A method for rapid identification of motor winding damage in mining machinery

Abstract

This invention belongs to the field of electrical equipment fault diagnosis and intelligent monitoring technology, specifically relating to a rapid identification method for motor winding damage in mining machinery. The method includes: collecting multi-source synchronous sensing signals of three-phase current, voltage, stator temperature distribution, and casing vibration acceleration during motor operation; constructing a high-dimensional feature vector integrating electromagnetic, thermal, and mechanical coupling information; establishing a nonlinear mapping model through a deep temporal convolutional network with an attention mechanism; constructing a dynamic health baseline that adaptively corrects according to load and environmental conditions; calculating the deviation by comparing the spatiotemporal temperature gradient and weighted harmonic characteristic index with the expected value of the dynamic health baseline; determining the level of early minor damage, mid-term development, or severe failure based on the magnitude of the deviation and the feature combination pattern; and outputting a diagnostic report including damage location tendency and development trend. This invention achieves highly robust and highly sensitive online rapid identification without machine shutdown, and is suitable for complex mining environments such as high humidity and strong interference.

Description

Technical Field

[0001] This invention belongs to the field of electrical equipment fault diagnosis and intelligent monitoring technology, specifically relating to a method for rapid identification of damage to motor windings in mining machinery. Background Technology

[0002] With the deepening of industrialization in the mining industry, the automation and intelligence level of large-scale mining machinery has become a core indicator for measuring production efficiency and operational safety. As the main power source for mining machinery, the operational stability of the motor directly affects the continuous operation of key processes such as hoisting, crushing, and transportation. Under harsh operating conditions of long-term high load, high humidity, and strong impact, the health of the motor windings becomes a decisive factor in ensuring efficient mining of mineral resources and reducing unplanned downtime losses.

[0003] Among them, the motor winding damage identification technology aims to predict and locate insulation failures, inter-turn short circuits, or mechanical stress damage within the winding by real-time monitoring of electromagnetic signals, vibration frequencies, or thermal distribution characteristics. In complex and variable mining conditions, this technology requires accurately mapping the logical relationship between the electromagnetic physical model inside the motor and real-time operating data. Especially during frequent heavy-load starts and variable frequency speed regulation, the electrical parameters of the winding exhibit highly dynamic and non-stationary evolution characteristics, placing extremely high demands on the response speed of the identification method and its ability to analyze subtle fault characteristics.

[0004] However, traditional winding inspection methods often rely on offline manual testing or periodic physical disassembly, making it difficult to meet the real-time and coverage requirements of fault identification for continuous production. Meanwhile, the strong electromagnetic coupling and multi-source background noise present in mining sites easily cause aliasing of characteristic signals, making it difficult for existing identification models to effectively isolate characteristic components related to early winding damage. Furthermore, existing analysis methods exhibit hysteresis when dealing with nonlinear current load fluctuations, failing to capture transient changes in damage states, allowing minor damage to easily escalate into severe burnout accidents. Summary of the Invention

[0005] The purpose of this invention is to provide a rapid identification method for motor winding damage in mining machinery, aiming to solve the problems of low sensitivity and susceptibility to interference in the identification of early winding damage under complex working conditions in existing technologies. This invention achieves highly robust and sensitive online rapid identification without requiring machine shutdown, and is suitable for complex mining environments such as high humidity and strong interference.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for rapid identification of damage to motor windings in mining machinery, comprising the following specific steps:

[0007] S1. Collect multi-source synchronous sensing signals during motor operation. The multi-source synchronous sensing signals include three-phase current signals, three-phase voltage signals, stator temperature distribution signals, and housing vibration acceleration signals. All signals are collected under a unified time reference by a high-precision synchronous sampling module, and the original signals are subjected to anti-aliasing filtering.

[0008] S2. Constructing the dynamic multi-physics field feature vector of the motor winding: Based on the multi-source synchronous sensing signal, extract the real-time phase difference, weighted harmonic characteristic index, zero-sequence component amplitude and negative-sequence component imbalance that characterize the electromagnetic state, fuse the stator temperature gradient spatiotemporal fusion index that characterizes the thermodynamic state and the energy concentration of a specific frequency band of the vibration spectrum that characterizes the mechanical state, and arrange the above multi-dimensional features in a predetermined sequence to form a high-dimensional feature vector containing electromagnetic, thermodynamic and mechanical coupling information.

[0009] S3. Establish a nonlinear mapping model of winding damage state. Use a deep temporal convolutional network that integrates physical mechanism constraints to jointly train historical normal operation data and known damage samples. The deep temporal convolutional network includes multiple causal convolutional layers and a self-attention weight mechanism based on physical prior guidance, which is used to capture the evolution law of feature vectors in the time dimension and the dynamic weights of key damage-sensitive features.

[0010] S4. Construct a dynamic health baseline that is adaptively corrected according to load and environmental conditions, and compare the currently extracted time temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state with the theoretical expected value of the dynamic health baseline in real time, and calculate the multidimensional deviation degree that characterizes the degree of deviation of the state.

[0011] The multidimensional deviation is quantified by introducing a weighted Mahalanobis distance with physical prior sensitivity weights. The weight coefficients are dynamically allocated based on the sensitivity of each physical feature to winding damage learned by the S3 model, so as to eliminate interference from dimensional differences and feature coupling, and accurately identify and exclude early damage signals after drift under normal operating conditions.

[0012] S5. Determine the winding damage level and output the identification result. When the multidimensional deviation exceeds the preset threshold, trigger the damage warning mechanism and divide the damage state into three levels: early minor damage, mid-term development or severe failure, based on the magnitude of the multidimensional deviation and the feature combination pattern. At the same time, generate a diagnostic report that includes the damage location tendency and development trend.

