Inter-turn short circuit fault diagnosis method and system for drilling environment permanent magnet synchronous motor based on sample expansion

Abstract

The present application relates to a kind of based on sample expansion's while drilling environment permanent magnet synchronous motor interturn short circuit fault diagnosis method and system, method includes: collecting the real current signal sample of permanent magnet synchronous motor under while drilling environment and pre-processing;Based on the entropy increase information of sample, entropy reduction generation network is trained;Obtain entropy reduction with normal distribution noise signal, input into entropy reduction generation network to generate noise prediction value, using constraint modulation sparse generalization generation algorithm according to noise prediction value, and combining accumulated time series attenuation coefficient is carried out across multi-step reverse entropy reduction deduction, output virtual current signal sample;Virtual current signal sample and real current signal sample are fused to the iterative training of fault diagnosis model, to diagnose interturn short circuit fault in real time.Compared with prior art, the present application realizes the balanced high-fidelity sample generation of multi-grade motor fault, improves the generation network training efficiency, convergence stability and diagnostic reasoning robustness, and verifies the effectiveness under complex while drilling environment.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet synchronous motor fault diagnosis technology, and in particular to a method and system for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors in a drilling environment based on sample expansion. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in oil drilling due to their advantages such as high power density, high efficiency, and excellent speed regulation stability. Stator inter-turn short circuits are a common and typical electrical fault in these motors. The fault evolves rapidly, directly disrupting the electromagnetic balance of the three-phase windings, inducing abnormal fluctuations in generator speed and unstable output torque. Once the fault worsens, it can easily lead to sudden motor shutdown, equipment damage, and in severe cases, injury or death to on-site personnel. Furthermore, the industry currently faces the challenge of scarce fault samples and high labeling costs due to the limited sample size. Therefore, conducting specialized research on accurate diagnosis technology for inter-turn short circuit faults in PMSMs under limited sample conditions has significant engineering value and academic significance.

[0003] Currently, the industry mostly uses unsupervised adversarial generative data augmentation technology to construct virtual fault samples to make up for the scarcity of real inter-turn short circuit annotation samples in industrial scenarios. However, traditional unsupervised generative models generally suffer from defects such as pattern collapse, fault feature distortion, and poor model training convergence stability. In addition, for edge deployment scenarios of fault diagnosis algorithms, iterative optimization of existing embedded diagnostic algorithms after deployment and solidification is difficult. When the motor is adapted to new operating conditions, the static model has not learned the fault features of the new operating conditions, which can easily lead to a decrease in fault diagnosis accuracy.

[0004] Based on the above-mentioned existing technologies for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors under small sample conditions, and considering the constraints of industrial field operating conditions, the existing technologies have the following three major shortcomings, as detailed below: 1. Conventional unsupervised adversarial generation methods have limited fault feature modeling capabilities and generate fault samples with low fidelity. Conventional unsupervised adversarial generative methods for data augmentation rely on a two-way adversarial game between the generator and the discriminator to complete iterative training. The training objective is only to reduce the global distribution difference between real and fake samples. Adversarial training only performs global binary classification to distinguish between real and fake samples. The final output fault samples have distorted operating condition features, large detail deviations, and insufficient sample fidelity, making it difficult to support high-precision fault diagnosis.

[0005] Causes of defects: The pure adversarial training mode takes global true / false discrimination as the only optimization goal, which belongs to the data-driven binary adversarial training paradigm, and the generator feature learning is one-sided; Industry solutions: Simply adjusting the training parameters such as the learning rate and iteration rounds of the adversarial generator network cannot change its underlying binary adversarial training logic, cannot guarantee the stability of small sample training and the fidelity of fault sample generation, and cannot meet the high-precision sample requirements of industrial diagnosis.

[0006] 2. Conventional unsupervised adversarial generative models have poor performance in synchronously generating multiple types of faults and are prone to mode collapse. Existing single-mode generation models tend to fit a single fault pattern. For multi-level gradient faults such as minor, moderate, and severe inter-turn short circuits, they cannot achieve simultaneous and balanced generation of fault samples of multiple categories and different damage levels. During training, the model prioritizes fitting severe short circuit samples with high feature recognition, ignoring weak early short circuit fault features. This leads to an imbalance in the number of gradient fault data generated and feature convergence, ultimately causing model mode collapse and only being able to output single-type fault samples.

[0007] For example, the invention disclosed in CN115688553A discloses a method for diagnosing inter-turn short-circuit faults in unbalanced samples of permanent magnet synchronous motors. It combines generative adversarial networks and sparse representation theory to perform sample expansion and noise reduction. Although the use of sparse theory improves the noise resistance and overfitting resistance of the deep residual shrinkage network model, the sample generation process is still based solely on generative adversarial networks. It still suffers from the problems of imbalance of multi-level fault samples and pattern collapse under small sample size. Furthermore, the training process is lengthy and inefficient.

[0008] Causes of the defects: Existing adversarial training generation does not incorporate multi-level fault category constraints throughout the entire process, lacks a differentiated category control mechanism in the training process, and the discriminator only distinguishes between real and fake samples, without differentiating between the severity of short-circuit faults. During model training, the model autonomously favors learning high-discrimination, severe short-circuit features, without mandatory fitting constraints on weak early short-circuit features. The category learning weights are completely autonomous and uncontrollable, ultimately leading to a bias towards generating a single fault after training convergence, inducing pattern collapse. Industry challenges: Manually splitting the model to generate various fault samples significantly increases computational costs and training time, making it difficult to achieve balanced generation of multiple types of inter-turn short-circuit faults at low cost.

[0009] 3. The fixed diagnostic model at the factory edge lacks autonomous iteration capability, has poor adaptability to new operating conditions, and the misdiagnosis rate is rising. The current mainstream architecture for inter-turn short circuit diagnosis of motor faults is a separate architecture with offline training in the cloud and fixed deployment at the edge. The diagnostic model parameters on the edge monitoring terminal are fixed after training and do not have the ability to learn and update with new operating condition data and new short circuit fault data. During long-term operation of industrial motors, the characteristics of inter-turn short circuit faults under new operating conditions will emerge due to load fluctuations and speed changes. The fixed model does not learn the new fault characteristics, and the original discrimination criteria become invalid, which directly leads to a significant increase in the probability of misdiagnosis and missed diagnosis of faults.

[0010] Causes of the Defect: The existing cloud-edge architecture suffers from fragmented data interaction, lacking a closed-loop linkage mechanism for incremental optimization in the cloud and dynamic iteration at the edge. The architecture is only adapted to static fault identification tasks and lacks a model iteration update interface adapted to real-time data streams. Industry Challenges: The existing cloud and edge are completely offline and separated, preventing real-time data interaction between the two ends. Traditional update methods require collecting all historical data, combining it with new data, retraining the model across the cloud, and then redeploying it to the edge. This cumbersome model update process necessitates equipment shutdown for adaptation training, making it impossible to achieve uninterrupted model iteration at the edge in industrial settings and failing to adapt to the dynamic and ever-changing operating conditions of motors.

[0011] 4. Existing motor fault simulation platforms struggle to achieve simultaneous simulation of high-temperature and high-vibration combined operating conditions, and the reliability of fault feature reproduction and algorithm verification cannot be guaranteed. Most current mainstream simulation platforms for inter-turn short-circuit faults in permanent magnet synchronous motors (PMSMs) can only independently apply high-temperature environments or single vibration excitations, making it difficult to replicate the real working conditions of oil drilling where high temperatures and strong vibrations coexist. The inter-turn short-circuit fault characteristics collected and verified under separate, single-environment conditions deviate from the fault characterization under complex and harsh downhole conditions. Diagnostic models trained based on such experimental data are prone to misjudgment and underjudgment, ultimately leading to ineffective maintenance decisions.

