A method and system for fault analysis and diagnosis of high-voltage inverters for mines

Abstract

The present invention relates to the technical field of fault diagnosis of power electronic equipment, and specifically discloses a fault analysis and diagnosis method and system for a high-voltage frequency converter for mining. The method collects the skin effect depth and parasitic parameter drift under high-frequency working conditions in real time through a multi-physics field sensor array; constructs the three-dimensional current density distribution of the conductor based on the non-Euclidean space grid partitioning technology, and dynamically corrects the conductor equivalent resistivity in combination with the metamaterial database to calculate the dynamic eddy current loss eigenvalue; utilizes the quantum annealing algorithm to optimize and solve the time-varying evolution equation of the parasitic parameter to obtain the voltage spike sensitivity eigenvalue; inputs the eigenvalue into a pre-trained deep residual shrinkage network diagnostic model for multimodal fusion analysis, and outputs the fault risk assessment result; finally, implements hierarchical protection according to the assessment result. The present invention innovatively solves the problem of modeling the time-varying characteristics of parasitic parameters in a vibration environment, and realizes early warning of hidden faults.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of power electronic equipment, and in particular to a fault analysis and diagnosis method and system for a high-voltage frequency converter for mining. Background Art

[0002] As the core power control device of mining equipment, high-voltage inverters for mining are subject to long-term operation in harsh environments with high temperature, high humidity, and strong vibration. Their failure rate is significantly higher than that of industrial environments. Traditional fault diagnosis methods mainly rely on the monitoring of conventional electrical parameters such as voltage and current, and use threshold alarms and simple logical judgment mechanisms, which have obvious technical limitations. Especially under high-frequency switching conditions, deep physical quantity changes such as local overheating of conductors caused by the skin effect and parasitic parameter drift caused by vibration are difficult to be effectively captured, resulting in problems such as delayed fault warning and high false alarm rates. The simulation methods based on finite element analysis in existing technologies usually use idealized models, which cannot accurately reflect the multi-physical field coupling effects in actual operation and perform poorly in predicting hidden faults such as voltage spikes and component aging.

[0003] The existing technology has the following deficiencies:

[0004] Continuous mechanical vibration in mining inverters causes deformation in the PCB wiring, leading to rapid drift in parasitic inductance and capacitance parameters at the μs level. Traditional frequency-domain impedance analysis methods, due to their insufficient sampling rate, cannot track these transient changes, while time-domain reflectometry is limited by signal interference in the complex electromagnetic environment of mines. This nonlinear drift in parasitic parameters can cause unpredictable voltage spikes, resulting in cumulative damage to IGBT modules.

[0005] The present invention uses a quantum tunneling probe to monitor parameter drift in real time, and combines the fractional-order time-varying evolution equation and quantum annealing optimization algorithm to achieve, for the first time, ns-level tracking accuracy of parasitic parameters in a vibration environment, solving the key technical bottleneck that has long plagued the improvement of the reliability of mining power electronic equipment. Summary of the Invention

[0006] The object of the present invention is to provide a method and system for analyzing and diagnosing faults of a high-voltage frequency converter for mining, so as to solve the above-mentioned problems.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A method for analyzing and diagnosing faults of a high-voltage frequency converter for mining comprises the following steps:

[0009] S1: During the operation cycle of the high-voltage inverter, real-time collection of operating status parameters of the high-voltage inverter is performed. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency working conditions;

[0010] S2: Based on the skin effect depth, a non-Euclidean spatial meshing technique is used to construct a three-dimensional current density distribution of the conductor cross section. The conductor equivalent resistivity is dynamically corrected using the metamaterial electromagnetic property database to calculate the dynamic eddy current loss characteristic value.

[0011] S3: Based on the parasitic parameter drift, a time-varying parameter evolution equation is established, and a quantum annealing algorithm is used to optimize and solve the real-time variation law of the parasitic inductance or capacitance, and the voltage spike sensitivity characteristic value is calculated;

[0012] S4: Input the dynamic eddy current loss characteristic value and the voltage spike sensitivity characteristic value into the pre-trained high-voltage inverter fault diagnosis model for analysis, and output the fault risk assessment result;

[0013] S5: Based on the fault risk assessment results, dynamically adjust the inverter operating parameters or trigger the hierarchical protection mechanism to achieve adaptive fault protection under high-frequency switching conditions.

[0014] As a further solution of the present invention: the three-dimensional current density distribution of the conductor cross section is constructed using the non-Euclidean space grid subdivision technology, specifically including:

[0015] A differential geometry model of the conductor surface is established based on Riemannian manifold theory, and an adaptive curvature mapping algorithm is used to map the conductor cross section into a non-Euclidean space.

[0016] Constructing a hexahedron-tetrahedron hybrid grid with a variable topology in the space, wherein the grid density in areas with significant skin effect is automatically refined to a sub-millimeter level;

[0017] The current density distribution is obtained by solving the Maxwell-Riemann equations, where the boundary conditions are dynamically modified using the equivalent magnetic permeability parameters in the metamaterial database.

[0018] As a further solution of the present invention: the dynamic correction process of the metamaterial electromagnetic property database includes:

[0019] Real-time collection of conductor surface temperature, vibration frequency and ambient humidity parameters;

[0020] A frequency-dependent impedance tensor model is established based on the measured data of the quantum magnetometer, and the anisotropic conductivity correction coefficient is obtained through tensor decomposition.

[0021] The correction coefficient is injected into the finite element solver, and the calculation is iterated until the relative error of the current density distribution is less than 1%.

[0022] As a further solution of the present invention: the calculation of the dynamic eddy current loss characteristic value includes:

[0023] The eddy current component in the skin layer is extracted from the three-dimensional current density distribution, and the loss power density equation is established using fractional calculus.

[0024] Calculate the instantaneous loss value of each conductor cross section through surface integral and superimpose the time domain integration results within the high-frequency switching cycle;

[0025] The vibration acceleration sensor data is introduced to dynamically compensate the calculation results and output the eddy current loss characteristic value corrected by environmental disturbance.