[0013] Preferably, in step S1,

[0014] Three-phase current signals are captured by electromagnetic transformers or Hall effect sensors deployed on the power cables at the input end of the motors of mining machinery, and three-phase voltage signals are captured by resistor-capacitor voltage divider circuits or voltage transformers.

[0015] The temperature distribution signal of the stator is obtained by using a temperature-sensing optical fiber that is spirally arranged along the entire length of the winding axis in the stator slot. The temperature-sensing optical fiber is connected to a demodulator, and the real-time temperature is calculated by analyzing the amount of wavelength shift or the intensity change of backscattered light.

[0016] The vibration acceleration signal of the housing is collected by a triaxial piezoelectric accelerometer installed on the surface of the motor bearing housing and the stator housing.

[0017] The hardware-level timestamp alignment mechanism generates a global trigger pulse through the master clock divider, synchronously triggering the analog-to-digital converters of each signal, ensuring that the time synchronization error of different physical quantity signals is less than one sampling period, and that the sampling frequency is not less than the requirement of 20 sampling points within the power frequency period.

[0018] The anti-aliasing filtering process uses a fourth-order Butterworth low-pass filter to filter out noise outside the sampling bandwidth and applies a mean filtering algorithm in the digital domain to suppress impulse noise.

[0019] Preferably, in step S2:

[0020] The three-phase current and three-phase voltage signals are subjected to full-cycle Fourier transform to extract the fundamental components of each phase and calculate the real-time phase difference between the current and voltage of each phase.

[0021] The extraction process of the weighted harmonic characteristic index reflecting the winding state includes: performing a discrete Fourier transform on the current sequence, extracting its spectral distribution, and then selecting the third, fifth, and seventh harmonic components for weighted synthesis, with the third harmonic having the highest weight coefficient, which is used to characterize the asymmetric current characteristics caused by the inter-turn short circuit.

[0022] The zero-sequence component amplitude is obtained by vector summation of the three-phase current signals and taking its magnitude, and is used to reflect the degree of deterioration of the winding insulation to ground.

[0023] The negative sequence unbalance is based on the symmetrical component method to extract the negative sequence component and calculate its ratio to the positive sequence component, which is used to characterize the degree of electrical asymmetry of the three-phase winding.

[0024] The spatiotemporal temperature gradient index is calculated from the temperature data along the groove obtained by the distributed optical fiber temperature measurement system. Specifically, it includes: calculating the ratio of the temperature difference between adjacent temperature measurement points to the physical distance to obtain the spatial temperature gradient, and calculating the ratio of the temperature change of the same temperature measurement point at adjacent sampling times to the sampling time interval to obtain the temporal temperature gradient. The fusion of the spatial temperature gradient and the temporal temperature gradient identifies local overheated areas.

[0025] The specific frequency band of the vibration spectrum refers to the frequency band within the range of the motor rotation frequency and its second to fifth harmonics. The energy concentration is obtained by integrating the power spectral density within this frequency band and is used to correlate the mechanical resonance characteristics caused by loosening of the stator winding, end deformation, or electromagnetic force imbalance.

[0026] Preferably, in step S3:

[0027] The causal convolutional layers in the deep temporal convolutional network employ an dilated convolutional structure, which expands the network's receptive field by gradually increasing the dilation rate in order to capture the evolution patterns of feature vectors over long time scales.

[0028] During the training phase of the model, an adversarial generative sample enhancement strategy based on physical constraints is introduced. The generator network generates synthetic damage data according to the multi-physics constraint equation of the motor, and the discriminator network distinguishes between scarce damage samples and synthetic samples. Through the alternating game between the two, a large number of weak damage samples with high realism and satisfying physical mechanisms are generated, thereby improving the generalization recognition ability of the deep learning model for early and hidden damage.

[0029] The self-attention weighting mechanism evaluates the correlation between the current input features and potential damage patterns by scaling the dot product attention module, and dynamically allocates attention intensity based on the historical contribution of each dimension of features.

[0030] More preferably, the self-attention weighting mechanism assigns a preset high initial weight to the time temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state during the initialization phase, so as to enhance the response to insulation aging and inter-turn short circuits. During motor operation, the model automatically fine-tunes the weight coefficients of each dimension according to different operating conditions. When in the heavy load start-up phase, the weight of transient vibration characteristics caused by current surge is reduced, and the attention weight of the weighted harmonic characteristic index reflecting the winding state and the time-space temperature gradient index is increased accordingly.

[0031] Preferably, in step S4:

[0032] The adaptively corrected dynamic health baseline is dynamically corrected based on the motor's current load rate, ambient temperature, ambient humidity, and cumulative running time.

[0033] The system uses a long short-term memory network module to record the characteristic evolution trajectory of the motor during the healthy operation phase throughout its entire life cycle. Combined with the current environmental parameters, the system uses a trained deep temporal convolutional network to deduce the theoretical expected values ​​and confidence intervals of each feature dimension, forming a healthy distribution manifold space.

[0034] Preferably, the quantization calculation process of the weighted Mahalanobis distance includes:

[0035] First, the feature vectors extracted in real time are centered and the mean of the healthy baseline is subtracted.

[0036] Subsequently, a diagonal matrix consisting of weight coefficients is introduced, which are dynamically adjusted according to the sensitivity of each feature dimension to winding damage. The sensitivity is obtained by mapping the physical prior attention weights learned by the model in step S3.

[0037] By eliminating interference caused by different units, magnitudes, and correlations among the various feature dimensions through matrix operations, the final output is a value reflecting the degree to which the current state of the motor deviates from the core healthy region.

[0038] Preferably, in step S5, the damage state is divided into three levels: early minor damage, intermediate development, and severe failure.

[0039] When the multidimensional deviation is between the first safety threshold and the second threshold, and the characteristic is that the weighted harmonic characteristic index reflecting the winding state increases or the spatiotemporal temperature gradient index increases, it is judged as an early micro-loss level.