[0012] Cause of defects: The temperature control unit and vibration excitation unit in the fault simulation system lack coordinated linkage, do not support the synchronous coupling application of temperature and vibration loads, and cannot reproduce the evolution mechanism of motor fault characteristics under multi-physics field coupling.

[0013] Industry challenges include: conducting single-environment tests in stages makes it difficult to characterize the fault feature shift caused by high-temperature-vibration coupling; the development threshold for integrated composite environment simulation devices is high, and existing commercial maintenance test platforms generally do not have the ability to simulate such composite working conditions, resulting in diagnostic algorithms lacking test verification conditions that closely resemble downhole working conditions. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of conventional unsupervised adversarial generation methods, such as limited fault feature modeling capabilities, low fidelity of generated fault samples, and low efficiency, and to provide a method and system for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors in a drilling environment based on sample expansion.

[0015] The objective of this invention can be achieved through the following technical solutions: A method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors in a drilling environment based on sample expansion includes the following steps: Real current signal samples of permanent magnet synchronous motors under normal operation and inter-turn short circuit fault conditions in the drilling environment were collected and preprocessed. Based on the preprocessed real current signal samples, a normal distribution noise signal for training is added, and entropy increase evolution is performed according to the preset time-series decay coefficient. The pre-constructed entropy decrease generation network is trained based on the entropy increase evolution data. The normally distributed noise signal used for entropy reduction is obtained and input into the trained entropy reduction generation network to generate noise prediction values. The constrained modulation sparse generalization generation algorithm is used to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the time-series decay coefficient accumulated during the entropy increase evolution process, and outputs the restored virtual current signal sample. The virtual current signal samples are fused with the real current signal samples to construct a complete fault dataset; the complete fault dataset is then used to iteratively update the pre-constructed fault diagnosis model. An updated fault diagnosis model is used to perform real-time diagnosis of the operating data of the permanent magnet synchronous motor to determine whether an inter-turn short circuit fault has occurred.

[0016] Furthermore, the processing steps of the constrained modulation sparse generalization generation algorithm include the following steps: S101: Regarding step size Initialize with the target fault category; S102: Subtract the randomly generated entropy from the normally distributed noise signal as... t The high-entropy perturbation state corresponding to time 1 , among which, time t The initial value is the total number of time periods of the entropy increase evolution; S103: Based on the high-entropy perturbation state Given the corresponding target fault category, constrained noise prediction values ​​are generated through an entropy reduction generation network. According to the high-entropy perturbation state With empty category labels, unconstrained noise predictions are generated through an entropy reduction generation network; The constrained noise prediction value and the unconstrained noise prediction value are weighted and fused to obtain the current... t The final network prediction noise corresponding to the given time; S104: If Then based on Gaussian random noise sampling and cross The mean state derived by the inverse entropy reduction step is determined. High-entropy perturbation state at time 1 Then return to step S103, where the state mean is determined based on the accumulated temporal decay coefficient and the final network prediction noise corresponding to the current moment; like Then the original state signal is determined. and the original state signal Output as a virtual current signal sample.

[0017] Furthermore, the cross The expression for calculating the state mean in the reverse entropy reduction derivation is: In the formula, From t The moment begins to cross The mean state derived from the reverse entropy reduction step. From the initial moment to The cumulative decay coefficient at time step, , for i The temporal decay coefficient at time step, From the initial moment to The cumulative decay coefficient at time step, The final network prediction noise is generated by entropy reduction in the network output. for t The square of the state variance at time t; The High-entropy perturbation state at time 1 The calculation expression is: In the formula, for t The state variance at time t, It is Gaussian random noise.

[0018] Furthermore, the entropy reduction generation network is a symmetric encoding and decoding architecture, including an encoding convolution module, a decoding convolution module, and a bottleneck convolution module. The encoding convolution module and the decoding convolution module are symmetric structures. The encoding convolution module receives the input of a high-entropy perturbation state and performs downsampling step by step through multi-level encoding convolution. Each level of encoding convolution is connected to the corresponding decoding convolution of the decoding convolution module through a residual jump fusion path. The bottom-level encoded features output by the encoding convolution module are input into the bottleneck convolution module to complete the deep feature mapping. Then, the decoding convolution module upsamples the features step by step and fuses the output features of the corresponding level of encoding convolution to finally output the noise prediction value. The entropy reduction generation network further includes a time-series information embedding module and a multi-gradient inter-turn short-circuit category coding module. The time-series information embedding module is used to extract prior knowledge of the operating conditions of the permanent magnet synchronous motor and form prior knowledge constraints. The multi-gradient inter-turn short-circuit category coding module is used to generate fault category modulation constraints based on the input target fault category. The prior knowledge constraints and fault category modulation constraints of the time-series information embedding module and the multi-gradient inter-turn short-circuit category coding module are injected in parallel into all network layers of the coding convolution module, the decoding convolution module, and the bottleneck convolution module.

[0019] Furthermore, the training process of the entropy reduction generation network includes the following steps: Obtain the entropy-incremented evolved data used for training the entropy-reduction generation network, and input it into the entropy-reduction generation network; The entropy reduction generation network is trained collaboratively through multiple distributed clusters, which employ a parallel architecture with multiple GPU nodes to split the training tasks. In each iteration, the final network prediction noise output by the entropy reduction generation network is compared with the normal distribution noise signal used for training to calculate the loss function; based on the calculated value of the loss function, gradient feedback is obtained to update the network parameters of the entropy reduction generation network. Construct a shadow generation network for the entropy reduction generation network, and update the parameters based on the network parameters of the entropy reduction generation network in each iteration. Use the network parameters of the shadow generation network obtained after the entropy reduction generation network completes its iteration as the network parameters trained by the entropy reduction generation network.

[0020] Furthermore, the shadow generation network uses an exponentially weighted average method for parameter updates, and the parameter update expression of the shadow generation network is: In the formula, Here are the network parameters of the smoothed shadow generation network in the k-th iteration. It is an exponential decay factor. These are the network parameters of the entropy-reduced generation network after the k-th iteration; The multi-distributed cluster calculates the global average gradient by aggregating the local gradients of each GPU node, and uses this as the gradient feedback for the entropy reduction generation network.

[0021] Furthermore, the convolutional layer in the entropy reduction generation network is a one-dimensional spatial channel separation convolution. This one-dimensional spatial channel separation convolution first fuses the spatial information of each channel, and then fuses the information output by each channel.

[0022] Furthermore, the entropy reduction generation network employs a cosine annealing modulation strategy to smoothly decay the learning rate. The expression for the smooth decay of the learning rate is as follows: In the formula, Let be the current learning rate in the k-th iteration. The initial learning rate, The minimum learning rate is set. It represents the total number of training rounds.