[0026] As a further solution of the present invention: the establishment of the time-varying parameter evolution equation includes:

[0027] The three-dimensional electromagnetic field distribution data of PCB wiring is collected in real time through terahertz time-domain spectrometer;

[0028] Construct a system of coupled differential equations involving parasitic inductance and capacitance, where the time-varying coefficients are dynamically updated by vibration acceleration sensor data;

[0029] Fractional-order derivatives are used to describe the relationship between conductor deformation and the memory effect of parasitic parameters, and a non-integer-order differential equation model is established, which is recorded as a time-varying parameter evolution equation.

[0030] As a further solution of the present invention: the optimization solution process of the quantum annealing algorithm includes:

[0031] The parasitic parameter identification problem is converted into the Ising model, and the design includes Hamiltonian encoding scheme for qubits;

[0032] Perform annealing operations in a quantum annealing processor and search for the global optimal solution through quantum tunneling effect;

[0033] The annealing results are post-processed using quantum error correction codes to output real-time change curves of parasitic inductance and capacitance.

[0034] As a further solution of the present invention: the calculation of the voltage spike sensitivity characteristic value includes:

[0035] Based on the parameter change curve output by the quantum annealing algorithm, a transient voltage response prediction model is constructed;

[0036] Analyze system stability through Lyapunov exponents and quantify the probability of voltage spikes;

[0037] A sensitivity grading standard is established based on historical fault data, and the normalized sensitivity characteristic value is output.

[0038] As a further solution of the present invention: the analysis process of the high-voltage inverter fault diagnosis model specifically includes:

[0039] Construct a deep residual shrinkage network based on multimodal feature fusion, where:

[0040] The first residual block receives the dynamic eddy current loss eigenvalue and eliminates the measurement noise through adaptive soft thresholding;

[0041] The second residual block receives the voltage spike sensitivity feature value and uses the channel attention mechanism to enhance the key frequency band features;

[0042] The outputs of the two residual blocks are fed into the spatiotemporal cross attention module to establish a coupling relationship model between eddy current loss and voltage spikes;

[0043] The cross-attention output is parsed through a quantum-derived decision tree to generate a triplet evaluation result containing fault type, location, and risk level;

[0044] The adversarial verification module is used to check the credibility of the evaluation results, and the feature recalculation mechanism is triggered when the confidence level is lower than the threshold.

[0045] As a further solution of the present invention: the dynamic adjustment of the inverter operating parameters or triggering of the hierarchical protection mechanism specifically includes:

[0046] Build an adaptive control strategy library based on reinforcement learning, which includes:

[0047] The first-level response targets low-risk faults and uses a fuzzy PID controller to dynamically adjust the carrier frequency and dead time;

[0048] The second level response targets medium-risk faults and activates the N+1 redundant switching mechanism of the power modules;

[0049] The third level response targets high-risk faults and executes an emergency shutdown sequence based on chaos synchronization;

[0050] The expected effects of each control strategy are simulated in real time through the digital twin system, and the optimal response plan is selected using the Bayesian optimization algorithm;

[0051] Implement nanosecond-level protection trigger logic at the FPGA hardware layer. When critical fault characteristics are detected, the protection action is directly executed, bypassing the main control system.

[0052] The protection decision-making process is recorded through blockchain smart contracts to establish an unalterable fault response traceability chain.

[0053] A fault analysis and diagnosis system for a high-voltage frequency converter used in a mine, comprising:

[0054] A multi-physics field real-time monitoring module, which is used to collect operating status parameters of the high-voltage inverter in real time during its operating cycle. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency conditions;

[0055] A three-dimensional electromagnetic field dynamic reconstruction module, which uses a non-Euclidean spatial meshing technique to construct a three-dimensional current density distribution of a conductor cross section based on the skin effect depth, dynamically corrects the conductor equivalent resistivity through a metamaterial electromagnetic property database, and calculates a dynamic eddy current loss characteristic value;

[0056] A parasitic parameter quantum optimization module, which establishes a time-varying parameter evolution equation based on the parasitic parameter drift, uses a quantum annealing algorithm to optimize and solve the real-time variation law of parasitic inductance or capacitance, and calculates the voltage spike sensitivity characteristic value;

[0057] a multimodal fault diagnosis module that inputs dynamic eddy current loss characteristic values ​​and voltage spike sensitivity characteristic values ​​into a pre-trained high-voltage inverter fault diagnosis model for analysis and outputs a fault risk assessment result;

[0058] An adaptive protection execution module dynamically adjusts the inverter operating parameters or triggers a hierarchical protection mechanism based on the fault risk assessment results to achieve adaptive fault protection under high-frequency switching conditions.

[0059] Beneficial effects of the present invention:

[0060] (1) The present invention achieves high-dimensional modeling and precise analysis of the complex electromagnetic behavior inside the high-voltage inverter by constructing a multi-physics field collaborative perception system and deeply integrating the conductor skin effect and the dynamic evolution characteristics of parasitic parameters under high-frequency working conditions. With the help of cutting-edge sensing technologies such as terahertz time-domain reflection and quantum tunneling effect, the system can accurately capture microscopic electromagnetic disturbances that are difficult to detect with traditional means, and reconstruct the three-dimensional current density distribution of the conductor cross section based on non-Euclidean space grid division and fractional-order differential model, and then dynamically calculate the eddy current loss characteristic value; at the same time, the quantum annealing algorithm and time-varying parameter evolution equation are introduced to achieve efficient solution of the real-time change law of parasitic inductance or capacitance and quantitative evaluation of voltage spike sensitivity. Furthermore, by integrating the intelligent diagnosis architecture of deep residual shrinkage network and quantum-derived decision tree, combined with adversarial verification mechanism and expert rule base, efficient fusion of multi-source heterogeneous features and high-confidence identification of fault modes are achieved, thereby significantly improving the sensitivity and robustness of early fault warning, and providing a solid guarantee for the stable operation of mining high-voltage inverters under complex working conditions.