[0040] When the multidimensional deviation exceeds the second threshold, and the characteristics include the growth of the negative sequence current component, the fluctuation of the zero sequence component, and the energy concentration of the vibration spectrum at the second harmonic, it is judged to be of the medium-term development level.

[0041] When the multidimensional deviation exceeds the third limit threshold, and the three-phase voltage and current show a severe imbalance, and the spatial temperature gradient shows a local temperature surge in the stator accompanied by violent shell oscillation, it is judged as a severe failure level.

[0042] Preferably, in step S5, the specific steps for generating a diagnostic report containing the predisposition and development trend of the injury are as follows:

[0043] By analyzing the imbalance of harmonic components of each phase current in the feature vector and combining the spatial coordinates of the abnormal temperature area fed back by the distributed optical fiber temperature measurement system, the phase and slot level of the winding are inferred, and the damage location is located in a tendency manner.

[0044] The damage development trend prediction in the diagnostic report is calculated based on the slope of the multidimensional deviation change over multiple consecutive monitoring periods. When the slope of the change shows an exponential growth trend, an emergency shutdown suggestion is output.

[0045] When the slope of the change is gentle, the recommended maintenance window period is automatically calculated and output by combining the historical maintenance data of the motor.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] This invention overcomes the problem of feature submersion caused by a single signal source in a noisy environment by constructing a high-dimensional feature vector by fusing synchronous signals from multiple physical fields, including electromagnetic, thermal, and mechanical fields. It employs a nonlinear mapping model combining a deep temporal convolutional network and an attention mechanism to effectively capture the weak and non-stationary feature evolution patterns in the early stages of winding damage, significantly improving recognition sensitivity. Through an adaptively corrected dynamic health baseline and a multi-dimensional deviation quantification mechanism, it avoids false alarms caused by changes in operating conditions, improving the robustness of recognition. The damage level classification and location tendency inference functions provide an operable decision-making basis for predictive maintenance of mining machinery, significantly shortening fault response time and reducing the risk of unplanned downtime. The entire recognition process does not require machine shutdown and can be completed while the motor is running normally, achieving truly rapid online recognition. It is suitable for complex mining environments such as high humidity, strong electromagnetic interference, and heavy-load frequency conversion. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the overall technical solution architecture of a method for rapid identification of damage to motor windings in mining machinery, according to an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, the technical solution adopted by the present invention is: a method for rapid identification of damage to motor windings in mining machinery, comprising the following specific steps:

[0052] S1. Collect multi-source synchronous sensing signals during motor operation. The multi-source synchronous sensing signals include three-phase current signals, three-phase voltage signals, stator temperature distribution signals, and housing vibration acceleration signals. All signals are collected under a unified time reference by a high-precision synchronous sampling module, and the original signals are processed by anti-aliasing filtering.

[0053] S2. Constructing the dynamic multi-physics feature vector of the motor winding: Based on the multi-source synchronous sensing signal, extract the real-time phase difference, weighted harmonic characteristic index, zero-sequence component amplitude and negative-sequence component imbalance that characterize the electromagnetic state, fuse the stator temperature gradient spatiotemporal fusion index that characterizes the thermodynamic state and the energy concentration of a specific frequency band of the vibration spectrum that characterizes the mechanical state, and arrange the above multi-dimensional features in a predetermined sequence to form a high-dimensional feature vector containing electromagnetic, thermodynamic and mechanical coupling information.

[0054] S3. Establish a nonlinear mapping model of winding damage state. Use a deep temporal convolutional network that integrates physical mechanism constraints to jointly train on historical normal operation data and known damage samples. The deep temporal convolutional network includes multiple causal convolutional layers and a self-attention weight mechanism based on physical prior guidance, which is used to capture the evolution law of feature vectors in the time dimension and the dynamic weights of key damage-sensitive features.

[0055] S4. Construct a dynamic health baseline that is adaptively corrected according to load and environmental conditions. In real time, compare the currently extracted time temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state with the theoretical expected value of the dynamic health baseline, and calculate the multidimensional deviation degree that characterizes the degree of deviation from the state.

[0056] The multidimensional deviation is quantified by introducing a weighted Mahalanobis distance with physical prior sensitivity weights. The weight coefficients are dynamically allocated based on the sensitivity of each physical feature to winding damage learned by the S3 model, so as to eliminate interference from dimensional differences and feature coupling, and accurately identify and exclude early damage signals after drift under normal operating conditions.

[0057] S5. Determine the winding damage level and output the identification result. When the multidimensional deviation exceeds the preset threshold, trigger the damage warning mechanism and divide the damage state into three levels: early minor damage, mid-term development or severe failure, based on the magnitude of the multidimensional deviation and the feature combination pattern. At the same time, generate a diagnostic report that includes the damage location tendency and development trend.

[0058] Step S1 involves acquiring multi-source synchronous sensing signals during motor operation. The acquisition of these signals relies on a series of sensor components deployed at key locations on the mining machinery motor. Specifically, three-phase current and three-phase voltage signals are captured using high-precision electromagnetic transformers or Hall effect sensors. The current transformer is installed on the motor's input power cable to convert the high-voltage, high-current signal into a small voltage signal within the input range of the sampling circuit. Simultaneously, the voltage signal acquisition employs a resistive-capacitive voltage divider circuit or a voltage transformer to ensure signal linearity. To cope with the complex high-frequency electromagnetic interference in the mining environment, electromagnetically shielded twisted-pair cables are connected to the transformer's output to reduce induced noise during signal transmission.

[0059] The stator temperature distribution signal is acquired using an embedded distributed fiber optic temperature measurement device, which utilizes the Bragg fiber grating sensing principle or the Raman scattering temperature measurement principle. During motor manufacturing or overhaul, the temperature-sensing fiber is spirally deployed along the entire axial length of the winding within the stator slot, achieving a spatial resolution of 0.1m, enabling precise capture of the temperature status of every local area within the winding. The end of the fiber is connected to a demodulator, which calculates the real-time temperature at the corresponding location by analyzing the wavelength shift or the intensity change of backscattered light.