[0023] This invention also provides an inter-turn short-circuit fault diagnosis system for implementing the above-described sample-expansion-based inter-turn short-circuit fault diagnosis method for permanent magnet synchronous motors in drilling environments, comprising: The data acquisition module is used to collect real current signal samples of the permanent magnet synchronous motor under normal operation and inter-turn short circuit fault conditions in the drilling environment, and to perform preprocessing. The entropy reduction generation network training module adds a normal distribution noise signal for training to the preprocessed real current signal samples, performs entropy increase evolution according to the preset time-series decay coefficient, and trains the pre-constructed entropy reduction generation network based on the entropy increase evolution data. The virtual current signal sample generation module is used to acquire the normally distributed noise signal for entropy reduction, input it into the trained entropy reduction generation network to generate noise prediction values, and use the constrained modulation sparse generalization generation algorithm to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the time-series decay coefficient accumulated during the entropy increase evolution process, and output the restored virtual current signal sample. A complete fault dataset generation module is used to fuse the virtual current signal samples with the real current signal samples to construct a complete fault dataset. The model update module is used to iteratively update the pre-built fault diagnosis model using the complete fault dataset; The real-time diagnostic module is used to perform real-time diagnostics on the operating data of the permanent magnet synchronous motor using the updated fault diagnosis model to determine whether an inter-turn short circuit fault has occurred.

[0024] Furthermore, the system is based on an edge and cloud processing architecture, with the data acquisition module and real-time diagnostic module running at the edge and uploading the acquired real current signal samples to the cloud; The complete fault dataset generation module and model update module run in the cloud and send the updated fault diagnosis model parameters to the fault diagnosis model at the edge.

[0025] Compared with the prior art, the present invention has the following advantages: (1) This invention proposes a sparse generalization generation network with multi-gradient fault category constraint modulation. It adopts k-step step-by-step inverse sparse entropy reduction sampling to realize high-fidelity sample generation for fault balance of multi-level permanent magnet synchronous motors, which can significantly reduce the number of iteration steps and improve training efficiency. Furthermore, by first performing entropy increase evolution based on real current signal samples to train the signal restoration process of the entropy reduction generation network, entropy reduction deduction is realized. This enables a scheme that can be directly restored to virtual current signal samples based on noise signals in practical applications, which greatly improves the richness and efficiency of sample amplification.

[0026] (2) This invention designs Markov entropy increase and entropy decrease sparse generalization denoising generation logic, which relies on the sparse generalization mechanism to enhance the sample intra-class generalization performance and replaces the traditional global binary classification adversarial mode from the bottom-level generation logic; it is paired with a temporal U-shaped residual encoding and decoding network to retain multi-scale fault details; it adopts a cluster collaborative annealing mixed precision training strategy, and stabilizes the entire training process through shadow weight smoothing, distributed gradient aggregation, cosine annealing learning rate, and BF16 mixed precision.

[0027] This method enables the generation of virtual fault time-series feature signals that closely match real motor faults under small sample conditions, effectively improving sample fidelity and intra-class generalization ability, and providing reliable data support for high-precision fault diagnosis.

[0028] (3) Existing fault sample expansion schemes mostly adopt unsupervised GAN-type adversarial networks, which rely solely on global binary classification optimization of real and fake samples, without a fault level differentiation constraint mechanism. This invention abandons the underlying logic of traditional binary adversarial training, embeds dual-condition modulation branches of working conditions and fault categories at all levels of the U-shaped generator network, and introduces differentiated category constraints by adopting a dual-branch noise weighted fusion algorithm. Combined with a step-back inverse unified sampling mechanism, a single model can uniformly control the generation process of multiple gradient faults such as light / medium / severe inter-turn short circuits, without the need to split multiple model training.

[0029] A single network can generate multi-level fault samples in a balanced and stable manner, completely suppressing pattern collapse and significantly reducing the computing power and time costs of expanding multi-category faults.

[0030] (4) The present invention establishes a cloud-edge bidirectional collaborative closed-loop architecture: the edge continuously transmits new working condition fault data, the cloud integrates real samples and virtual expanded samples, and continuously iterates and optimizes the diagnostic model based on incremental learning; the optimized model parameters can be sent to the edge terminal in real time to realize hot update of edge devices without shutdown.

[0031] Implementation: The edge diagnostic model can dynamically adapt to varying loads, rotational speeds, and drilling environments under conditions of high temperature and strong vibration, continuously reducing the rate of misdiagnosis and missed diagnosis of faults, and avoiding production losses caused by equipment downtime for upgrades.

[0032] (5) Existing maintenance test equipment for inter-turn short circuit faults of permanent magnet synchronous motors only supports independent loading of a single high-temperature environment or a single vibration condition. The temperature control unit and the vibration excitation unit are independent of each other and cannot work together. They cannot replicate the real complex working conditions of high temperature-strong vibration coupling in oil drilling. The collected fault features are seriously distorted, and the reliability of algorithm verification is low. This invention innovatively builds an integrated drilling coupled working condition simulation maintenance platform to achieve synchronous coupling and accurate loading of temperature field and vibration load. It can truly reproduce the evolution mechanism and feature shift law of inter-turn short circuit faults under the coupling effect of multiple physical fields in the well. It can accurately collect multi-level fault sample data under complex and harsh working conditions, filling the technical defects of existing test equipment that cannot carry out drilling coupled working condition fault verification. It is fundamentally different from the existing single-environment test platform in terms of equipment architecture, working condition simulation method, fault data acquisition and verification logic. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a method for diagnosing inter-turn short-circuit faults in a permanent magnet synchronous motor in a drilling environment based on sample expansion, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the training process of an entropy reduction generation network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a constrained modulation encoding and decoding temporal U-shaped entropy reduction network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a cloud-edge bidirectional collaborative incremental training and diagnostic architecture provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a sample fidelity experiment result provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of an anti-mode collapse experiment provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the verification of inter-class generalization ability provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the inter-turn short circuit classification results of a permanent magnet synchronous motor provided in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the diagnostic accuracy after expanding the training set based on different generation algorithms, provided in an embodiment of the present invention. Figure 10 This is a schematic diagram of a training loss convergence result provided in an embodiment of the present invention; Figure 11 This is a schematic diagram illustrating the structural differences between one-dimensional traditional convolution and one-dimensional spatial channel separation convolution provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0036] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0037] Example 1 like Figure 1 As shown, this embodiment provides a method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors in a drilling environment based on sample expansion, including the following steps: S1: Collect real current signal samples of the permanent magnet synchronous motor under normal operation and inter-turn short circuit fault conditions in the drilling environment, and perform preprocessing. S2: Based on the preprocessed real current signal samples, add training normal distribution noise signals, perform entropy increase evolution according to the preset time-series decay coefficient, and train the pre-constructed entropy decrease generation network based on the entropy increase evolution data; S3: Obtain the normally distributed noise signal for entropy reduction, input it into the trained entropy reduction generation network to generate noise prediction values, and use the constrained modulation sparse generalization generation algorithm to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the accumulated time-series decay coefficient during the entropy increase evolution process, and output the restored virtual current signal sample. S4: Fuse virtual current signal samples with real current signal samples to construct a complete fault dataset; use the complete fault dataset to iteratively update the pre-constructed fault diagnosis model; S5: The optimized fault diagnosis model is used to perform real-time diagnosis of the operating data of the permanent magnet synchronous motor to determine whether an inter-turn short circuit fault has occurred.

