph neural network, analyze the voiceprint features and physical field changes of different parts, determine the source of the fault, and locate the specific position of the device fault;

[0071] A three-dimensional visualization early warning module is configured to send early warning information to the operation and maintenance personnel, and present the internal damage evolution process of the device and the parameter threshold through a visualization platform.

[0072] Embodiment 2, as shown in Figure 1 , Figure 2 The present application provides a technical solution based on embodiment 1: preferably, the digital twin simulation deduction module includes a digital twin model construction unit and a virtual simulation deduction unit;

[0073] The digital twin model construction unit is configured to embed a sound-heat-force coupled finite element simulation model to construct a digital twin model of the device, simulate the physical field interaction under normal and abnormal states of the device, embed a sound-heat-force coupled finite element simulation model to construct a digital twin model of the device based on the geometric structure and material parameters of the energy device, combine digital twin technology, integrate sound field wave equation, heat field heat conduction equation and force field structural mechanics equation to construct a multi-physical field partial differential equation set, discretize and solve the multi-physical field partial differential equation set through a numerical solver, form a bottom simulation framework of the physical behavior of the device, map real-time monitoring data (voiceprint, temperature and vibration) to dynamic input parameters of the simulation model, calibrate the simulation output through data assimilation technology, reduce model error, and construct a state parameter database to record the physical field distribution under different working conditions. Based on the calibrated digital twin model, the physical field evolution process of the energy device under normal working conditions and preset abnormal modes is simulated to generate a simulation data set containing sound-heat-force multi-dimensional responses;

[0074] The specific work content of the digital twin model construction unit is: based on the geometric structure parameters and material properties of the energy equipment, a finite element simulation framework of sound-heat-force coupling is established, a multi-physical field partial differential equation set is constructed by integrating the sound field wave equation, the heat field heat conduction equation and the force field structural mechanics equation, and the dynamic interaction behavior of the energy equipment under complex working conditions is described, wherein the sound field simulates the propagation and reflection of sound waves in the equipment structure through the wave equation, the heat field calculates the temperature distribution and heat flow transfer through the heat conduction equation, and the stress and strain distribution is analyzed based on the structural mechanics equation of the force field, and a numerical solver is used to discretize and solve the multi-physical field partial differential equation set (the Galerkin weighted residual method is used for spatial discretization of the multi-physical field partial differential equation set, and the Newmark-β method is used for dynamic solution), forming a full-dimensional simulation framework of the physical behavior of the energy equipment; the real-time monitoring soundprint, temperature and vibration signals of the energy equipment are mapped into dynamic input parameters of the simulation model through data assimilation technology, and the Kalman filtering algorithm is used for parameter inversion, taking the monitoring data as the observation value, iteratively updating the model boundary conditions and material parameters, minimizing the residual error between the simulation output and the actual monitoring data, improving the model accuracy, and synchronously constructing a state parameter database to record the sound field distribution, thermal gradient and stress concentration area inside the energy equipment under different working conditions, forming a two-way mapping relationship between historical and real-time data; based on the calibrated digital twin model, the physical field evolution process of the energy equipment under normal working conditions and preset abnormal modes is simulated, the abnormal characteristics are introduced through parameterized design, the progressive damage model is introduced, and a simulation data set containing sound-heat-force multi-dimensional response is generated, covering the whole process from early weak failure (sound emission signal frequency band expansion, local temperature rise) to serious failure (structure resonance, thermal runaway);

[0075] The expression of the sound field wave equation is as follows:

[0076]

[0077] In the formula, p is the sound pressure (a function of space and time), c0 is the sound speed in the medium, KM is the bulk modulus, ρ is the material density, β is the thermal-acoustic coupling coefficient, T is the temperature field distribution, F v is the viscous force term (simulating sound energy dissipation), is the spatial second derivative of sound pressure (describing sound wave diffusion), is the thermal acoustic source term (temperature change excites sound waves);

[0078] The expression of the heat field heat conduction equation is as follows:

[0079]