[0060] The vibration acceleration signal of the housing is acquired by a triaxial piezoelectric accelerometer mounted on the motor bearing housing and the surface of the stator housing. The accelerometer's range is set according to the heavy-load vibration characteristics of the mining machinery, typically selecting a dynamic range of ±50g, with a response frequency of 0.1~10000Hz. The triaxial piezoelectric accelerometer is fixed to a location with high housing rigidity using a magnetic mount or bolts to ensure accurate transmission of the vibration signal.

[0061] At the hardware execution level of data acquisition, all signals enter the processing system through a multi-channel, high-precision synchronous sampling module. This sampling module incorporates a hardware-level timestamp alignment mechanism. Within each sampling cycle, the sampling module's master clock divider generates a global trigger pulse, which is simultaneously sent to the trigger terminals of the analog-to-digital converters for current, voltage, temperature, and vibration signals. This master-slave synchronous architecture ensures that different types of physical quantity signals have completely consistent sampling instants on the time axis. The sampling frequency is set to a requirement of no less than 20 sampling points within the power frequency cycle. In a 50Hz power frequency environment, this means a sampling rate of at least 1000 times per second, while to capture high-frequency harmonic characteristics, the actual sampling rate is typically set to 10240 times per second or higher. The time synchronization error of each signal is strictly limited to within one sampling cycle, avoiding spurious power fluctuations or correlation analysis errors caused by phase shifts.

[0062] The acquired raw electrical signals are first processed by anti-aliasing filtering. This filtering process employs a fourth-order Butterworth low-pass filter at the hardware level, with its cutoff frequency set below half the sampling frequency to filter out noise outside the sampling bandwidth. Subsequently, the analog signal is converted into a digital sequence via a 16-bit or 24-bit analog-to-digital converter. In the digital domain, the system further applies digital signal processing algorithms, including DC offset component removal and impulse noise suppression based on median filtering, to obtain high-quality basic data.

[0063] Step S2 involves constructing a dynamic multiphysics feature vector for the motor windings. This process transforms the multi-source raw signals into a structured feature set that characterizes the evolution of the motor's internal physical state. First, for the acquired three-phase current and three-phase voltage signals, the fundamental components of each phase are extracted using a full-cycle Fourier transform algorithm. Based on the complex representation of the fundamental components, the real-time phase difference between the current and voltage of each phase is calculated. This phase difference reflects the changes in the load's power factor and the winding's inductive reactance. When an inter-turn short circuit or insulation degradation occurs in the winding, changes in local impedance cause a shift in the phase difference of the damaged phase.

[0064] Furthermore, a weighted harmonic characteristic index reflecting the winding state is constructed, and the current sequence is subjected to discrete Fourier transform to extract its spectral distribution. To improve the sensitivity to winding damage, the third, fifth, and seventh harmonic components are selected for weighted synthesis. In the weighting process, the third harmonic is assigned the highest weight coefficient, typically accounting for more than 60%. This is because early asymmetrical faults in the motor stator winding, such as minor inter-turn short circuits, significantly increase the magnetomotive force spatial harmonics, thereby inducing specific third current harmonics.

[0065] For example, the weighted synthesis is calculated using the following formula:

[0066]

[0067] Where A1 is the fundamental current amplitude, and A3, A5, and A7 are the amplitudes of the third, fifth, and seventh harmonics, respectively; k3, k5, and k7 are the corresponding weighting coefficients. In practical applications, the weighting coefficients are set to k3=0.6, k5=0.25, and k7=0.15. By calculating the winding health characteristic index at the current moment and comparing it with a preset threshold, if it exceeds the threshold, it is determined that there is a potential inter-turn short circuit. The preset threshold is 1.5 times the benchmark value.

[0068] The calculation of the zero-sequence component amplitude is based on the symmetrical component method, which involves vector summation of the fundamental phasors of the three-phase currents and taking their magnitudes. In a perfectly balanced and healthy motor state, the vector sum of the three-phase currents is theoretically zero; however, when the winding insulation deteriorates to ground or during the early evolution of a single-phase ground fault, the zero-sequence component amplitude will show a significant increase, thus serving as a characteristic indicator of ground faults.

[0069] Simultaneously, the system extracts the negative-sequence component based on the symmetrical component method and calculates its imbalance relative to the positive-sequence component. The magnitude of the negative-sequence current directly characterizes the degree of electrical asymmetry in the three-phase windings. Since internal faults such as inter-turn short circuits can disrupt the symmetry of winding parameters without necessarily causing a change in zero-sequence current, the negative-sequence component is a key specific indicator for identifying winding asymmetric damage. For example, winding asymmetric damage occurs in the early stages of inter-turn short circuits or phase-to-phase short circuits.

[0070] Regarding thermal characteristics, the system calculates the spatiotemporal fusion index of the stator temperature gradient. Using temperature data along the groove acquired by a distributed fiber optic temperature measurement device, the gradients in two dimensions are first calculated: the spatial temperature gradient and the temporal temperature gradient. The spatial temperature gradient is calculated as the temperature difference ΔT between adjacent measurement points. space Dividing by its physical distance Δx reflects the spatial non-uniformity of the temperature field; the time-temperature gradient is used to calculate the temperature change ΔT at the same temperature measurement point at adjacent sampling times. time The temperature rise rate is reflected by dividing the result by the sampling time interval Δt. By fusing these two values, abnormal hotspot regions caused by local winding damage are identified. If the spatiotemporal fusion index of a certain region continuously exceeds a set threshold for a preset duration, it is determined to be local winding damage and overheating. For example, a weighted fusion algorithm synthesizes the two gradients into a spatiotemporal thermal fault index I. thermal The calculation formula is as follows:

[0071]

[0072] Among them, K s_base and K t_base These are the spatial and temporal gradient reference values ​​under normal operating conditions, w s and w t The weighting coefficient is usually w. s =w t =0.5 or adjusted according to the fault mechanism. This fusion mechanism can effectively eliminate single-dimensional interference: a high spatial gradient alone may indicate poor sensor contact, a high temporal gradient alone may indicate load fluctuations, and a significant increase in both simultaneously confirms a local overheating fault.