[0061] (2) The present invention constructs a multi-level, adaptive protection control system based on fault risk assessment, deeply integrating reinforcement learning, chaos synchronization theory and hardware-level real-time protection mechanism, and realizes intelligent decision-making and rapid response of mining high-voltage inverters under high-frequency switching conditions. The system dynamically switches the control strategy according to the risk level of the diagnostic output: in the low-risk state, a fuzzy PID controller is used in combination with a deep reinforcement learning algorithm to optimize the operating parameters online, improve energy efficiency and suppress potential disturbances; in the medium-risk state, the N+1 redundant module seamless switching mechanism is used to achieve automatic load transfer and fault isolation to ensure the continuous and stable operation of the system; in the high-risk scenario, an emergency shutdown sequence based on the principle of chaos synchronization is activated to quickly guide the system state to converge to a safe area to prevent energy accumulation from causing secondary damage. At the same time, the FPGA hardware-level protection layer provides the last physical line of defense with a sub-microsecond response speed, ensuring that the source of danger can be reliably cut off under extreme fault conditions; the digital twin decision-making system verifies the feasibility and safety of various control strategies in advance through high-fidelity simulation modeling and Bayesian optimization algorithm, significantly reducing the probability of misoperation. This comprehensive protection architecture not only improves the robustness and availability of the system, but also provides solid support for the intelligent and autonomous operation and maintenance of high-voltage frequency conversion equipment in the complex electromagnetic environment of mines. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention will be further described below with reference to the accompanying drawings.

[0063] Figure 1 This is a flow chart of a method for analyzing and diagnosing faults of a high-voltage frequency converter for mining according to the present invention;

[0064] Figure 2 The present invention is a flowchart of a fault analysis and diagnosis system for a high-voltage frequency converter for mining. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] See also Figure 1 As shown, the present invention is a method for analyzing and diagnosing faults of a high-voltage frequency converter for mining, comprising the following steps:

[0067] S1: During the operation cycle of the high-voltage inverter, real-time collection of operating status parameters of the high-voltage inverter is performed. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency working conditions;

[0068] S2: Based on the skin effect depth, a non-Euclidean spatial meshing technique is used to construct a three-dimensional current density distribution of the conductor cross section. The conductor equivalent resistivity is dynamically corrected using the metamaterial electromagnetic property database to calculate the dynamic eddy current loss characteristic value.

[0069] S3: Based on the parasitic parameter drift, a time-varying parameter evolution equation is established, and a quantum annealing algorithm is used to optimize and solve the real-time variation law of the parasitic inductance or capacitance, and the voltage spike sensitivity characteristic value is calculated;

[0070] S4: Input the dynamic eddy current loss characteristic value and the voltage spike sensitivity characteristic value into the pre-trained high-voltage inverter fault diagnosis model for analysis, and output the fault risk assessment result;

[0071] S5: Based on the fault risk assessment results, dynamically adjust the inverter operating parameters or trigger the hierarchical protection mechanism to achieve adaptive fault protection under high-frequency switching conditions.

[0072] In S1, during the operation cycle of the high-voltage inverter, the operating state parameters of the high-voltage inverter are collected in real time. The operating state parameters include: skin effect depth and parasitic parameter drift under high-frequency conditions, specifically including:

[0073] At the beginning of each high-voltage inverter operation cycle, the system first initializes a multi-physics sensor array deployed at key nodes. This array consists of twelve distributed intelligent sensor nodes, each equipped with a high-frequency electromagnetic field probe, a vibration accelerometer, and a temperature sensor, connected to the central processing unit via fiber-optic Ethernet. The sensor nodes are arranged in an evenly spaced topology at key locations, such as the power busbar surface, the IGBT module heat sink substrate, and the DC bus capacitor bank. The installation spacing is determined by the golden ratio based on the conductor geometry to ensure comprehensive measurement coverage.

[0074] To measure the skin effect depth, the system uses terahertz time-domain reflectometry. Each sensor node transmits a terahertz pulse signal with a picosecond pulse width. This pulse exhibits characteristic attenuation as it propagates across the conductor surface. The receiver analyzes the amplitude attenuation and phase offset of the echo signal, combining it with a pre-stored database of material electromagnetic properties to calculate the equivalent skin depth under the current operating conditions in real time. During measurement, the system automatically compensates for the effect of ambient temperature on the propagation characteristics of the terahertz wave. The compensation coefficient is derived from a laboratory-calibrated temperature-dielectric constant curve.

[0075] Parasitic parameter drift is detected through the quantum tunneling effect. Nanoscale tunneling probes installed at each power device pin monitor the fluctuations in electron tunneling current to infer real-time changes in parasitic inductance and capacitance. The probes utilize a gallium nitride on diamond heterojunction structure, ensuring stable operation under high-temperature and high-pressure environments. The system samples the tunneling current waveform at microsecond intervals, extracts characteristic harmonic components through fast Fourier transform, and then uses a pre-calibrated conversion algorithm to determine parasitic parameter drift.

[0076] All sensor data is transmitted to the central processing unit via a dual-redundant fiber optic network. The transmission protocol utilizes the time-triggered Ethernet standard, ensuring strict synchronization of data between nodes with nanosecond-level time synchronization accuracy. The CPU's built-in self-diagnostic module verifies the operating status of each sensor in real time and automatically activates a backup sensing channel when an abnormal node is detected, ensuring continuous data collection.

[0077] During data preprocessing, the system first applies a sliding average filter to the raw data from each sensor, with the window width dynamically adjusted based on the current switching frequency. A wavelet transform algorithm is then used to eliminate pulse interference unique to mining environments, preserving valid characteristic frequency band information. The preprocessed data is categorized and stored in a ring buffer, with skin depth data marked with millisecond timestamps and parasitic parameter data stored synchronously by switching cycle.

[0078] During system operation, the human-machine interface displays dynamic curves and statistical characteristics of each measured parameter in real time. Operators can access detailed data records for any time period at any time and set alarm thresholds for key parameters. All collected data is simultaneously uploaded to a cloud database for subsequent fault diagnosis model training and system performance optimization.