[0038] Specifically, the processing steps of the constrained modulation sparse generalization generation algorithm in step S2 include the following steps: S101: Regarding step size Initialize with the target fault category; S102: Based on randomly generated entropy reduction using normally distributed noise signal as... t The high-entropy perturbation state corresponding to time 1 , among which, time t The initial value is the total number of time periods of entropy increase evolution in step S2; S103: This refers to the processing of the entropy-reducing generation network, which is consistent with the training process of the entropy-reducing generation network in step S2: based on the high-entropy perturbation state... Given the corresponding target fault category, constrained noise prediction values ​​are generated through an entropy reduction generation network. According to the high-entropy perturbation state With empty category labels, unconstrained noise predictions are generated through an entropy reduction generation network; The constrained noise predictions and unconstrained noise predictions are weighted and fused to obtain the current... t The final network prediction noise corresponding to the given time; S104: If Then based on Gaussian random noise sampling and cross The mean state derived by the inverse entropy reduction step is determined. High-entropy perturbation state at time 1 and will Assigned value Then return to step S103, where the state mean is determined based on the accumulated temporal decay coefficient and the final network prediction noise corresponding to the current moment; like Then the restoration is complete, and the high-entropy perturbation state at this point is taken as the original state signal. and the original state signal Output as a virtual current signal sample.

[0039] The following is a detailed description of this plan: Entropy increase execution process: This entropy increase evolution process is based on Markov chains for state updates, satisfying the Markov assumption of no aftereffect: the hidden state at time t The distribution is determined solely by the state at the previous time step. With random noise The decision is independent of earlier historical states; the recursive formula for the state is: (1) The meanings of the variables in the formula are as follows: The current state of the Markov chain at step t; : The historical state at step t-1; Independent and identically distributed standard random noise (usually Gaussian noise, providing the source of randomness); : Time decay coefficient, which controls the weight of retaining historical states; : Entropy increment step size coefficient, directly adjusts the entropy increase rate at each step; the maximum value of t will be initialized to 1000. The values ​​will be uniformly distributed based on the maximum value of t, between 0.0001 and 0.02. Therefore, it can be obtained.

[0040] Entropy reduction process: Equation (2) constructs a reverse Markov transition distribution using Bayes' theorem to control the entropy reduction process; Equation (3) combines the closed nature of the Gaussian distribution under linear transformation and the temporal consistency condition of the Markov chain to give the Gaussian analytical expressions for various probability distributions in Equation (2). The forward entropy increase process will continuously add random noise, constantly expanding the distribution variance and increasing the system information entropy; entropy reduction is its reverse derivation process: with a high-entropy perturbation state Compared with the original sample Given the conditions, deduce the intermediate state with less noise and lower entropy by working backwards. .

[0041] (2) (3) in The conditional probability distribution representing the signal at different times; Multivariate Gaussian (normal) distribution It is the mean vector. Let be the covariance matrix. The cumulative decay coefficient from the initial time to time t. , representing the total retention rate of the initial information after multiple evolution steps.

[0042] Based on formulas (1)-(3), the posterior distribution can be derived. It satisfies formula (4).

[0043] (4) in , Representative and Irrelevant items.

[0044] Equation (5) gives the mean of the posterior distribution P. and variance P follows a Gaussian distribution .variance Based solely on the time series coefficients and entropy increase step size Determine and control the noise fluctuation range of the reverse entropy reduction; mean. Generative Network with Entropy Reduction Predict and cancel positive superposition noise to restore the low-entropy signal from the high-entropy state, thus completing a single-step entropy reduction.

[0045] (5) Equation (6) is derived from equations (2) to (4) to obtain the conditional distribution. mean The variance corresponding to this distribution can be solved recursively using the variance formula in equation (5).

[0046] (6) in the formula The mean value of the state with k-step reverse entropy reduction is obtained through the cumulative decay coefficient. Scaling the current state Network noise prediction value This completes multi-step denoising and mean estimation.

[0047] Traditional generative models only support single-step inverse denoising, requiring iterative derivation round by round, resulting in a significant decrease in generation efficiency with increasing step count. This invention, based on a single-step entropy-reduction Gaussian analytical framework, derives a closed-form solution for k-step inverse entropy reduction. It eliminates the need for layer-by-layer calculation of intermediate states, directly obtaining low-entropy samples at k-step intervals from high-entropy states in one step. Relying on the cumulative decay coefficient, it completes multi-step denoising estimation in a single step, significantly reducing the number of iterations and substantially improving sample generation efficiency.

[0048] Table 1 As shown in Table 1, Algorithm 1 is a constrained modulation sparse generalization generation algorithm sampling algorithm. It takes a trained noise prediction network and noise scheduling parameters as input, uses standard Gaussian noise as the initial high-entropy signal, and iteratively performs multi-step entropy reduction and restoration from the maximum time step T with a step size k. During the iteration, the network predicts two classes of noise: constrained and unconstrained, and then performs weighted... The final predicted noise is obtained by fusing the data, and then the mean of the inverse distribution across k steps is calculated. Intermediate states are updated with Gaussian noise superimposed on non-terminating time series, iterating until... The virtual fault characteristic signal is generated directly at that time.

[0049] The algorithm uses step-back sampling to reduce computation, relies on dual-branch noise fusion to achieve controllable modulation of fault categories, and uses Markov Gaussian posterior to achieve stable entropy reduction, thus efficiently generating fault features with good generalization.

[0050] The network is configured to output two types of noise prediction values: constrained noise prediction. Introducing fault category priors to characterize noise distribution guided by fault labels; unconstrained noise prediction. It does not rely on fault labels and characterizes the general noise distribution. By weightedly fusing two prediction results, it takes into account both fault prior information and general noise features to obtain the final predicted noise, thereby increasing the generalization ability of the final model in generating single-category data and avoiding overly simplistic generated data.

[0051] The constrained noise calculation value is the current state Timing step size t, target fault category, etc., are input as follows: Figure 2 The constrained modulation encoding and decoding temporal U-shaped entropy reduction network shown is used to obtain the value. The unconstrained noise calculation value is obtained by taking the current state... The time step size t and the empty category label are input as follows: Figure 2 The constraint modulation encoding and decoding timing U-shaped entropy reduction network shown is obtained.

[0052] Existing fault feature generation schemes based on entropy reduction generation networks mostly employ unconstrained prediction modes, lacking a category-constrained input mechanism. During sampling, the model cannot incorporate prior information about fault category and severity, making controllable generation difficult, and a single network struggles to stably output inter-turn short-circuit samples with different damage levels. To address this, this invention introduces a constrained noise prediction branch to achieve controllable generation of fault samples of varying severity. A constrained modulation scheme is proposed, enabling a single model to accurately generate multiple fault features based on category constraints, ensuring the generalization of the generated features.

[0053] In addition, the entropy reduction generation network is a symmetric encoding and decoding architecture, including an encoding convolution module, a decoding convolution module and a bottleneck convolution module. The encoding convolution module and the decoding convolution module are symmetric structures. The encoding convolution module receives the input of a high-entropy perturbation state and performs downsampling step by step through multi-level encoding convolution. Each level of encoding convolution is connected to the corresponding decoding convolution of the decoding convolution module through a residual skip fusion path. The bottom-level encoded features output by the encoding convolution module are input into the bottleneck convolution module to complete the deep feature mapping. Then, the decoding convolution module upsamples the features step by step and fuses the output features of the corresponding encoding convolution to finally output the noise prediction value. The entropy reduction generation network also includes a time-series information embedding module and a multi-gradient inter-turn short-circuit category coding module. The time-series information embedding module is used to extract prior knowledge of the operating conditions of the permanent magnet synchronous motor and form prior knowledge constraints. The multi-gradient inter-turn short-circuit category coding module is used to generate fault category modulation constraints based on the input target fault category. The prior knowledge constraints and fault category modulation constraints of the time-series information embedding module and the multi-gradient inter-turn short-circuit category coding module are injected in parallel into all network layers of the coding convolution module, the decoding convolution module, and the bottleneck convolution module.