[0073] In terms of mechanical characteristics, the system analyzes the vibration acceleration signal mounted on the motor housing in real time. First, the power spectral density of the vibration signal is estimated, and then the frequency band energy integration method is used to calculate the vibration energy concentration in the range of the motor rotation frequency and its second to fifth harmonics.

[0074] When the stator winding becomes loose, deformed at the ends, or experiences electromagnetic force imbalance, it will trigger abnormal vibration responses at specific frequencies. By calculating the relative proportion of energy in a specific frequency band to the total vibration energy, the extent of damage to the winding at the mechanical structural level can be assessed. Finally, the extracted electromagnetic, thermal, and mechanical features are standardized and normalized according to their physical meaning, forming a high-dimensional feature vector containing multi-dimensional physical field coupling information.

[0075] Step S3 involves establishing a nonlinear mapping model between the winding damage state and the multidimensional feature vector. The core of this nonlinear mapping model adopts an improved temporal convolutional network, which integrates multiple causal convolutional layers and a self-attention weight mechanism.

[0076] The design of the causal convolutional layers ensures that the output at any given time depends only on the input information prior to that time, which aligns with the logic of online real-time monitoring. Each convolutional layer employs dilated convolution, gradually increasing the dilation rate to enable the network to obtain a large receptive field without increasing the number of parameters. This allows the network to capture the evolution patterns of winding damage characteristics over long time scales, such as the long-range coupling relationship between slow temperature drift and current distortion.

[0077] During the model training phase, an adversarial generative sample enhancement strategy based on physical constraints is employed. This strategy introduces a generator network and a discriminator network. The generator network takes random noise and operating parameters as input and embeds the multiphysics constraint equations of the motor into its loss function, thereby generating synthetic damage data that conforms to both statistical distribution and strict physical logic, simulating complex scenarios including various degrees of inter-turn short circuits, insulation aging, and heat dissipation obstruction. The discriminator network attempts to distinguish between the scarce, real-world damage samples and the synthetic samples. Through alternating game optimization between the two, the generator can produce a large number of highly realistic and physically consistent weak damage samples. This step effectively solves the problems of extremely scarce damage samples and highly imbalanced sample categories in actual operation of mining machinery, greatly improving the generalization ability of deep learning models to identify early and hidden damage.

[0078] The training phase of the model specifically employs a physical constraint-based adversarial generative sample enhancement strategy. This strategy constructs a game-theoretic architecture comprising a generator network and a discriminator network. The generator network takes random noise and operating parameters as input, aiming to synthesize time-series data of various winding damages. Examples of these damages include inter-turn short circuits, insulation aging, and impaired heat dissipation. Its core innovation lies in introducing a physical constraint loss term into the training objective function, forcing the generated data to strictly adhere to the two major physical mechanisms of motor operation:

[0079] 1. Voltage balance constraints in electrical circuits:

[0080] The generated instantaneous current sequence i(t) and instantaneous voltage sequence u(t) must satisfy the requirement of minimizing the residuals of the stator voltage balance equation:

[0081]

[0082] In the formula: u(t) is the stator terminal voltage, R is the winding resistance, and L is the winding inductance. Let e(t) be the rate of change of current, and e(t) be the back electromotive force.

[0083] 2. Thermodynamic energy conservation constraint:

[0084] The evolution of the generated instantaneous temperature sequence T(t) must satisfy the minimization of the residuals of the heat balance equation:

[0085]

[0086] In the formula: C is the equivalent heat capacity of the winding. For the rate of temperature change, T represents the Joule heat power generated by the current, h represents the overall heat dissipation coefficient, and T represents the total heat dissipation. env The ambient temperature is used as the constraint. This constraint ensures that the generated "high temperature" samples are necessarily caused by "high current" or "impeded heat dissipation," thus eliminating false data that violates the law of conservation of energy.

[0087] The generator's total loss function consists of adversarial loss and weighted physical constraint loss. Adversarial loss measures the realism of the data, while weighted physical constraint loss measures physical compliance. To avoid excessively strong physical constraints in the early stages of training, which could hinder model convergence, this embodiment employs a "phased incremental" strategy: In the early training phase, a smaller physical constraint weight coefficient is set, allowing the generator to prioritize learning the statistical distribution characteristics of real data and quickly establish basic data generation capabilities. In the later training phase, as the number of iterations increases, the weight coefficient is increased linearly or stepwise, forcing the generator to focus on correcting residuals that violate the aforementioned physical equations. For example, in the early training phase, the physical constraint weight coefficient is set to λ=0.1, and in the later training phase, as the number of iterations increases, the weight coefficient is linearly and gradually increased to λ=0.5.

[0088] This strategy ensures that the final generated samples are both statistically similar to real data and physically conform to the operating rules of motors.

[0089] The discriminator network is responsible for distinguishing between the real, scarce samples and the aforementioned synthetic samples. Through the alternating game optimization between the two, this step effectively solves the problems of extremely scarce mining machinery damage samples and class imbalance, and significantly improves the model's ability to generalize and identify early and hidden damage.

[0090] A self-attention weighting mechanism is embedded between convolutional layers to enhance the model's ability to capture multi-dimensional physical features. This mechanism adaptively allocates attention intensity by calculating the dynamic correlation of the input feature sequence under different operating conditions. Specifically, the algorithm utilizes a scaled dot product attention module to evaluate the causal correlation between the high-order physical features of the current input and potential damage patterns.