[0079] In S2, based on the skin effect depth, a non-Euclidean spatial meshing technique is used to construct the three-dimensional current density distribution of the conductor cross section. The conductor equivalent resistivity is dynamically corrected using the metamaterial electromagnetic property database, and the dynamic eddy current loss characteristic value is calculated, specifically including:

[0080] First, the system initiates the conductor surface geometry modeling program. A high-precision 3D laser scanner acquires actual surface topography data for the power busbar and cable connector, achieving micron-level scanning resolution. Based on Riemannian manifold theory, this measured geometric data is converted into a differential geometry model. The model specifically annotates the microscopic concave and convex features of the conductor surface caused by long-term vibration wear. The system employs an adaptive curvature mapping algorithm, automatically adjusting the degree of parameterization based on the curvature radius of different conductor locations, accurately mapping complex surfaces into non-Euclidean space. During this process, the algorithm prioritizes preserving geometric details in areas of concentrated electric fields, such as conductor edges and connections.

[0081] After completing the spatial mapping, the system initiates the intelligent mesh generation process, employing a hybrid hexahedron-tetrahedron mesh strategy. Structured hexahedral meshes are used in the center of the conductor cross-section to ensure computational efficiency, while unstructured tetrahedral meshes are used in the surface area to accommodate complex geometries. Based on the skin depth values ​​monitored in real time, the system automatically implements mesh refinement within the skin effect affected area, with the minimum mesh size adjustable to 0.1 mm. During the mesh generation process, the algorithm dynamically evaluates the numerical discretization error and automatically triggers local mesh reconstruction when areas with drastic changes in the electric field gradient are detected.

[0082] The system solves the modified Maxwell-Riemann equations. The solver uses a finite element algorithm based on non-Euclidean space to expand the traditional electromagnetic field equations to a form that takes into account the geometric properties of curved surfaces. Boundary conditions are set based on a database of metamaterial electromagnetic properties, which stores equivalent electromagnetic parameters for various conductive materials under different temperature and frequency conditions. During the solution process, the system monitors residual convergence in real time and automatically activates artificial damping technology when oscillation is detected. After completing the field distribution calculation, the system generates a three-dimensional current density cloud map containing vector direction and amplitude information, and displays it to the operator in a visual form.

[0083] During the dynamic correction phase of the metamaterial database, the system first collects real-time environmental parameters from multiple sensor nodes. Temperature data comes from thin-film platinum resistors embedded in the conductor surface, vibration data is provided by a three-axis MEMS accelerometer, and ambient humidity is monitored by a capacitive sensor. These parameters are fed into the quantum magnetometer's control system to adjust the compensation parameters of the measurement circuit. The quantum magnetometer collects the microscopic magnetic field distribution on the conductor surface at a rate of 1000 times per second, inverting the frequency-dependent impedance characteristics through inverse problem solving. The system uses high-order tensor decomposition technology to extract characteristic components related to anisotropic conductivity from the impedance data and generate corresponding material correction coefficients.

[0084] The finite element solver's parameter updates utilize an incremental injection method. A dual buffering mechanism maintains the stability of the current calculation while preparing a new set of material parameters in the background. When a parameter update is triggered, the system retains 50 percent of the previous calculation results as initial values, gradually transitioning to the solution under the new parameters. During the iteration process, the algorithm simultaneously compares the relative errors of multiple physical quantities, including current density amplitude error, phase angle error, and energy conservation error. Convergence is considered complete only when all three meet a threshold of less than 1 percent.

[0085] To extract eddy current components, the system first delineates the skin effect region within the 3D field distribution data. The thickness of this region is determined to be 1.5 times the measured skin depth to ensure complete coverage of the eddy current activity. The system employs an improved vector decomposition algorithm to separate the total current density into two orthogonal components: conduction current and eddy current. Eddy current components are identified based on their unique spatial distribution and phase characteristics. During the extraction process, the algorithm automatically filters out spurious eddy current components caused by numerical discretization.

[0086] The loss power density is calculated using a fractional-order differential model. The system automatically selects the optimal differential order based on material properties and operating frequency, typically adjusting dynamically between 0.5 and 1.5. During the calculation, the algorithm considers the fractal dimensions of the eddy current paths and accurately evaluates non-integer-order differential terms using a recursive subdivision algorithm. The system employs an adaptive integration strategy, using larger step sizes in areas of gradual loss variation and automatically increasing sampling points in areas of drastic change.

[0087] The integration calculation of instantaneous loss values ​​utilizes a dual-integration scheme in both time and space. In the spatial dimension, the system performs a surface integral on each conductor cross section. The integration region is automatically divided into multiple subdomains based on the current density distribution for parallel calculation. In the temporal dimension, the integration covers the entire high-frequency switching cycle, and the system intelligently identifies key characteristic points of the PWM waveform as integration nodes. To improve calculation accuracy, the sampling density is automatically increased near current zero crossings and peaks.

[0088] The environmental disturbance compensation module receives real-time data from vibration sensors. The system establishes an empirical model of vibration frequency and eddy current loss correction factors, trained using extensive experimental data. The compensation algorithm analyzes the spectral characteristics of the vibration signal, identifies the primary vibration modes, and then adjusts the loss calculations based on pre-defined correspondences. For abnormal vibrations outside the model range, the system triggers an expert system for special processing and automatically adds the new case to the training database.

[0089] The final output of the dynamic eddy current loss characteristic value contains multiple dimensions of information: basic loss value, temperature influence coefficient, vibration correction factor, and confidence index. This data is transmitted to the fault diagnosis module in a standardized data structure and displayed in real time as a trend curve on the human-machine interface. The system continuously monitors the rate of change of the characteristic value and automatically triggers a data review mechanism when an abnormal sudden change is detected to ensure the reliability of the output results.