[0080] In the formula, C p is the specific heat capacity, k r is the thermal conductivity (tensor when anisotropic), a heat source for mechanical work, a heat source for acoustic energy dissipation, σ is a stress tensor, a strain rate tensor, denotes the change of internal energy with time, a heat conduction term (Fourier's law);

[0081] The expression of the force field structure mechanics equation is as follows:

[0082]

[0083] In the formula, u is a displacement vector field, σ X is a Cauchy stress tensor, is a fourth-order elastic stiffness tensor, ∈ is a strain tensor, F th is a thermal expansion force, α T is a thermal expansion coefficient, F ac is an acoustic radiation force;

[0084] When the voiceprint AI detects an anomaly, the virtual simulation deduction unit automatically calls the digital twin model of the device to deduce the fault evolution path, and combines the Weibull distribution to predict the remaining life of the device according to the voiceprint degradation trend. When the voiceprint AI model detects that the device voiceprint feature deviates from the normal voiceprint feature and determines that it is abnormal, it automatically triggers virtual simulation deduction, calls the digital twin model matching the real-time state of the energy device, loads the current energy device geometric parameters, material properties and operating condition data, provides a high-fidelity simulation environment for fault evolution deduction, ensures that the deduction result is consistent with the actual physical behavior, based on the abnormal voiceprint feature, implants the corresponding fault mode in the digital twin model, dynamically deduces the fault evolution path through multi-physical field coupling simulation (acoustic-thermal-force interaction), outputs the voiceprint change trend, stress distribution evolution and temperature field migration in the fault development process, forms the full-dimensional time sequence data of fault evolution, combines the voiceprint degradation trend data, that is, the abnormal score time sequence, constructs the device life degradation model using the Weibull distribution, quantifies the degradation rate through parameter estimation, and based on the deduced fault terminal state, reversely calculates the probability distribution of the remaining life of the device, and outputs the life expectancy and confidence interval;

[0085] The specific working content of the virtual simulation deduction unit is: when the voiceprint AI model detects that the device voiceprint feature deviates from the normal voiceprint feature and determines that it is abnormal, a virtual simulation deduction process is automatically triggered, a digital twin model matching the real-time state of the device is called according to the device identity (device code), it is ensured that the model geometric parameters, material properties and operating condition data are completely synchronized with the physical device, the device monitoring data is obtained in real time through the data interface, the model boundary conditions and initial state are dynamically updated, a high-fidelity simulation environment is constructed, a virtual test field consistent with the actual physical behavior is provided for fault evolution deduction, and deviation of the deduction result caused by model mismatch is avoided; based on the fault mode corresponding to the abnormal voiceprint feature, a parameterized fault model is implanted in the digital twin model, a sound-heat-force multi-physical field coupling simulation engine is activated, a multi-physical field partial differential equation group of a sound field wave equation, a heat field heat conduction equation and a force field structural mechanics equation is solved, and a fault evolution path is dynamically deduced, wherein the sound field part simulates the spectrum migration of the acoustic emission signal caused by the fault; the heat field part calculates the temperature field diffusion caused by local friction heat; the force field part analyzes the stress concentration area expansion, and finally outputs the full-dimensional time sequence data of fault evolution, including the voiceprint feature curve, the stress nephogram sequence and the temperature gradient migration graph; combined with the voiceprint degradation trend data (i.e. abnormal score time sequence), a device life degradation model is constructed by using Weibull distribution, the shape parameter (reflecting the degradation rate) and the scale parameter (characterizing the fault dispersion) are fitted by using the maximum likelihood estimation method, the degradation rate and the fault dispersion are quantified, based on the fault terminal state obtained by digital twin deduction, the residual life probability distribution of the device from the current state to failure is calculated through Monte Carlo simulation, and the life expectancy and confidence interval (life range under 95% confidence level) are output, wherein for the abnormal device detected by the voiceprint AI, based on the historical fault data and the voiceprint degradation trend (abnormal score time sequence), the applicability of Weibull distribution is verified, whether the data conforms to the two-parameter Weibull distribution is confirmed through probability graph test, if the data presents the typical characteristics of early failure (shape parameter <1), random failure (shape parameter ≈1) or wear-out failure (shape parameter >1), then Weibull distribution is selected as the basic framework of the life degradation model; the abnormal score time sequence is converted into equivalent life data, a likelihood function is constructed, combined with the probability density function of Weibull distribution, expressed as the product form of the parameter shape parameter and the scale parameter, the maximum value of the likelihood function is solved through Newton-Raphson method, the estimated value of the parameter is obtained, the model fitting degree is verified through K-S test, the accuracy of parameter estimation is ensured, the voiceprint degradation trend and the device life degradation are directly related, and a quantitative model is formed.The historical data is divided into a training set and a test set, the parameters are estimated through the training set, the prediction error is calculated on the test set to evaluate the generalization ability of the model, if the prediction error exceeds the threshold, the model form is adjusted or the parameter estimation method is optimized, the parameters of the equipment life degradation model are dynamically updated combined with the fault end state deduced by the digital twin in real time, the new monitoring data is integrated into the prior distribution through the Bayesian update framework, the posterior distribution of the shape parameter and the scale parameter is iteratively corrected, the model is adapted to the dynamic changes of the equipment operating state, and a trained equipment life degradation model is obtained;