[0091] During the model initialization phase, the system introduces a physical prior knowledge-guided strategy: based on the motor fault mechanism, high initial attention biases are assigned to the two core features with the strongest causal correlation to insulation aging and inter-turn short circuits. Specifically, these are: the temporal temperature gradient, the spatial temperature gradient, and the weighted harmonic characteristic index reflecting the winding state. The temporal temperature gradient directly corresponds to the heat accumulation rate in the thermodynamic equation, sensitively reflecting changes in heat capacity or abnormal overall temperature rise caused by insulation aging; the spatial temperature gradient corresponds to the non-uniform heat dissipation term in the heat conduction equation, accurately locating the distribution of "hot spots" caused by local inter-turn short circuits, compensating for the shortcomings of single-point temperature monitoring; the weighted harmonic characteristic index integrates the distortion information of the third harmonic and other higher harmonics, corresponding to the frequency domain response caused by changes in nonlinear inductance and resistance in the voltage balance equation. Compared to single harmonics, this weighted index can more robustly characterize electromagnetic field distortion caused by slight winding deformation or early short circuits.

[0092] As training progresses, the model automatically fine-tunes the attention weights across various dimensions under the joint drive of the physical constraint loss function. For example, during high-current transients at heavy-load startup, the model automatically suppresses noise feature weights caused by normal mechanical vibrations, instead focusing on anomalous shifts in the "weighted harmonic characteristic index" and whether the evolution of the "spatiotemporal temperature gradient" violates the laws of heat conduction. This mechanism ensures that the model consistently focuses on core evidence consistent with physical fault mechanisms, achieving a leap from "data-driven" to "physical-data dual-driven," significantly enhancing the accuracy and interpretability of identifying early, hidden damage.

[0093] In step S4, the aim is to compare the currently extracted feature vector with the expected health state output by the nonlinear mapping model in real time to quantify the degree to which the current state of the motor winding deviates from the normal baseline. This process constructs a dynamic health baseline that adapts to the operating conditions and combines it with a weighted Mahalanobis distance algorithm to achieve highly sensitive identification of early, minor damage. The specific implementation process is as follows:

[0094] First, a dynamic health baseline based on physical mechanism constraints is constructed; then, real-time feature vector extraction and centering are performed; subsequently, weighted Mahalanobis distance calculation based on damage sensitivity weights is performed; finally, state assessment and output of a multidimensional deviation degree representing the degree of state deviation are performed.

[0095] Traditional health baselines are typically fixed thresholds or static statistical distributions, which cannot adapt to the nonlinear changes in motors under different loads and complex environments. This embodiment proposes a manifold space adaptive correction mechanism:

[0096] Input parameters: The system obtains the motor's current load rate, environmental operating parameters, and cumulative running time in real time. Among them, environmental operating parameters include, but are not limited to, ambient temperature, humidity, and ventilation coefficient.

[0097] Baseline extrapolation: Using the deep temporal convolutional network with integrated physical mechanism constraints trained in step S3, the above input parameters are used as conditional variables to extrapolate the theoretical expected values ​​and confidence intervals of the temporal temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state under the current specific operating conditions.

[0098] Physical Significance: This dynamic health baseline is essentially a multidimensional manifold space, whose central trajectory is determined by data fine-tuning of the theoretical solutions to the voltage balance equation and the heat conduction equation. This means that the baseline can automatically "drift" to follow load fluctuations and environmental changes, thereby mathematically eliminating characteristic offsets caused by changes in normal operating conditions and retaining only abnormal residuals caused by substantial damage to the windings. For example, mathematically eliminating characteristic offsets caused by the natural temperature rise due to a sudden increase in load.

[0099] Historical memory mechanism: The system embeds a long short-term memory network module to record the characteristic evolution trajectory of the motor during the healthy operation stage throughout its entire life cycle. Combined with the current environmental parameters, the theoretical expected value is corrected a second time to ensure that the baseline conforms to the laws of physics and fits the actual aging characteristics of the individual motor.

[0100] Then, real-time monitoring data is collected and the corresponding temporal temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state are extracted to form a real-time feature vector. The weighted Mahalanobis distance between the real-time feature vector and the center point of the dynamic health baseline is calculated as a quantitative indicator characterizing the degree of winding damage. The calculation of the weighted Mahalanobis distance introduces a damage sensitivity weight matrix, whose diagonal elements are obtained by mapping the physical prior attention weights learned by the model in step S3. This matrix is ​​used to characterize the differentiated sensitivity of the temporal temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index to early winding damage. The weighted Mahalanobis distance eliminates the interference of different physical dimensions and inter-feature coupling correlations on the state assessment, outputting the pure fault deviation after excluding normal operating condition drift. If the weighted Mahalanobis distance is less than a preset dynamic threshold, the motor is determined to be in a healthy state, and the current feature fluctuation is attributed to normal load or environmental changes. If the weighted Mahalanobis distance is greater than the preset dynamic threshold, the winding is determined to have potential damage. The dynamic threshold is also determined by the confidence interval of the dynamic health baseline.

[0101] Step S5 involves determining the winding damage level and outputting the identification result. When the multidimensional deviation exceeds a preset first safety threshold, the system immediately triggers a damage warning mechanism. Based on the magnitude of the multidimensional deviation value and the specific feature combination pattern exhibited in the feature vector, which includes components such as time temperature gradient, spatial temperature gradient, weighted harmonic characteristic index reflecting the winding state, and frequency band energy concentration, the damage state is precisely divided into three continuous stages:

[0102] The first stage is early-stage micro-loss. In this stage, the multidimensional deviation falls between the first and second thresholds. Characteristically, this manifests as a slight increase in the weighted harmonic characteristic index, reflecting the winding condition, or a slow increase in local temperature gradients within the temporal and spatial temperature gradients, while the mechanical vibration components and three-phase balance based on vibration spectrum analysis remain within normal ranges. This typically corresponds to the initial aging of the winding's enameled wire insulation or extremely minor inter-turn leakage.