[0090] In S3, based on the parasitic parameter drift, a time-varying parameter evolution equation is established, and a quantum annealing algorithm is used to optimize and solve the real-time variation law of the parasitic inductance or capacitance, and the voltage spike sensitivity characteristic value is calculated, specifically including:

[0091] During the development phase of the time-varying parameter evolution equations, the system first deploys a terahertz time-domain spectroscopy measurement system. The system's transmitter uses a fiber-coupled femtosecond laser to generate terahertz pulses with a pulse width of less than 100 femtoseconds. These pulses are focused onto the PCB trace area under test via a quasi-optical system. The receiver utilizes electro-optical sampling technology to record the reflected waveform with sub-picosecond temporal resolution. During the measurement process, the system sets up multiple scanning points along the trace path, capturing a 50-picosecond time window at each point. A complete electromagnetic field distribution image is reconstructed using a three-dimensional interpolation algorithm. To eliminate environmental interference, the system employs a nitrogen purge optical path and performs multiple sampling averages.

[0092] When constructing the coupled differential equations, the system first extracts the original parasitic parameters from the circuit layout file. The initial inductance matrix is ​​calculated using the partial element equivalent circuit method, while the capacitance matrix is ​​solved using the boundary element method. These static parameters are then imported into the dynamic equations as baseline values. The vibration accelerometer data is processed through a digital filter to generate time-varying coefficient update instructions. The system establishes a nonlinear mapping relationship between vibration amplitude and parasitic parameter changes, which was obtained through previous vibration table test calibration. During the solution of the equations, the system uses the variable step-size Runge-Kutta method, automatically reducing the integration step size when a sudden change in the parameters is detected.

[0093] The fractional derivative model is implemented using a memory kernel method. The system allocates a separate memory window for each parasitic parameter, and the window length is dynamically adjusted based on the material's fatigue properties. For PCB substrate materials, a typical memory window is on the order of 10 seconds; for solder joints, a shorter window of 3 seconds is used. The derivative order is determined based on material microstructural analysis results, and the system includes a built-in library of fractional-order reference values ​​for common materials. During the calculation process, the algorithm automatically records historical parameter changes to update the weight distribution of the memory kernel function.

[0094] During the implementation phase of the quantum annealing algorithm, the system first transforms the problem. Each parasitic parameter is discretized into 256 possible values, corresponding to the eight qubits of the quantum processor. The Hamiltonian construction takes into account the electromagnetic coupling effects between adjacent wiring and the additional constraints introduced by device packaging. The system uses a penalty term design that embeds prior knowledge to ensure physical plausibility. The encoding scheme is specifically optimized to ensure that the qubits corresponding to key parameters are located on the critical path of the processor topology, improving computational efficiency.

[0095] Annealing is performed in a superconducting quantum processor in a dilution refrigerator. The system employs a segmented annealing strategy, rapidly reducing the temperature in the initial phase to overcome local minima, followed by a slower convergence to improve accuracy. During annealing, the coherence time of the qubits is monitored in real time, and dynamic decoupling pulses are automatically inserted when decoherence risk is detected. Each annealing operation is repeated 1024 times, and the probability distribution of the measurement results is statistically analyzed. The top three solutions with the highest frequency of occurrence are selected as the candidate solution set.

[0096] Quantum error correction utilizes a surface code scheme. The system allocates nine physical qubits of error correction resources to each logical qubit. The error correction decoder, based on a minimum weight matching algorithm, corrects bit and phase flip errors that occur during measurement in real time. In the post-processing phase, the system maps the raw output of the quantum processor back into the physical parameter space, eliminating anomalous solutions through confidence interval testing, and ultimately outputs parameter combinations that satisfy the fundamental laws of the circuit.

[0097] The transient voltage response prediction model is constructed using the transmission line theory framework. The system divides the entire power loop into multiple microsegments, with parameters for each segment derived from the latest results of quantum annealing. The model specifically considers wave propagation effects at high frequencies and employs the finite-difference time-domain method. To account for nonlinear components, the system incorporates the Newton iteration method for local linearization. Boundary conditions are set based on the actual measured drive signal waveform, including detailed characteristics of the rising and falling edges.

[0098] Lyapunov exponent analysis is performed on a dedicated computing accelerator card. The system first establishes a state-space equation, using voltage fluctuations, current changes, and temperature drift as key state variables. Using the QR decomposition algorithm, the eigenvalues ​​of the Jacobian matrix are calculated, tracking the evolution of the maximum Lyapunov exponent. During the analysis, the system identifies strange attractor structures in the phase space and assesses the likelihood of the system entering a chaotic state. The calculated results are quantified as a stability coefficient between 0 and 1, reflecting the potential risk of voltage spikes.

[0099] The sensitivity grading standard was established based on big data analysis. The system collected and annotated failure cases of similar inverters over the past five years, extracting parameter change characteristics preceding each failure. Using a machine learning algorithm, a classification model incorporating 12 key features was established. The grading process utilizes fuzzy logic, taking into account multiple factors such as the absolute value, rate of change, and duration of the parameter. The resulting sensitivity characteristic values ​​are scaled to five levels, corresponding to normal, caution, warning, danger, and emergency status. Each status is associated with clear action recommendations.

[0100] In S4, the dynamic eddy current loss characteristic value and the voltage spike sensitivity characteristic value are input into the pre-trained high-voltage inverter fault diagnosis model for analysis, and the fault risk assessment results are output, including:

[0101] The deep residual shrinkage network is constructed using a dual-branch parallel architecture. The first residual block specifically processes dynamic eddy current loss eigenvalues. Its core innovation lies in the design of an adaptive soft thresholding layer. This layer automatically adjusts the shape of the threshold curve based on the statistical properties of the input features, applying a steep suppression curve for small-amplitude noise components while maintaining a smooth transition for valid signal components. The threshold parameter is dynamically calculated based on the root mean square amplitude of the eigenvalue itself, achieving an optimal balance between noise suppression and signal preservation. Residual connections utilize a cross-layer weighting mechanism, assigning fusion weights to feature maps at different depths based on their contribution to the final diagnosis.