[0054] Specifically, Figure 3This is a diagram of the constrained modulation encoding and decoding timing U-shaped entropy reduction generation network structure of the present invention, used for small sample data enhancement of one-dimensional timing fault signals of inter-turn short circuits in permanent magnet synchronous motors. The network adopts a symmetric encoding and decoding architecture. The diagram distinguishes the operation paths such as encoding convolution, decoding convolution, bottleneck convolution, internal residual, residual jump fusion, multi-condition constraint modulation, and global data connection. The orange waveform on the left is the real motor timing input signal, which is downsampled and compressed step by step through four levels of encoding convolution to enhance the timing and feature channels. The output features of each encoding layer are reserved for the decoding side through the residual jump fusion path. The bottom layer encoding features are fed into a double-layer bottleneck convolution to complete the depth feature mapping, and then upsampled step by step through symmetric decoding convolution. After fusing the shallow features of the corresponding encoding layers, the virtual fault timing noise reduction sample of the orange waveform on the right is finally output. The feature channels and timing length dimensions are labeled at each level.

[0055] The network is equipped with two major condition control branches: the left vertical block is the time-series information embedding module, which extracts prior operating conditions such as speed and load; the right hierarchical dashed box is the multi-gradient inter-turn short-circuit category encoding, corresponding to fault types such as normal, light, medium and severe short circuits. Both the prior operating conditions and fault types are transformed into coded information that the model can recognize through the encoding function. The two are injected into the encoding, bottleneck and decoding of all network layers in parallel through the gray dashed multi-condition constraint modulation path. Relying on differentiated category constraints, the single model is able to generate multiple types of fault samples in a balanced manner to suppress mode collapse; the bottom orange data connection module gathers the time-series operating conditions, fault categories and full network features to complete the unified global condition control.

[0056] This network differs from traditional unsupervised adversarial generative frameworks and convolutional networks. It relies on multi-layer residual structures and temporal condition embedding to improve the fidelity of virtual fault samples. Combined with category-restricted modulation, it achieves balanced expansion of multi-level short-circuit samples. The generated complete hybrid dataset is supplied to the cloud for training of incremental diagnostic models, supporting the entire cloud-edge collaborative incremental diagnostic method to complete the iterative diagnosis of inter-turn short circuits in edge motors without stopping operation.

[0057] Category-constrained modulation information, combined with temporal information, is fed into the encoding and decoding modules at each level of the network via the bottom data connection path. Category embedding information persists throughout the encoding and decoding stages, continuously impacting the bottleneck convolutional module and all levels of the encoding and decoding convolutional units. This category-conditional embedding introduces prior constraints on fault categories, guiding the network to learn the feature distributions corresponding to different fault types, thus achieving controllable feature generation.

[0058] Table 2 shows an example of a working condition with varying speed and load.

[0059] Table 2. Operating Conditions with Varying Speed ​​and Load Preferably, the convolutional layers in the aforementioned entropy reduction generation network are one-dimensional spatial channel-separating convolutions. These one-dimensional spatial channel-separating convolutions first fuse spatial information from each channel, and then fuse the information output from each channel. That is, by using spatial channel separation technology, the convolutional modules in the constraint modulation sparse generalization generation algorithm are made lighter, requiring fewer trainable parameters, thus ensuring generation quality while reducing training time.

[0060] The difference between one-dimensional conventional convolution and one-dimensional spatial channel-separating convolution is as follows: Figure 11 As shown, traditional convolution directly fuses spatial and channel information simultaneously, such as... Figure 11 As shown in 'a'; while one-dimensional spatial channel-separated convolution first fuses spatial information and then fuses the information from each channel, as shown in 'a'. Figure 11 As shown in b in the diagram. Theoretically, for a one-dimensional traditional convolutional layer with C input channels, D output channels, and k kernels, the number of trainable parameters is: For a one-dimensional spatial channel-separating convolutional layer, the number of trainable parameters is... Therefore, the total number of trainable parameters is reduced. .

[0061] For the actual use of this solution, such as Figure 3 Regarding the constrained modulation sparse generalization generation algorithm shown, if a traditional one-dimensional convolutional layer is used, the total number of trainable parameters is 51,186,177; if a one-dimensional spatial channel separation convolutional layer is used, the total number of trainable parameters is 28,123,137. Based on dual Tesla A100 GPUs, training 5000 generations of the constrained modulation sparse generalization generation algorithm using traditional one-dimensional convolutional layers takes 16 hours, 14 minutes, and 48 seconds, while training 5000 generations of the constrained modulation sparse generalization generation algorithm using one-dimensional spatial channel separation convolutional layers takes 8 hours, 33 minutes, and 2 seconds.

[0062] like Figure 2 As shown, the training process of the entropy-reducing generative network includes the following steps: Obtain the entropy increase evolution data for training the entropy decrease generation network and input it into the entropy decrease generation network; The entropy reduction generation network is trained collaboratively through multiple distributed clusters, which use a parallel architecture with multiple GPU nodes to split the training tasks. In each iteration, the loss function is calculated by comparing the final network prediction noise (generated in step S103) output by the entropy reduction generation network with the training noise signal; the gradient feedback is obtained based on the calculated value of the loss function to update the network parameters of the entropy reduction generation network. Construct a shadow generation network for the entropy-reducing generation network, and update the parameters based on the network parameters of the entropy-reducing generation network in each iteration. Use the network parameters of the shadow generation network obtained after the entropy-reducing generation network completes its iteration as the network parameters of the trained entropy-reducing generation network.

[0063] Preferably, the shadow generation network uses an exponentially weighted average method for parameter updates. Equation (7) is the iterative smoothing calculation formula for robust exponential parameter averaging, which is used to smoothly update the shadow weights after each optimization step, thereby improving the robustness of the inference stage. This exponentially weighted average update method can weaken the disturbance caused by the fluctuation of parameters in a single iteration, making the shadow parameters used for inference more stable and more robust.

[0064] (7) in It is the smoothed shadow parameter of the k-th iteration. This represents the network parameters of the entropy-reduced generation network after the k-th iteration. As an exponential decay factor, historical parameters from more distant iterations will be affected. Higher powers are assigned smaller weights, achieving a smooth effect of retaining the emphasis on historical parameters and weakly fusing new real-time parameters; It serves as a normalization mechanism, keeping the magnitude of the weighted sum stable.

[0065] The weights of the shadow network are used as the network parameters of the final entropy-reduction generator network. Because the weights of the shadow network are updated differently, their updates are smoother, ensuring that the quality of the generated samples before and after iterative optimization does not differ significantly. This avoids sudden changes in sample generation quality caused by problems such as gradient explosion and vanishing gradients encountered during model training. Preferably, the multi-distributed clusters aggregate the local gradients of each GPU node to solve for the global average gradient, which serves as the gradient feedback for the fault diagnosis model and the shadow generation network.

[0066] Multi-distributed cluster collaborative training adopts a multi-node parallel architecture to split training tasks, synchronously aggregate gradients, break through the limitations of single-card memory and computing power, and improve training throughput and computational efficiency; Equation (8) is used to summarize the local gradients of each node and solve the global average gradient: (8) In the formula, This represents the local gradient calculated at the i-th node. To achieve the average gradient after global synchronization, all nodes update their network parameters synchronously based on this unified gradient.

[0067] Preferably, the entropy reduction generation network uses a cosine annealing modulation strategy to smoothly decay the learning rate, and Equation (9) is the corresponding learning rate update formula. This formula utilizes the periodic smooth decline characteristic of the cosine function to allow the learning rate to continuously and slowly decay from the initial value to the lower limit value with the number of iterations, avoiding training oscillations caused by a sudden drop in the learning rate, and effectively improving the model training stability and convergence effect. The learning rate is relatively large in the early stage of training, resulting in rapid convergence; the learning rate is gradually reduced in the later stage of training to finely optimize the model parameters.