[0103] The second stage is the intermediate development stage. At this point, the multidimensional deviation exceeds the second threshold. Characteristic manifestations include a significant increase in the negative-sequence current component, fluctuations in the zero-sequence component, and energy concentration in the vibration spectrum at the second harmonic. This indicates that the damage has evolved from localized point-like issues to regional faults, posing a risk of inter-turn short circuits, and has begun to affect the electromagnetic force distribution of the motor, leading to an accelerated response rate to the time-temperature gradient.

[0104] The third stage is severe failure. The multidimensional deviation exceeds the third limit threshold. A severe imbalance occurs in the three-phase voltage and current, and the spatial temperature gradient shows a sharp rise in localized stator temperature, accompanied by violent casing oscillations. This stage indicates that the windings are about to burn out or a large-scale short circuit has already occurred.

[0105] While determining the damage level, the system analyzes the imbalance of harmonic components of each phase current in the characteristic vector and combines this with the spatial coordinates of the abnormal temperature region fed back by the distributed fiber optic temperature measurement system. These spatial coordinates are calculated based on the spatial temperature gradient field, enabling a tendency inference of the damage location. For example, if the third harmonic increment of the A phase current is significantly higher than that of phases B and C, and the distributed fiber optic temperature measurement system shows that slots 15 to 20 near the A phase winding at the front end of the stator have both excessive temporal and spatial temperature gradients, the system infers that the damage point is located at that specific phase and slot level.

[0106] Finally, the system generates a detailed diagnostic report. This report includes not only the current damage level and location, but also a damage development trend prediction calculated based on the slope of multidimensional deviation changes over multiple consecutive monitoring periods. If the slope shows an exponential increase, the system will issue an emergency shutdown recommendation; if the slope is gentle, it will automatically suggest an optimal maintenance window based on the motor's maintenance history, thereby achieving preventative maintenance.

[0107] The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for quickly identifying damage of a motor winding of a mining machine, characterized in that, Includes the following steps: S1. Collect multi-source synchronous sensing signals during motor operation. The multi-source synchronous sensing signals include three-phase current signals, three-phase voltage signals, stator temperature distribution signals, and housing vibration acceleration signals. All signals are collected under a unified time reference by a high-precision synchronous sampling module, and the original signals are subjected to anti-aliasing filtering. S2. Constructing the dynamic multi-physics field feature vector of the motor winding: Based on the multi-source synchronous sensing signal, extract the real-time phase difference, weighted harmonic characteristic index, zero-sequence component amplitude and negative-sequence component imbalance that characterize the electromagnetic state, fuse the stator temperature gradient spatiotemporal fusion index that characterizes the thermodynamic state and the energy concentration of a specific frequency band of the vibration spectrum that characterizes the mechanical state, and arrange the above multi-dimensional features in a predetermined sequence to form a high-dimensional feature vector containing electromagnetic, thermodynamic and mechanical coupling information. S3. Establish a nonlinear mapping model of winding damage state. Use a deep temporal convolutional network that integrates physical mechanism constraints to jointly train historical normal operation data and known damage samples. The deep temporal convolutional network includes multiple causal convolutional layers and a self-attention weight mechanism based on physical prior guidance, which is used to capture the evolution law of feature vectors in the time dimension and the dynamic weights of key damage-sensitive features. S4. Construct a dynamic health baseline that is adaptively corrected according to load and environmental conditions, and compare the currently extracted time temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state with the theoretical expected value of the dynamic health baseline in real time, and calculate the multidimensional deviation degree that characterizes the degree of deviation of the state. The multidimensional deviation is quantified by introducing a weighted Mahalanobis distance with physical prior sensitivity weights. The weight coefficients are dynamically allocated based on the sensitivity of each physical feature to winding damage learned by the S3 model, so as to eliminate interference from dimensional differences and feature coupling, and accurately identify and exclude early damage signals after drift under normal operating conditions. S5. Determine the winding damage level and output the identification result. When the multidimensional deviation exceeds the preset threshold, trigger the damage warning mechanism and divide the damage state into three levels: early minor damage, mid-term development or severe failure, based on the magnitude of the multidimensional deviation and the feature combination pattern. At the same time, generate a diagnostic report that includes the damage location tendency and development trend.

2. The method according to claim 1, characterized in that, In step S1, Three-phase current signals are captured by electromagnetic transformers or Hall effect sensors deployed on the power cables at the input end of the motors of mining machinery, and three-phase voltage signals are captured by resistor-capacitor voltage divider circuits or voltage transformers. The temperature distribution signal of the stator is obtained by using a temperature-sensing optical fiber that is spirally arranged along the entire length of the winding axis in the stator slot. The temperature-sensing optical fiber is connected to a demodulator, and the real-time temperature is calculated by analyzing the amount of wavelength shift or the intensity change of backscattered light. The vibration acceleration signal of the housing is collected by a triaxial piezoelectric accelerometer installed on the surface of the motor bearing housing and the stator housing. The hardware-level timestamp alignment mechanism generates a global trigger pulse through the master clock divider, synchronously triggering the analog-to-digital converters of each signal, ensuring that the time synchronization error of different physical quantity signals is less than one sampling period, and that the sampling frequency is not less than the requirement of 20 sampling points within the power frequency period. The anti-aliasing filtering process uses a fourth-order Butterworth low-pass filter to filter out noise outside the sampling bandwidth and applies a mean filtering algorithm in the digital domain to suppress impulse noise.