[0102] The second residual block processes the voltage spike sensitivity eigenvalues. Its core is an improved channel attention module. This module first performs a multi-scale decomposition of the input features to extract feature subbands containing information from different frequency bands. It then calculates the importance score for each frequency band using a learnable weight matrix, prioritizing the feature channels with high correlation with typical fault modes. A temperature-adjusted softmax function is introduced during the attention weight generation process to avoid excessive focus on a single feature dimension. Local response normalization is also performed before the module output to maintain feature scale stability.

[0103] The innovation of the spatiotemporal cross-attention module lies in its modeling of dynamic correlations between multiple physical quantities. This module first rearranges the output features of the two residual blocks into a spatiotemporal cube data structure, where the time dimension corresponds to the sampling sequence and the spatial dimension corresponds to the topological relationship of the inverter's physical locations. A separable convolution operation is used to calculate a cross-modal similarity matrix, identifying the potential causal relationship between eddy current loss changes and voltage spikes. The module also incorporates a built-in memory mechanism that tracks the evolution of key features, enhancing the ability to identify progressive faults.

[0104] The quantum-derived decision tree is implemented using a superconducting quantum processor. This tree structure transforms the judgment conditions of traditional decision nodes into probability measurements under quantum state superposition. Each decision node corresponds to a carefully designed sequence of quantum gates, which selectively rotate qubits to achieve nonlinear partitioning of the feature space. The tree's depth expansion is automatically optimized using quantum amplitude amplification technology to avoid overfitting. The final classification results are obtained through quantum state tomography, outputting the probability distribution of different fault categories.

[0105] The adversarial verification module consists of two components: a generator and a discriminator. The generator, based on a variational autoencoder architecture, learns the distribution of feature vectors under normal operating conditions. The discriminator, a multilayer perceptron network, is responsible for evaluating the degree of deviation between input features and generated samples. During verification, the system calculates the reconstruction error of the original features in latent space and combines this with the discriminator's output score to comprehensively judge the credibility of the diagnostic result. When an anomaly is detected, the system prioritizes checking the integrity of the sensor data and, if necessary, re-triggers the feature calculation process.

[0106] The fault risk assessment results are generated using a multi-level fusion strategy. The system first compares the output of the quantum decision tree with the expert rule base to eliminate obvious logical contradictions. It then uses historical equipment maintenance records to perform a weighted assessment of the recurrence risk of similar faults. The resulting structured report includes information on three dimensions: estimated fault probability, potential impact, and recommended remedial measures. The report uses a standardized JSON format for easy integration into the factory's asset management system.

[0107] In S5, based on the fault risk assessment results, the inverter operating parameters are dynamically adjusted or the hierarchical protection mechanism is triggered to achieve adaptive fault protection under high-frequency switching conditions. Specifically, the following are included:

[0108] The reinforcement learning adaptive control strategy library is constructed using a layered architecture. The first-level response is for low-risk faults with a risk assessment value below 0.3, where the system initiates the dynamic parameter adjustment mechanism of the fuzzy PID controller. This controller receives the fault feature vector as input in real time, maps it to the fuzzy rule space via a membership function, and outputs the carrier frequency adjustment and dead-time correction values. The control rule library contains 256 empirical rules and is implemented using a TSK fuzzy inference system. The parameter adjustment process follows the principle of gradual change, with each adjustment not exceeding 5% of the baseline value to avoid secondary impacts on the system. The controller has a built-in self-learning module that dynamically optimizes rule weights based on the adjustment results.

[0109] The second-level response handles medium-risk faults with a risk assessment value between 0.3 and 0.7. The system first identifies the power module group corresponding to the fault location and then initiates the N+1 redundant switching process. The switching process uses pre-charge synchronization technology, and the standby module completes DC bus voltage tracking through a current-limiting resistor before switching in. After the switching command is issued, the system monitors the output current difference between the new and old modules and automatically adjusts the PWM phase compensation when it detects that the circulating current exceeds the threshold. The entire switching process is completed within three switching cycles, and the output voltage fluctuation is controlled within 2% of the rated value. After the redundant module is switched in, the system automatically initiates a 48-hour aging monitoring period to record its performance degradation curve.

[0110] The third-level response targets high-risk faults with a risk assessment value exceeding 0.7. The system uses the principle of chaotic synchronization to design an emergency shutdown sequence, quickly activating a specific form of disturbance signal to enable the system to quickly exit a dangerous state. The shutdown process is divided into three stages: first, the carrier frequency is reduced to 1 / 10 of the fundamental frequency within 1 millisecond, then an inverted braking voltage is applied, and finally, the drive signal is completely shut down. The parameters of the chaos controller are dynamically adjusted based on the real-time operating conditions to ensure that the system state variables converge to a safe zone in the shortest possible time. During the shutdown process, the system continuously monitors the changing trajectory of key parameters and, when necessary, injects auxiliary damping signals to prevent oscillations.

[0111] The digital twin decision-making support system runs on a separate real-time simulation server. It builds a high-fidelity virtual model encompassing circuit, thermal, and mechanical coupling, enabling short-term predictions of various control strategies with a 10-microsecond step size. During simulation, the system pays particular attention to the evolving trends of key indicators such as IGBT junction temperature and DC bus voltage ripple. A Bayesian optimization algorithm uses Gaussian process regression to construct a proxy model for the objective function, balancing exploration and exploitation through acquisition functions. The optimization results include evaluation data on three dimensions: expected effect score, risk index, and execution priority.

[0112] The FPGA hardware protection layer utilizes a triple-mode redundant design to ensure reliability. The protection logic continuously monitors raw signals from the high-speed ADC, including bus voltage, output current, and gate drive status. When a voltage spike exceeding a safety threshold or an abnormal current rise rate is detected, the protection circuit initiates an emergency procedure within 80 nanoseconds. This process is completely independent of the main control system and directly operates the power device's gate driver IC via a dedicated optocoupler channel. Three protection modes are available: immediate soft shutdown, active clamping, and rapid discharge, automatically selected based on fault severity. Once a protection event is triggered, the system locks the current status data for subsequent analysis.