[0086] Let the abnormal score time sequence of the voiceprint degradation trend data be t1, t2, …, t n (representing abnormal scores at different times, which can be regarded as a measure of the degree of equipment degradation), a two-parameter Weibull distribution is used to construct a device life degradation model, and a probability density function of the device life degradation model is:

[0087]

[0088] Beta and eta are fitted by the maximum likelihood estimation method, and the likelihood function is:

[0089]

[0090] In the formula, f (t i ; beta, eta) is the probability density function of the Weibull distribution, which represents the probability density of observing the data point t i under the parameters beta and eta, L (beta, eta) is the likelihood function, which represents the joint probability density of all observed data t1, t2, …, t n , and t i represents the abnormal score data point at the i-th time, that is, the degree of equipment degradation at the i-th time, and n represents the total number of observed abnormal score data points.

[0091] The likelihood function is taken as a logarithm and the partial derivative is taken, and the maximum likelihood estimates of beta and eta are obtained by setting the partial derivative to 0. And

[0092] Let the abnormal score corresponding to the fault end state deduced based on the digital twin be T f (that is, the critical value of equipment failure), and the abnormal score at the current time be T0. A large number of samples are randomly generated from the fitted Weibull distribution through Monte Carlo simulation to simulate the process of the equipment from the current state T0 to the failure state T f , and the probability distribution of the remaining life R of the equipment from the current state to failure is calculated, and the calculation formula of the life expectancy value is:

[0093]

[0094] In the formula, E(R) is the expected remaining life of the equipment from its current state to failure, representing the average remaining operating time of the equipment; η is the scaling parameter of the Weibull distribution, obtained through maximum likelihood estimation, characterizing the fault dispersion; β is the shape parameter of the Weibull distribution, obtained through maximum likelihood estimation, reflecting the degradation rate; Γ(·) is the gamma function; z is the independent variable of the gamma function; and t is the integral variable in the integral expression of the gamma function. When β increases, if β>1, the degradation rate accelerates and E(R) decreases; if 0<β<1, the degradation rate slows down and E(R) increases. When η increases, the fault dispersion increases, and E(R) increases.