[0068] (9) In the formula, This represents the current learning rate in the k-th iteration; The initial learning rate, The minimum learning rate set; It represents the total number of training rounds.

[0069] Preferably, the forward computation process is automatically converted to bfloat16 format to improve computational efficiency, while model parameters and gradients retain the FP32 high-precision format to ensure numerical stability. This strategy can reduce memory usage, accelerate model convergence, and stabilize training results.

[0070] In this embodiment, the training process of the entropy-reducing generation network uses entropy-increasing evolution data as input, which is then fed into the generation network for cluster-coordinated multi-GPU parallel training. The training process employs BF16 mixed precision for forward computation to reduce memory overhead. After forward propagation, loss calculation is performed, and the generation network parameters are updated based on gradient feedback. Then, the shadow generation network parameters are updated based on historical generation network parameters. Subsequently, a cosine annealing modulation strategy is introduced to achieve smooth learning rate decay adjustment. The updated network parameters are then fed back to the multi-GPU cluster for the next iteration. The backward gradient propagation stage uses FP32 precision to ensure gradient numerical stability. The entire process, through a dual-parameter update mechanism for the main network and shadow network combined with a dynamic learning rate strategy, stably optimizes the training process and alleviates the problems of training oscillations and mode collapse in the generation model.

[0071] The specific process is shown in Table 2. in, For the entropy reduction network to be trained, Represents all learnable parameters of the network; The predicted noise is the output noise of the network. denoted as t, where t is the time step of the diffusion model; con is the conditional embedding vector; ucon is the unconditional vector. This is normally distributed noise; To train the loss function, The learning rate; The loss function with respect to network parameters The gradient is used to perform gradient descent to complete the iterative update of network parameters.

[0072] The noisy signal and time step information are determined by the entropy increase process; signal regularization involves subtracting the mean of the collected fault sample signal and dividing by the standard deviation; the ut and time step t from the entropy increase process, along with the class determination information con and empty class information, are input into the entropy reduction network to obtain the predicted noise. ,calculate The mean square error of the normally distributed noise signal used for training is used as the loss, and the entropy reduction network parameters are updated based on the loss value.

[0073] Preferred, such as Figure 4 As shown, the above-mentioned cloud-edge bidirectional collaborative incremental training and diagnostic architecture performs real-time diagnosis of inter-turn short circuits in drilling motors and training of entropy reduction generation networks. The architecture is divided into two functional areas: edge execution and cloud execution. The two areas rely on a bidirectional data interaction channel to achieve real-time information exchange, and build a closed-loop iterative link of data upload - incremental training - model distribution - real-time diagnosis, overcoming the technical defects of traditional cloud-edge separation architectures, such as model solidification and the need for shutdown and full-domain retraining for model updates.

[0074] Four functional units are deployed in parallel on the left edge: the data acquisition module acquires the original current timing signal of the permanent magnet synchronous motor, the operating condition acquisition module acquires the operating condition parameters such as load and speed simultaneously; the acquired signals and operating condition information are pushed to the cloud through the fault feature upload channel; at the same time, the edge loads the latest model sent from the cloud to continuously complete the real-time diagnosis of inter-turn short circuits of the drilling motor.

[0075] On the right, the cloud relies on a multi-GPU computing cluster to complete various parallel tasks: receiving fault data and operating condition information transmitted from the edge, combining virtual fault samples generated by the constrained modulation U-shaped network to form a hybrid dataset for multi-GPU collaborative training and real-time incremental training of the model; the updated training parameters are sent to the edge terminal to realize iterative optimization of the edge diagnostic model; the cloud can also formulate maintenance strategies based on the diagnostic results.

[0076] Figure 4 The red double-headed arrows represent the continuous bidirectional data flow between the edge device and the cloud. The edge device continuously uploads new fault samples and operating condition information, while the cloud continuously distributes optimized model parameters. The entire iterative process does not require the edge device to be shut down. Relying on the incremental training mechanism, the diagnostic model is driven to dynamically adapt to the changing operating conditions of oil drilling, continuously reducing the probability of misdiagnosis and missed diagnosis of inter-turn short circuit faults in the permanent magnet synchronous motor while drilling.

[0077] Figure 5To constrain the sample fidelity verification results of the modulation sparse generalization generation algorithm, a three-layer subgraph is used to intuitively verify that the generation network can produce high-fidelity virtual fault samples of inter-turn short circuits in motors, solving the feature distortion problem of traditional generation models. Top-layer waveform comparison: The blue solid line represents the original signal of the real motor fault, and the red dashed line represents the virtual fault signal generated by this method. The two waveforms have highly overlapping trends and amplitudes, and the time-domain waveforms have almost no obvious offset, indicating that the generated samples can accurately replicate the temporal fluctuation characteristics of the real fault. Middle-layer error curve: The curve represents the point-by-point error between the original signal and the generated signal. The error magnitude is only one-thousandth, and the overall fluctuation amplitude is extremely small, with no continuous large deviation, proving that the deviation between the generated signal and the real signal is weak. Bottom-layer error distribution histogram: The error values ​​are concentrated around 0, showing an approximately symmetrical distribution. The positive and negative error ranges are narrow, and there are no large-area extreme errors, further quantitatively demonstrating the low degree of distortion of the generated samples.

[0078] The experimental results confirm that the sparse generalization generative network with multi-fault category constraints proposed in this invention can generate high-fidelity virtual samples of inter-turn short circuits by relying on the complete entropy increase-entropy decrease inverse sampling process. This can expand the small sample motor fault dataset, provide high-quality training data for cloud incremental diagnostic models, and support the stable operation of cloud-edge collaborative closed-loop diagnostic methods.

[0079] Figure 6 To compare three types of signal waveforms—original real data, data generated by the constrained modulation sparse generalization generation algorithm proposed in this invention, and data generated by a traditional adversarial generative network—covering four states: healthy, fault level 1, fault level 2, and fault level 3, this study verifies that the proposed method can effectively suppress mode collapse and generate multi-gradient inter-turn short-circuit fault samples in a balanced manner. The table is divided horizontally into healthy and three different levels of inter-turn short-circuit faults; vertically, it is divided into three rows. The first row shows the original timing signals of the motor, with clear distinctions in waveform characteristics for each operating condition, and amplitude and waveform shape showing regular changes with the fault degree. The second row, in red, shows samples generated by the algorithm of this invention. The waveform contours, amplitude ranges, and fluctuation trends for each operating condition highly match the corresponding original signals, with clear distinctions between different fault levels and no category confusion or feature loss. The third row, in red, shows the output results of the traditional adversarial generative network. All operating condition waveforms are superimposed with a large amount of chaotic high-frequency noise, resulting in severe waveform contour distortion, blurred waveform differences between different fault levels, and the typical mode collapse problem, failing to accurately reproduce the differentiated characteristics of multi-level faults.

[0080] Experimental results show that, relying on the multi-fault category constraint modulation entropy reduction inverse sampling mechanism, the network generated by this invention can stably generate multi-gradient fault samples with good discrimination and complete features, solving the pain points of mode collapse and imbalance in the generation of multi-level fault samples in traditional adversarial generative models. It can provide a balanced and complete extended dataset for cloud-edge collaborative incremental diagnosis methods.