3. The method according to claim 1, characterized in that, In step S2: Full-cycle Fourier transform is performed on the collected three-phase current and three-phase voltage signals to extract the fundamental components of each phase and calculate the real-time phase difference between the current and voltage of each phase. The extraction process of the weighted harmonic characteristic index reflecting the winding state includes: performing a discrete Fourier transform on the current sequence, extracting its spectral distribution, and then selecting the third, fifth, and seventh harmonic components for weighted synthesis, with the third harmonic having the highest weight coefficient, which is used to characterize the asymmetric current characteristics caused by the inter-turn short circuit. The zero-sequence component amplitude is obtained by vector summation of the three-phase current signals and taking its magnitude, and is used to reflect the degree of deterioration of the winding insulation to ground. The negative sequence unbalance is based on the symmetrical component method to extract the negative sequence component and calculate its ratio to the positive sequence component, which is used to characterize the degree of electrical asymmetry of the three-phase winding. The spatiotemporal temperature gradient index is calculated from the temperature data along the groove obtained by the distributed optical fiber temperature measurement system. Specifically, it includes: calculating the ratio of the temperature difference between adjacent temperature measurement points to the physical distance to obtain the spatial temperature gradient, and calculating the ratio of the temperature change of the same temperature measurement point at adjacent sampling times to the sampling time interval to obtain the temporal temperature gradient. The fusion of the spatial temperature gradient and the temporal temperature gradient identifies local overheated areas. The specific frequency band of the vibration spectrum refers to the frequency band within the range of the motor rotation frequency and its second to fifth harmonics. The energy concentration is obtained by integrating the power spectral density within this frequency band and is used to correlate the mechanical resonance characteristics caused by loosening of the stator winding, end deformation, or electromagnetic force imbalance.

4. The method of claim 1, wherein, In step S3: The causal convolutional layers in the deep temporal convolutional network employ an dilated convolutional structure, which expands the receptive field of the network by gradually increasing the dilation rate in order to capture the evolution patterns of feature vectors over long time scales. During the training phase of the model, an adversarial generative sample enhancement strategy based on physical constraints is introduced. The generator network generates synthetic damage data according to the multi-physics constraint equation of the motor, and the discriminator network distinguishes between scarce damage samples and synthetic samples. Through the alternating game between the two, a large number of weak damage samples with high realism and satisfying physical mechanisms are generated, thereby improving the generalization recognition ability of the deep learning model for early and hidden damage. The self-attention weighting mechanism evaluates the correlation between the current input features and potential damage patterns by scaling the dot product attention module, and dynamically allocates attention intensity based on the historical contribution of each dimension of features.

5. The method of claim 4, wherein the method further comprises: The self-attention weighting mechanism assigns preset high initial weights to the temporal temperature gradient, spatial temperature gradient, and weighted harmonic characteristic index reflecting the winding state during the initialization phase to enhance the response to insulation aging and inter-turn short circuits. During motor operation, the model automatically fine-tunes the weight coefficients of each dimension according to different operating conditions. When in the heavy-load start-up phase, it reduces the weight of transient vibration characteristics caused by current surges and correspondingly increases the attention weights of the weighted harmonic characteristic index and the temporal temperature gradient index reflecting the winding state.

6. The method of claim 1, wherein, In step S4: The adaptively corrected dynamic health baseline is dynamically corrected based on the motor's current load rate, ambient temperature, ambient humidity, and cumulative running time. The system uses a long short-term memory network module to record the characteristic evolution trajectory of the motor during the healthy operation phase throughout its entire life cycle. Combined with the current environmental parameters, the system uses a trained deep temporal convolutional network to deduce the theoretical expected values ​​and confidence intervals of each feature dimension, forming a healthy distribution manifold space.

7. The method of claim 1, wherein, The quantization calculation process of the weighted Mahalanobis distance includes: First, the feature vectors extracted in real time are centered and the mean of the healthy baseline is subtracted. Subsequently, a diagonal matrix consisting of weight coefficients is introduced, which are dynamically adjusted according to the sensitivity of each feature dimension to winding damage. The sensitivity is obtained by mapping the physical prior attention weights learned by the model in step S3. By eliminating interference caused by different units, magnitudes, and correlations among the various feature dimensions through matrix operations, the final output is a value reflecting the degree to which the current state of the motor deviates from the core healthy region.

8. The method of claim 1, wherein, In step S5, the damage state is divided into three levels: early minor damage, intermediate development, and severe failure. When the multidimensional deviation is between the first safety threshold and the second threshold, and the characteristic is that the weighted harmonic characteristic index reflecting the winding state increases or the spatiotemporal temperature gradient index increases, it is judged as an early micro-loss level. When the multidimensional deviation exceeds the second threshold, and the characteristics include the growth of the negative sequence current component, the fluctuation of the zero sequence component, and the energy concentration of the vibration spectrum at the second harmonic, it is judged to be of the medium-term development level. When the multidimensional deviation exceeds the third limit threshold, and the three-phase voltage and current show a severe imbalance, and the spatial temperature gradient shows a local temperature surge in the stator accompanied by violent shell oscillation, it is judged as a severe failure level.

9. The method for rapid identification of damage to motor windings in mining machinery according to claim 1, characterized in that, In step S5, the specific steps for generating a diagnostic report containing the predisposition and development trend of the injury are as follows: By analyzing the imbalance of harmonic components of each phase current in the feature vector and combining the spatial coordinates of the abnormal temperature area fed back by the distributed optical fiber temperature measurement system, the phase and slot level of the winding are inferred, and the damage location is located in a tendency manner. The damage development trend prediction in the diagnostic report is calculated based on the slope of the multidimensional deviation change over multiple consecutive monitoring periods. When the slope of the change shows an exponential growth trend, an emergency shutdown suggestion is output. When the slope of the change is gentle, the recommended maintenance window period is automatically calculated and output by combining the historical maintenance data of the motor.

Images