[0095] For the Weibull distribution, let the 95% confidence interval for the remaining lifetime R be [R]. L R U The remaining lifetime R is sorted to obtain the sample sequence R. (1) R (2) , ..., R (m) (where m is the number of samples in the Monte Carlo simulation), then the lower bound R L With upper limit R U The calculation formula is:

[0096]

[0097] In the formula, R L R represents the lower limit of the 95% confidence interval for the remaining lifespan of the equipment from its current state to failure, indicating a 95% confidence that the remaining lifespan of the equipment will not be less than this value. U This represents the upper limit of the 95% confidence interval for the remaining lifespan of the equipment from its current state to failure, indicating a 95% certainty that the remaining lifespan of the equipment will not exceed this value. This represents the floor function. R represents the floor function, where the confidence interval narrows as the sample size m increases. U -R L This will decrease, and the estimate of remaining lifespan will be more accurate;

[0098] The root cause localization module includes:

[0099] The integrated digital twin simulation deduction result and the case features in the historical fault library are time and space aligned, the device topology relationship is stored by using a graph structure, a node represents a key component and an edge represents a physical coupling relationship, feature normalization and dynamic time warping are used to eliminate the influence of working condition fluctuations, to ensure that the simulation data and the historical fault mode are comparable in the same metric space, a heterogeneous graph neural network is constructed, the node embedding contains sound, heat and force multi-modal features, the edge weight reflects the coupling strength of the physical field, and the message passing mechanism is used to aggregate neighborhood information, the graph similarity of the current state and the historical fault mode is calculated, the attention mechanism is used to focus on the area with high abnormality, the Top-3 candidate fault modes are output, a fault hypothesis set is formed, based on the candidate fault modes, the multi-physical field equations are solved reversely, it is verified whether the abnormal propagation path of sound, heat and force is consistent with the simulation deduction trend, the earliest physical field with feature mutation and the spatial coordinates are located by using causal reasoning analysis, and finally a three-dimensional visual report of the fault source component, failure mechanism and influence range is output;

[0100] The specific working content of the fault root cause positioning module is as follows: the fault root cause positioning module first time and space aligns the digital twin simulation deduction result and the historical fault library, stores the device topology relationship by using a graph structure, a node represents a key component and an edge represents a physical coupling relationship, normalizes sound, heat and force multi-modal features, and uses dynamic time warping to eliminate the time sequence feature stretching problem caused by speed and load fluctuations, to ensure that the simulation data and the historical fault case are comparable in the same metric space, and then stores the time and space aligned data in a graph database to form a heterogeneous network containing component attributes, coupling strength and fault labels; based on the aligned data, a heterogeneous graph neural network is constructed, the node embedding fuses multi-physical field features, the edge weight is dynamically calculated from the physical field coupling strength, and the message passing mechanism (MPNN) is used to aggregate neighborhood information, the multi-head attention mechanism is used to focus on the area with high abnormality, and then the similarity (based on node embedding cosine similarity) between the current state graph and the historical fault mode graph is calculated, the Top-3 candidate fault modes are output, and a fault hypothesis set is formed; for the candidate fault modes, the sound-heat-force coupling equations are solved reversely, it is verified whether the abnormal propagation path is consistent with the simulation deduction trend, the causal relationship of the time sequence data of each physical field is analyzed by using causal reasoning, the earliest physical field with feature mutation and the spatial coordinates are located, and finally a three-dimensional visual report is generated, marking the fault source component, failure mechanism and influence range;

[0101] The calculation formula of the graph similarity is as follows:

[0102]

[0103] In the formula, Sim(G c ,G h ) is the graph similarity, is the node embedding cosine similarity, Let G be the neighborhood distribution JS divergence. c For the current state graph (built in real time), G h This is a historical failure mode diagram, where V is the set of all nodes in the diagram. This is the embedding vector of the current graph node v. Let be the embedding vector of the historical graph node v, λ be the weight coefficient (set to 0.7), and M be the neighborhood feature distribution of the current state graph node v. Distribution of neighborhood features of node v corresponding to the historical failure mode diagram The average distribution of is an intermediate reference distribution between two probability distributions, used for calculating the symmetric KL divergence. The distribution of neighborhood features (k-nearest neighbor statistics) of the current graph node v. Let α be the neighborhood feature distribution of historical graph node v. v The average of the multi-head attention weights for node v;

[0104] The 3D visualization early warning module includes:

[0105] Integrating real-time monitoring data with digital twin simulation results, the system maps multi-physics data onto the equipment's 3D geometric model using spatial registration technology. Employing voxel modeling, it transforms sound field energy distribution, temperature gradient, and stress cloud maps into mesh data with physical properties. Simultaneously, threshold values ​​for each parameter are overlaid as semi-transparent isosurfaces, creating an interactive mixed reality scene. Based on time-series simulation data, it generates frame-by-frame animations of internal equipment damage evolution, displaying heat flow / stress transmission paths using streamline diagrams and marking the spatial coordinates and deterioration rates of abnormal areas. Maintenance personnel can use the timeline control to trace the entire fault development process, automatically focusing the viewpoint to the highest-risk area. A line graph simultaneously displays the deviation of each parameter from the threshold, providing comprehensive spatiotemporal warnings, generating alarm signals, and pushing 3D diagnostic reports. These reports include fault source location markers, impact range heatmaps, and handling suggestions. All operation logs and diagnostic conclusions are automatically archived, supporting post-event review and model optimization.

[0106] The specific work content of the three-dimensional visualization early warning module is: through the spatial registration technology, the real-time monitored voiceprint, temperature, vibration data and the sound field, thermal field, force field data output by the digital twin simulation are time and space aligned, the data and the equipment geometric model are accurately matched, the voxel modeling method is used to convert the sound field energy distribution, temperature gradient and stress nephogram into three-dimensional grid data with physical properties, each voxel stores the numerical attribute of the corresponding physical quantity, each parameter threshold is dynamically superimposed on the model in the form of semi-transparent isosurface, forming an interactive mixed reality scene; based on the time series simulation data, the device internal damage evolution animation is generated frame by frame, the heat flow path and the stress transmission direction are presented in the form of dynamic streamline diagram, the spatial coordinates and the deterioration rate of the abnormal area are labeled in real time, the time axis control allows to backtrack the fault whole life cycle, and automatically focuses the visual angle to the highest risk area, the broken line chart synchronously displayed compares the deviation trend of the key parameters and the threshold, realizes the comprehensive early warning in time-space two dimensions, helps to quickly identify the abnormal root cause; generates an alarm signal and automatically generates a three-dimensional diagnostic report, including fault source positioning mark, influence range thermal diagram and disposal suggestion based on historical cases, the alarm signal is pushed to the operation and maintenance terminal in real time, triggers the emergency response process, all operation logs, diagnostic conclusions and simulation data are automatically archived to the knowledge base, supports multi-dimensional retrieval according to time stamp and fault mode.

[0107] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis, comprising a device fault early warning platform, characterized in that, The device fault early warning platform is communicatively connected with the following modules, wherein: A data perception fusion module is configured to collect high-frequency soundprint signals, temperature field distribution and vibration data of the energy equipment in real time through a distributed sensor network to form a comprehensive data set; A soundprint AI analysis module is configured to extract soundprint features from the comprehensive data set, and identify whether the equipment emits abnormal soundprints by using a pre-trained soundprint AI model; A digital twin simulation deduction module is configured to construct a digital twin model of the equipment, deduce a fault evolution path when the soundprint AI detects an abnormality, and predict the remaining life according to the soundprint degradation trend; A fault root cause positioning module is configured to fuse the simulation deduction result and the historical fault library, match the fault mode by using a graph neural network, analyze the soundprint features and physical field changes of different parts, and determine the source of the fault; A three-dimensional visualization early warning module is configured to send early warning information to the operation and maintenance personnel, and present the internal damage evolution process of the equipment and the parameter threshold values through a visualization platform.

2. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 1, characterized in that: The data perception fusion module comprises: A distributed sensor including a soundprint sensor, a temperature sensor and a vibration sensor is pre-deployed at key nodes of the energy equipment to form a distributed sensor network, and the operation state of the equipment is monitored through the distributed sensor network to collect high-frequency soundprint signals, temperature field distribution and vibration data of the equipment; A timestamp synchronization technology is used to mark the accurate time for the collected data at different times and by different sensors, and a spatial registration technology is used to unify the spatial coordinates of the data of different sensors; The high-frequency soundprint signals, temperature field distribution and vibration data after time and space processing are integrated, redundant and error data are removed, format unification and standardization processing are performed, and a comprehensive data set containing multi-physical field information is formed. 3.The energy source device anomaly voiceprint AI analysis-based device fault intelligent early warning system according to claim 2, characterized in that: The soundprint AI analysis module comprises a soundprint feature extraction unit and an abnormal soundprint identification unit; The soundprint feature extraction unit extracts soundprint features from the high-frequency soundprint signals of the comprehensive data set by using a signal processing technology, and determines normal soundprint features in combination with a standard soundprint mode library of each component of the energy equipment; The abnormal soundprint identification unit uses a soundprint AI model pre-trained based on a support vector machine to perform real-time analysis and judgment on the high-frequency soundprint signals in combination with the extracted soundprint features, judges whether the equipment emits abnormal soundprints, identifies early sub-health states, and determines the type of soundprint abnormalities.

4. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 3, characterized in that: The soundprint feature extraction unit comprises: The high-frequency soundprint signals in the comprehensive data set are subjected to noise reduction and filtering processing, a band-pass filter is used to retain the equipment characteristic frequency band, time-domain signals are converted into time-frequency domain representations by using short-time Fourier transform, dynamic frequency spectrum characteristics of the soundprint are extracted, and the high-frequency soundprint signals are subjected to frame windowing processing; Based on the processed time-frequency domain representations, mel-frequency cepstral coefficients of the soundprint are extracted to depict the frequency spectrum envelope characteristics, linear predictive coding coefficients are combined to capture sound channel response characteristics, time-domain features including spectral centroid and frequency band energy are calculated to quantify the energy distribution and frequency gravity center of the soundprint, signals are decomposed by using wavelet transform to extract multi-scale time-frequency detail features, and a multi-dimensional feature vector covering the frequency domain, time domain and scale space is formed; According to the design standards of each component of the energy equipment, the standard acoustic fingerprint features of each component of the energy equipment are determined, a standard acoustic fingerprint mode library is integrated and constructed, the standard acoustic fingerprint modes of each component of the energy equipment are determined, and the normal acoustic fingerprint features of each component of the energy equipment are synchronously marked.

5. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 3, characterized in that: The abnormal acoustic fingerprint recognition unit comprises: The extracted multi-dimensional feature vector of the acoustic fingerprint features covering the frequency domain, the time domain and the scale space is preprocessed, and the acoustic fingerprint features are input into the acoustic fingerprint AI model pre-trained based on the support vector machine, the current acoustic fingerprint features are matched with the normal acoustic fingerprint features in the standard acoustic fingerprint mode library, the similarity score is calculated, and the similarity of the current acoustic fingerprint and the standard mode is quantified; In combination with the support vector machine classifier, the matching results are classified according to the decision boundary trained based on the historical data, the decision boundary of the normal and abnormal states is divided, for the acoustic fingerprint deviating from the standard mode, the abnormal score is calculated by the Mahalanobis distance, the deviation degree is quantified, in combination with the preset abnormal score threshold, the health level to which the acoustic fingerprint belongs is evaluated, and the health level is divided into the normal health level, the sub-health level and the abnormal health level; In combination with the sliding window detection, the matching results of the continuous three fixed time periods are analyzed, and whether the abnormality is continuous or deteriorated is judged.

6. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 3, characterized in that: The digital twin simulation deduction module comprises a digital twin model construction unit and a virtual simulation deduction unit; The digital twin model construction unit is configured to embed a digital twin model of the device constructed by the acoustic-thermal-mechanical coupled finite element simulation model, and simulate the physical field interaction under the normal and abnormal states of the device. The virtual simulation deduction unit is configured to automatically call the digital twin model of the device to deduce the fault evolution path when the acoustic fingerprint AI detects the abnormality, and predict the remaining life of the device according to the acoustic fingerprint degradation trend in combination with the Weibull distribution.

7. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 6, characterized in that: The digital twin model construction unit comprises: Based on the geometric structure and material parameters of the energy equipment, the digital twin model of the device constructed by the acoustic-thermal-mechanical coupled finite element simulation model is embedded by combining the digital twin technology, the multi-physical field partial differential equation set is constructed by integrating the acoustic field wave equation, the thermal field heat conduction equation and the mechanical field structural mechanics equation, the multi-physical field partial differential equation set is discretely solved by a numerical solver, and a bottom simulation framework of the physical behavior of the device is formed; Real-time monitoring data are mapped as dynamic input parameters of the simulation model, the simulation output is calibrated by data assimilation technology, a state parameter database is constructed, and the physical field distribution under different working conditions is recorded; Based on the calibrated digital twin model, the physical field evolution process of the energy equipment under the normal working condition and the preset abnormal mode is simulated, and a simulation data set containing acoustic-thermal-mechanical multi-dimensional responses is generated.

8. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 6, characterized in that: The virtual simulation deduction unit comprises: When the acoustic fingerprint AI model detects that the acoustic fingerprint features of the device deviate from the normal acoustic fingerprint features and determines that the deviation is abnormal, the virtual simulation deduction is automatically triggered, the digital twin model matched with the real-time state of the energy equipment is called, and the current geometric parameters, material properties and operating condition data of the energy equipment are loaded. Based on abnormal acoustic signature features, corresponding failure modes are implanted in the digital twin model, and the failure evolution path is dynamically deduced through multi-physical field coupling simulation, outputting the acoustic signature change trend, stress distribution evolution and temperature field migration during the failure development process, forming a full-dimensional time sequence data of failure evolution; Combined with acoustic signature degradation trend data, i.e. abnormal score time series, a device life degradation model is constructed using Weibull distribution, the degradation rate is quantified through parameter estimation, and based on the deduced failure terminal state, the residual life probability distribution of the device is calculated in reverse, outputting the life expectancy and confidence interval.

9. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 6, characterized in that: The fault root cause positioning module includes: Integrate the digital twin simulation deduction results and the case features in the historical fault library, perform spatio-temporal alignment, use graph structure to store the device topology relationship, and use nodes to represent key components and edges to represent physical coupling relationship; Build a heterogeneous graph neural network, node embedding contains acoustic, thermal and force multi-modal features, edge weight reflects the coupling strength of physical fields, and use message passing mechanism to aggregate neighborhood information, calculate the graph similarity between the current state and historical failure modes, focus on the area with high abnormality through attention mechanism, output Top-3 candidate failure modes, form a fault hypothesis set; Based on the candidate failure modes, solve the multi-physical field equations in reverse, verify whether the acoustic-thermal-force abnormal propagation path conforms to the simulation deduction trend, combine the causal reasoning analysis to locate the earliest feature mutation of the physical field and spatial coordinates, and finally output a three-dimensional visualization report of the fault source component, failure mechanism and impact range.

10. The device fault intelligent early warning system based on energy equipment abnormal voiceprint AI analysis according to claim 9, characterized in that: The three-dimensional visualization warning module includes: Integrate real-time monitoring data and digital twin simulation results, map multi-physical field data to the device three-dimensional geometric model through spatial registration technology, use voxel modeling method to convert acoustic field energy distribution, temperature gradient and stress contour into grid data with physical properties, at the same time, the threshold values of each parameter are displayed in the form of semi-transparent isosurface, forming an interactive mixed reality scene; Based on time series simulation data, generate device internal damage evolution animation frame by frame, combine stream chart to show heat flow / stress transfer path, and label the spatial coordinates and deterioration rate of abnormal area, operation and maintenance personnel can backtrack the whole process of failure development through time axis control, automatically focus on the highest risk area, at the same time, display the deviation degree of each parameter relative to the threshold value in the line chart, and conduct comprehensive warning in time and space dimensions; Generate alarm signal and push three-dimensional diagnosis report, three-dimensional diagnosis report includes fault source positioning mark, impact range thermal map and disposal suggestion, all operation logs and diagnosis conclusions are automatically archived.

Images