[0081] Figure 7 The model is divided into four subgraphs: (a) healthy state, (b) fault level 1, (c) fault level 2, and (d) fault level 3. These subgraphs are used to verify the inter-class generalization generation capability of the proposed constrained modulation sparse generalization generation algorithm for inter-turn short-circuit fault samples of motors. In each subgraph, the blue solid line represents the real original signal, and the orange, purple, and yellow dashed lines represent multiple sets of virtual samples generated by the model.

[0082] Experiments show that, relying on the multi-sparse generalization mechanism, the generative network can stably produce diverse virtual samples of the same class under a single model, taking into account both sample fidelity and intra-class diversity. It has excellent generalization performance and can evenly expand the multi-level inter-turn short-circuit fault dataset, solving the problems of scarce fault samples and imbalance of the number of various types of samples in industrial scenarios, and providing rich and high-quality expansion data for cloud-based incremental diagnostic models.

[0083] Figure 8 Feature scatter plots were used to compare the feature discrimination effect before and after the diagnostic model training. The samples were divided into four categories: healthy, fault level 1, fault level 2, and fault level 3. Figure 8 Subgraph (a) shows the feature distribution when the model is not trained. The scatter points of the four types of samples are mixed together and the category boundaries are blurred. The model cannot effectively distinguish different fault conditions. Figure 8 Subgraph (b) shows the feature output results after incremental training is completed in the cloud and the optimized model is deployed to the edge terminal. The four types of samples each form an independent and compact cluster, with no overlap and clear boundaries between the clusters.

[0084] The results verify that after the balanced fault dataset with network expansion generated by this invention is completed and incremental iteration is performed in the cloud, the diagnostic model sent to the edge has a strong ability to distinguish multi-level inter-turn short circuit features. It can accurately distinguish between healthy states and short circuit faults of different degrees, effectively reduce the probability of misdiagnosis and missed diagnosis of faults in industrial fields, and meet the real-time diagnostic needs of the edge.

[0085] Figure 9 To compare the motor fault classification accuracy of different sample generation methods under varying load and speed conditions, four groups of dynamic speed and load conditions were set up. Five schemes were compared: adversarial generation, W-adversarial generation, gradient spectrum adversarial generation, traditional diffusion model, and the algorithm proposed in this invention. The vertical axis represents the fault classification accuracy, and the table simultaneously provides the mean and ± standard deviation of each group as quantitative indicators.

[0086] The two GAN frameworks, ordinary adversarial generation and W-adversarial generation, achieved an accuracy of only 56%–65% across all scenarios, with a standard deviation generally greater than 2.7%, indicating low precision and large fluctuations. Gradient spectrum adversarial generation showed some improvement, but the average accuracy was only 76%–83%, still exhibiting a significant shortcoming in precision. The root cause is that traditional adversarial networks are prone to pattern collapse, resulting in insufficient fidelity and balance in the generated fault samples, leading to poor recognition performance of the diagnostic model. The standard diffusion model achieved a classification accuracy of 93%–98% across four scenarios, with significantly better stability than the GAN series methods, validating the advantages of the entropy increase / decrease inverse sampling architecture in generating temporal fault samples. Under all four dynamic load conditions, the proposed method maintained an average accuracy of 96.43%–97.86%, with a standard deviation of only 0.14–0.29%, the smallest among all schemes, representing extremely strong model robustness. Under varying industrial operating conditions with wide ranges of speed and load changes, this algorithm, relying on a constrained modulation sparse generalization mechanism, can generate high-fidelity, class-balanced multi-level inter-turn short-circuit samples. The edge diagnostic model trained after expanding the dataset exhibits stronger generalization capabilities and can stably adapt to dynamically changing load industrial scenarios, significantly reducing the probability of false and false fault diagnoses. Under four sets of differentiated speed and load conditions, the proposed algorithm consistently maintains top-tier and stable recognition accuracy, unaffected by fluctuations in operating conditions. Compared to other comparative methods, it possesses stronger industrial field adaptability, aligning with the application requirements of cloud-edge collaborative incremental diagnostic methods for variable motor operating conditions.

[0087] Figure 10 To compare the loss convergence characteristics of the generated network under a hybrid precision training strategy with and without clustered cooperative annealing. Among them, Figure 10 Subgraph (a) shows the loss curve using the complete integrated training strategy of this invention. The enlarged logarithmic graph shows that the loss steadily decreases with each iteration, maintaining a slight decay trend even after 5000 iterations, without any oscillations or rebounds, and finally converges to... It exhibits excellent magnitude, convergence stability, and fitting accuracy. Figure 10 Subgraph (b) shows the control group without this integrated training strategy. After only 543 iterations, the training loss changed and overflowed into a NAN, causing the training to diverge completely and fail to achieve effective fitting. Experiments verify that the integrated training strategy, consisting of clustered distributed synchronous gradients, BF16 mixed precision, cosine annealing learning rate, and exponentially smoothed shadow weights, can stably constrain numerical fluctuations, avoid training numerical overflow and divergence, and ensure stable convergence of the generator network to achieve high-precision fitting results.

[0088] This embodiment also provides an inter-turn short-circuit fault diagnosis system for implementing the above-described sample-expansion-based inter-turn short-circuit fault diagnosis method for permanent magnet synchronous motors in drilling environments, comprising: The data acquisition module is used to collect real current signal samples of the permanent magnet synchronous motor under normal operation and inter-turn short circuit fault conditions in the drilling environment, and to perform preprocessing. The entropy reduction generation network training module adds a normal distribution noise signal for training to the preprocessed real current signal samples, performs entropy increase evolution according to the preset time-series decay coefficient, and trains the pre-constructed entropy reduction generation network based on the entropy increase evolution data. The virtual current signal sample generation module is used to obtain the normally distributed noise signal for entropy reduction, input it into the trained entropy reduction generation network to generate noise prediction values, and use the constrained modulation sparse generalization generation algorithm to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the accumulated time-series decay coefficient during the entropy increase evolution process, and output the restored virtual current signal sample. A complete fault dataset generation module is used to fuse virtual current signal samples with real current signal samples to construct a complete fault dataset. The model update module is used to iteratively update the pre-built fault diagnosis model using a complete fault dataset; The real-time diagnostic module is used to perform real-time diagnostics on the operating data of the permanent magnet synchronous motor using the updated fault diagnosis model to determine whether an inter-turn short circuit fault has occurred.

[0089] Preferably, a high-temperature, high-vibration coupling simulation platform for the aforementioned inter-turn short-circuit fault diagnosis system is also constructed, mainly comprising a measurement and control system, an environmental temperature control module, and a vibration excitation base. The measurement and control system includes an oscilloscope, a control circuit board, a prototype of the drilling permanent magnet synchronous motor under test, and a monitoring computer on the vibration table. An electric vibration table is used as the vibration excitation base. The environmental temperature control module measures the surface temperature of the motor. This platform relies on the vibration table to reproduce the strong downhole vibration load, and, in conjunction with the environmental temperature control module, simulates the high-temperature environment of the motor. It can simultaneously reproduce the high-temperature, high-vibration coupling conditions of drilling, and can collect motor operating signals under different degrees of inter-turn short-circuit damage to obtain the real fault characteristics under coupled conditions, supporting the experimental verification of the inter-turn short-circuit diagnosis algorithm.

[0090] It should be noted that the specific content and beneficial effects of the system in this application can be found in the above method embodiments, and will not be repeated here.