[0113] The blockchain evidence storage system is deployed using a consortium chain architecture. After each protection decision is generated, the system automatically extracts its key features into standardized event descriptors, including timestamp, fault type, response measures, and effectiveness evaluation. This data is hashed and written to the blockchain, along with the Merkle tree root of the relevant sensor data. Smart contracts implement a multi-level signature verification mechanism, ensuring that only authorized maintenance personnel can add comments or update status. The blockchain explorer provides an intuitive visual interface and supports multi-dimensional searches by time, device ID, or fault type.

[0114] See also Figure 2 As shown, a fault analysis and diagnosis system for a high-voltage frequency converter for mining includes:

[0115] A multi-physics field real-time monitoring module, which is used to collect operating status parameters of the high-voltage inverter in real time during its operating cycle. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency conditions;

[0116] A three-dimensional electromagnetic field dynamic reconstruction module, which uses a non-Euclidean spatial meshing technique to construct a three-dimensional current density distribution of a conductor cross section based on the skin effect depth, dynamically corrects the conductor equivalent resistivity through a metamaterial electromagnetic property database, and calculates a dynamic eddy current loss characteristic value;

[0117] A parasitic parameter quantum optimization module, which establishes a time-varying parameter evolution equation based on the parasitic parameter drift, uses a quantum annealing algorithm to optimize and solve the real-time variation law of parasitic inductance or capacitance, and calculates the voltage spike sensitivity characteristic value;

[0118] a multimodal fault diagnosis module that inputs dynamic eddy current loss characteristic values ​​and voltage spike sensitivity characteristic values ​​into a pre-trained high-voltage inverter fault diagnosis model for analysis and outputs a fault risk assessment result;

[0119] An adaptive protection execution module dynamically adjusts the inverter operating parameters or triggers a hierarchical protection mechanism based on the fault risk assessment results to achieve adaptive fault protection under high-frequency switching conditions.

[0120] The working principle of the present invention is as follows: the present invention uses a multi-physics field sensor array to collect the skin effect depth and parasitic parameter drift under high-frequency working conditions in real time; based on the skin effect depth, the non-Euclidean space grid partitioning technology is used to construct the three-dimensional current density distribution of the conductor, and the dynamic eddy current loss characteristic value is calculated by combining the metamaterial database to dynamically correct the resistivity; based on the parasitic parameter drift, a time-varying parameter evolution equation is established, and the quantum annealing algorithm is used to solve the parasitic parameter change law to obtain the voltage spike sensitivity characteristic value; the above characteristic value is input into the pre-trained deep residual shrinkage network diagnostic model, and the fault risk assessment result is output through multi-modal feature fusion and quantum-derived decision tree analysis; finally, according to the evaluation results, the reinforcement learning strategy library is used to implement hierarchical protection, including fuzzy PID parameter adjustment, power module redundant switching and chaotic synchronous emergency shutdown, and the response plan is optimized through digital twin simulation, FPGA hardware realizes nanosecond-level protection, and blockchain technology ensures that the decision is traceable, so as to achieve precise fault protection under high-frequency switching conditions.

[0121] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for analyzing and diagnosing faults of a high-voltage frequency converter for mining, characterized in that: The following steps are involved: S1: During the operation cycle of the high-voltage inverter, real-time collection of operating status parameters of the high-voltage inverter is performed. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency working conditions; S2: Based on the skin effect depth, a non-Euclidean spatial meshing technique is used to construct the three-dimensional current density distribution of the conductor cross section. The conductor equivalent resistivity is dynamically corrected using the metamaterial electromagnetic property database to calculate the dynamic eddy current loss characteristic value, specifically including: Establish a differential geometry model of the conductor surface and map the conductor cross section into a non-Euclidean space; Construct hexahedral-tetrahedral hybrid meshes with variable topology in non-Euclidean space; The three-dimensional current density distribution is obtained by solving the Maxwell-Riemann equations; The calculation of the dynamic eddy current loss characteristic value includes: The eddy current component in the skin layer is extracted from the three-dimensional current density distribution, and the loss power density equation is established using fractional calculus. Calculate the instantaneous loss value of each conductor cross section through surface integral and superimpose the time domain integration results within the high-frequency switching cycle; The vibration acceleration sensor data is introduced to dynamically compensate the calculation results and output the eddy current loss characteristic value corrected by environmental disturbance; S3: Based on the parasitic parameter drift, a time-varying parameter evolution equation is established, and a quantum annealing algorithm is used to optimize and solve the real-time variation law of the parasitic inductance or capacitance, and the voltage spike sensitivity characteristic value is calculated; The establishment of the time-varying parameter evolution equation includes: The three-dimensional electromagnetic field distribution data of PCB wiring is collected in real time through terahertz time-domain spectrometer; Formulate a system of coupled differential equations involving parasitic inductances and capacitances; Fractional derivatives are used to describe the relationship between conductor deformation and the memory effect of parasitic parameters, and a non-integer order differential equation model is established, which is recorded as a time-varying parameter evolution equation. The optimization solution process of the quantum annealing algorithm includes: The parasitic parameter identification problem is converted into the Ising model, and the design includes Hamiltonian encoding scheme for qubits; Perform annealing operations in a quantum annealing processor and search for the global optimal solution through quantum tunneling effect; The annealing results are post-processed using quantum error correction codes to output real-time curves of parasitic inductance and capacitance. The calculation of the voltage spike sensitivity characteristic value includes: Based on the parameter change curve output by the quantum annealing algorithm, a transient voltage response prediction model is constructed; Analyze system stability through Lyapunov exponents and quantify the probability of voltage spikes; Establish sensitivity grading standards based on historical fault data and output normalized sensitivity characteristic values; S4: Input the dynamic eddy current loss characteristic value and the voltage spike sensitivity characteristic value into the pre-trained high-voltage inverter fault diagnosis model for analysis, and output the fault risk assessment result; The analysis process of the high-voltage inverter fault diagnosis model specifically includes: Construct a deep residual shrinkage network based on multimodal feature fusion, where: The first residual block receives the dynamic eddy current loss eigenvalue and eliminates the measurement noise through adaptive soft thresholding; The second residual block receives the voltage spike sensitivity feature value and uses the channel attention mechanism to enhance the key frequency band features; The outputs of the two residual blocks are fed into the spatiotemporal cross attention module to establish a coupling relationship model between eddy current loss and voltage spikes; The cross-attention output is parsed through a quantum-derived decision tree to generate a triplet evaluation result containing fault type, location, and risk level; The adversarial verification module is used to check the credibility of the evaluation results. When the confidence level is lower than the threshold, the feature recalculation mechanism is triggered. S5: Based on the fault risk assessment results, dynamically adjust the inverter operating parameters or trigger the hierarchical protection mechanism to achieve adaptive fault protection under high-frequency switching conditions.