[0091] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors in drilling environments based on sample expansion, characterized in that, Includes the following steps: Real current signal samples of permanent magnet synchronous motors under normal operation and inter-turn short circuit fault conditions in the drilling environment were collected and preprocessed. Based on the preprocessed real current signal samples, a normal distribution noise signal for training is added, and entropy increase evolution is performed according to the preset time-series decay coefficient. The pre-constructed entropy decrease generation network is trained based on the entropy increase evolution data. The normally distributed noise signal used for entropy reduction is obtained and input into the trained entropy reduction generation network to generate noise prediction values. The constrained modulation sparse generalization generation algorithm is used to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the time-series decay coefficient accumulated during the entropy increase evolution process, and outputs the restored virtual current signal sample. The virtual current signal samples are fused with the real current signal samples to construct a complete fault dataset; the complete fault dataset is then used to iteratively update the pre-constructed fault diagnosis model. An updated fault diagnosis model is used to perform real-time diagnosis of the operating data of the permanent magnet synchronous motor to determine whether an inter-turn short circuit fault has occurred.

2. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 1, is characterized in that... The processing steps of the constrained modulation sparse generalization generation algorithm include the following steps: S101: Regarding step size Initialize with the target fault category; S102: Subtract the randomly generated entropy from the normally distributed noise signal as... t The high-entropy perturbation state corresponding to time 1 , among which, time t The initial value is the total number of time periods of the entropy increase evolution; S103: Based on the high-entropy perturbation state Given the corresponding target fault category, constrained noise prediction values ​​are generated through an entropy reduction generation network. According to the high-entropy perturbation state With empty category labels, unconstrained noise predictions are generated through an entropy reduction generation network; The constrained noise prediction value and the unconstrained noise prediction value are weighted and fused to obtain the current... t The final network prediction noise corresponding to the given time; S104: If Then based on Gaussian random noise sampling and cross The mean state derived by the inverse entropy reduction step is determined. High-entropy perturbation state at time 1 Then return to step S103, where the state mean is determined based on the accumulated temporal decay coefficient and the final network prediction noise corresponding to the current moment; like Then the original state signal is determined. and the original state signal Output as a virtual current signal sample.

3. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 2, is characterized in that... The cross The expression for calculating the state mean in the reverse entropy reduction derivation is: In the formula, From t The moment begins to cross The mean state derived from the reverse entropy reduction step. From the initial moment to The cumulative decay coefficient at time step, , for i The temporal decay coefficient at time step, From the initial moment to The cumulative decay coefficient at time step, The final network prediction noise is generated by entropy reduction in the network output. for t The square of the state variance at time t; The High-entropy perturbation state at time 1 The calculation expression is: In the formula, for t The state variance at time t, It is Gaussian random noise.

4. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 2, is characterized in that... The entropy reduction generation network is a symmetric encoding and decoding architecture, including an encoding convolution module, a decoding convolution module, and a bottleneck convolution module. The encoding convolution module and the decoding convolution module are symmetric structures. The encoding convolution module receives the input of a high-entropy perturbation state and performs downsampling step by step through multi-level encoding convolution. Each level of encoding convolution is connected to the corresponding decoding convolution of the decoding convolution module through a residual jump fusion path. The bottom-level encoded features output by the encoding convolution module are input into the bottleneck convolution module to complete the deep feature mapping. Then, the decoding convolution module upsamples the features step by step and fuses the output features of the corresponding level of encoding convolution to finally output the noise prediction value. The entropy reduction generation network further includes a time-series information embedding module and a multi-gradient inter-turn short-circuit category coding module. The time-series information embedding module is used to extract prior knowledge of the operating conditions of the permanent magnet synchronous motor and form prior knowledge constraints. The multi-gradient inter-turn short-circuit category coding module is used to generate fault category modulation constraints based on the input target fault category. The prior knowledge constraints and fault category modulation constraints of the time-series information embedding module and the multi-gradient inter-turn short-circuit category coding module are injected in parallel into all network layers of the coding convolution module, the decoding convolution module, and the bottleneck convolution module.

5. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 2, is characterized in that... The training process of the entropy reduction generation network includes the following steps: Obtain the entropy-incremented evolved data used for training the entropy-reduction generation network, and input it into the entropy-reduction generation network; The entropy reduction generation network is trained collaboratively through multiple distributed clusters, which employ a parallel architecture with multiple GPU nodes to split the training tasks. In each iteration, the final network prediction noise output by the entropy reduction generation network is compared with the normal distribution noise signal used for training to calculate the loss function; based on the calculated value of the loss function, gradient feedback is obtained to update the network parameters of the entropy reduction generation network. Construct a shadow generation network for the entropy reduction generation network, and update the parameters based on the network parameters of the entropy reduction generation network in each iteration. Use the network parameters of the shadow generation network obtained after the entropy reduction generation network completes its iteration as the network parameters trained by the entropy reduction generation network.

6. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 5, is characterized in that... The shadow generation network updates its parameters using an exponentially weighted average method. The parameter update expression for the shadow generation network is as follows: In the formula, Here are the network parameters of the smoothed shadow generation network in the k-th iteration. It is an exponential decay factor. These are the network parameters of the entropy-reduced generation network after the k-th iteration; The multi-distributed cluster calculates the global average gradient by aggregating the local gradients of each GPU node, and uses this as the gradient feedback for the entropy reduction generation network.

7. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 5, is characterized in that... The convolutional layer in the entropy reduction generation network is a one-dimensional spatial channel separation convolution. This one-dimensional spatial channel separation convolution first fuses the spatial information of each channel, and then fuses the information output by each channel.

8. The method for diagnosing inter-turn short-circuit faults of permanent magnet synchronous motors in drilling environments based on sample expansion, as described in claim 5, is characterized in that... The entropy reduction generation network employs a cosine annealing modulation strategy to smoothly decay the learning rate. The expression for the smooth decay of the learning rate is as follows: In the formula, Let be the current learning rate in the k-th iteration. The initial learning rate, The minimum learning rate is set. It represents the total number of training rounds.

9. A system for diagnosing inter-turn short-circuit faults in a permanent magnet synchronous motor in a drilling environment, as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect real current signal samples of the permanent magnet synchronous motor under normal operation and inter-turn short circuit fault conditions in the drilling environment, and to perform preprocessing. The entropy reduction generation network training module adds a normal distribution noise signal for training to the preprocessed real current signal samples, performs entropy increase evolution according to the preset time-series decay coefficient, and trains the pre-constructed entropy reduction generation network based on the entropy increase evolution data. The virtual current signal sample generation module is used to acquire the normally distributed noise signal for entropy reduction, input it into the trained entropy reduction generation network to generate noise prediction values, and use the constrained modulation sparse generalization generation algorithm to perform multi-step inverse entropy reduction deduction based on the noise prediction values ​​and the time-series decay coefficient accumulated during the entropy increase evolution process, and output the restored virtual current signal sample. A complete fault dataset generation module is used to fuse the virtual current signal samples with the real current signal samples to construct a complete fault dataset. The model update module is used to iteratively update the pre-built fault diagnosis model using the complete fault dataset; The real-time diagnostic module is used to perform real-time diagnostics on the operating data of the permanent magnet synchronous motor using the updated fault diagnosis model to determine whether an inter-turn short circuit fault has occurred.

10. The system according to claim 9, characterized in that, The system is based on an edge and cloud processing architecture. The data acquisition module and real-time diagnostic module run at the edge and upload the acquired real current signal samples to the cloud. The complete fault dataset generation module and model update module run in the cloud and send the updated fault diagnosis model parameters to the fault diagnosis model at the edge.

Images