2. A method for analyzing and diagnosing faults of a high-voltage frequency converter for mining according to claim 1, characterized in that: The dynamic adjustment of the inverter operating parameters or triggering of the hierarchical protection mechanism specifically includes: Build an adaptive control strategy library based on reinforcement learning, which includes: The first-level response targets low-risk faults and uses a fuzzy PID controller to dynamically adjust the carrier frequency and dead time; The second level response targets medium-risk faults and activates the N+1 redundant switching mechanism of the power modules; The third level response targets high-risk faults and executes an emergency shutdown sequence based on chaos synchronization; The expected effects of each control strategy are simulated in real time through the digital twin system, and the optimal response plan is selected using the Bayesian optimization algorithm; Implement nanosecond-level protection trigger logic at the FPGA hardware layer. When critical fault characteristics are detected, the protection action is directly executed, bypassing the main control system. The protection decision-making process is recorded through blockchain smart contracts to establish an unalterable fault response traceability chain.

3. A fault analysis and diagnosis system for a high-voltage frequency converter used in a mine, characterized in that: A method for analyzing and diagnosing faults of a high-voltage frequency converter for mining as claimed in any one of claims 1 to 2, comprising: A multi-physics field real-time monitoring module, which is used to collect operating status parameters of the high-voltage inverter in real time during its operating cycle. The operating status parameters include: skin effect depth and parasitic parameter drift under high-frequency conditions; A three-dimensional electromagnetic field dynamic reconstruction module, based on the skin effect depth, uses a non-Euclidean space meshing technique to construct a three-dimensional current density distribution of the conductor cross section, dynamically corrects the conductor equivalent resistivity through a metamaterial electromagnetic property database, and calculates the dynamic eddy current loss characteristic value, specifically including: Establish a differential geometry model of the conductor surface and map the conductor cross section into a non-Euclidean space; Construct hexahedral-tetrahedral hybrid meshes with variable topology in non-Euclidean space; The three-dimensional current density distribution is obtained by solving the Maxwell-Riemann equations; The calculation of the dynamic eddy current loss characteristic value includes: The eddy current component in the skin layer is extracted from the three-dimensional current density distribution, and the loss power density equation is established using fractional calculus. Calculate the instantaneous loss value of each conductor cross section through surface integral and superimpose the time domain integration results within the high-frequency switching cycle; The vibration acceleration sensor data is introduced to dynamically compensate the calculation results and output the eddy current loss characteristic value corrected by environmental disturbance; A parasitic parameter quantum optimization module, which establishes a time-varying parameter evolution equation based on the parasitic parameter drift, uses a quantum annealing algorithm to optimize and solve the real-time variation law of parasitic inductance or capacitance, and calculates the voltage spike sensitivity characteristic value; The establishment of the time-varying parameter evolution equation includes: The three-dimensional electromagnetic field distribution data of PCB wiring is collected in real time through terahertz time-domain spectrometer; Formulate a system of coupled differential equations involving parasitic inductances and capacitances; Fractional derivatives are used to describe the relationship between conductor deformation and the memory effect of parasitic parameters, and a non-integer order differential equation model is established, which is recorded as a time-varying parameter evolution equation. The optimization solution process of the quantum annealing algorithm includes: The parasitic parameter identification problem is converted into the Ising model, and the design includes Hamiltonian encoding scheme for qubits; Perform annealing operations in a quantum annealing processor and search for the global optimal solution through quantum tunneling effect; The annealing results are post-processed using quantum error correction codes to output real-time curves of parasitic inductance and capacitance. The calculation of the voltage spike sensitivity characteristic value includes: Based on the parameter change curve output by the quantum annealing algorithm, a transient voltage response prediction model is constructed; Analyze system stability through Lyapunov exponents and quantify the probability of voltage spikes; Establish sensitivity grading standards based on historical fault data and output normalized sensitivity characteristic values; a multimodal fault diagnosis module that inputs dynamic eddy current loss characteristic values ​​and voltage spike sensitivity characteristic values ​​into a pre-trained high-voltage inverter fault diagnosis model for analysis and outputs a fault risk assessment result; The analysis process of the high-voltage inverter fault diagnosis model specifically includes: Construct a deep residual shrinkage network based on multimodal feature fusion, where: The first residual block receives the dynamic eddy current loss eigenvalue and eliminates the measurement noise through adaptive soft thresholding; The second residual block receives the voltage spike sensitivity feature value and uses the channel attention mechanism to enhance the key frequency band features; The outputs of the two residual blocks are fed into the spatiotemporal cross attention module to establish a coupling relationship model between eddy current loss and voltage spikes; The cross-attention output is parsed through a quantum-derived decision tree to generate a triplet evaluation result containing fault type, location, and risk level; The adversarial verification module is used to check the credibility of the evaluation results. When the confidence level is lower than the threshold, the feature recalculation mechanism is triggered. An adaptive protection execution module dynamically adjusts the inverter operating parameters or triggers a hierarchical protection mechanism based on the fault risk assessment results to achieve adaptive fault protection under high-frequency switching conditions.

Images