A method and system for evaluating the health status of power equipment based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based method and system for assessing the health status of power equipment. The method includes acquiring and labeling power equipment health status data, and assessing the health status of power equipment. The present invention employs a learnable spectral modulation vector to improve robustness to complex environments. Through Lie group symmetry operations, the model maintains consistent feature extraction under different installation angles. Small random perturbations are introduced to enhance the ability to explore abnormal samples and noisy data. Real-time monitoring of training status improves training efficiency and avoids overfitting specific fault modes. The model's ability to identify faults under geometric distortion and complex operating conditions is enhanced. By selectively pruning redundant channels, the model structure is made more compact, computational costs are reduced, and high classification accuracy is maintained.

Description

Technical Field

[0001] The present invention relates to the field of power equipment health monitoring, and in particular to an artificial intelligence-based power equipment health status assessment method and system. Background Art

[0002] As power systems continue to expand and become more intelligent, monitoring the health of power equipment has become increasingly important. During long-term operation, equipment can be affected by mechanical wear, environmental factors, and varying operating loads, leading to faults such as bearing wear, rotor imbalance, looseness, and overheating. Failure to detect and address these faults in a timely manner can lead to decreased equipment performance, increased energy consumption, and even serious safety incidents. Therefore, accurate and efficient power equipment health assessment technology is crucial to ensuring grid stability and the normal operation of industrial systems.

[0003] Among the required technologies, the technical solutions of patents with publication numbers CN119295442B, CN118962253B, and CN115428231B have the following problems: using fixed filters (such as STFT or Gabor filters), they cannot adapt to spectrum changes under different working conditions; they mainly rely on data enhancement (such as rotation, flipping, etc.) to improve the robustness of the model to geometric transformations, and cannot fundamentally solve the signal distortion caused by changes in sensor installation position or angle; traditional gradient descent methods are prone to falling into local optimality, resulting in overfitting or training stagnation; fixed learning rates or preset strategies are often used to adjust the learning rate, which makes it difficult to adapt to changes in data distribution; traditional convolutional neural networks mostly use a single loss (such as cross entropy), and have limited generalization capabilities when dealing with complex failure modes of power equipment; traditional pruning methods are mainly based on weight size or sensitivity, which may lose key information. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In order to solve the above technical problems, the present invention provides an artificial intelligence-based power equipment health status assessment method and system.

[0006] (2) Technical solution

[0007] In order to solve the above-mentioned technical problems and achieve the purpose of the invention, the present invention is implemented through the following technical solutions:

[0008] An artificial intelligence-based power equipment health status assessment method includes the following steps:

[0009] S1: Acquisition of power equipment health status data, including collecting vibration monitoring signals of power equipment during actual operation;

[0010] S2: Label the acquired vibration signal data based on the equipment's operating status and fault type;

[0011] S3: Assessment of the health status of power equipment, including the health status of power equipment based on convolutional neural networks and training of the network; the training process includes initializing the weights of the neural network based on the He normal initialization function; learning and modulating the spectral components of the features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; applying perturbations to the gradient field; updating the convolution kernel weights based on dynamic adjustment factors; pruning based on weight importance; dynamically adjusting the learning rate based on real-time performance during training; and calculating the loss function by integrating classification error, feature comparison consistency, and dynamic sparse regularization.

[0012] Furthermore, the spectrum components based on the adaptive spectrum transformation learning and modulation features adjust the amplitude of each frequency component through element-by-element multiplication.

[0013] Furthermore, the processing of intermediate features based on Lie group symmetry operations includes performing specific geometric transformations on input features through Lie group symmetry operation functions, and correcting features by applying an operation representing the geometric transformation and its corresponding inverse operation before and after the feature input.

[0014] Furthermore, the convolution kernel weight update based on the dynamic adjustment factor dynamically focuses on the area with larger loss by backpropagating the perturbed gradient.

[0015] Furthermore, the dynamic adjustment factor is calculated as follows:

[0016]

[0017] Where, is the adjustment coefficient of the dynamic adjustment factor; The loss function of the convolutional neural network is The gradient of the convolution kernel weight of the layer, is the loss sensitivity, is the loss function value.

[0018] Furthermore, the weight importance-based pruning calculation method calculates the sum of the absolute values of all convolution kernel weights of the layer and compares it with a preset threshold, and uses a binary variable to represent retention or pruning.

[0019] Furthermore, the loss function of the convolutional neural network is expressed as follows:

[0020]

[0021] Where, is the classification loss, To compare the consistency loss, is the dynamic regularization loss; is the classification loss weight; is the comparison consistency loss weight; is the dynamic regularization loss weight.

[0022] Furthermore, the classification loss is obtained by calculating weighted cross entropy; the contrast consistency loss uses Lie group transformation to generate a geometrically enhanced view of the same signal; and the dynamic regularization loss combines pruning strategy with gradient perturbation to impose sparsity constraints and parameter smoothing constraints.

[0023] The present invention also provides an artificial intelligence-based power equipment health status assessment system, which includes:

[0024] The power equipment health status data acquisition module is used to obtain the vibration monitoring signal of the power equipment during actual operation;

[0025] Power equipment data annotation module, which is used to annotate data based on the equipment's operating status and fault type;

[0026] The power equipment health status assessment module is used to construct and train a convolutional neural network. The training process includes initializing the weights of the neural network based on the He normal initialization function; learning and modulating the spectral components of the features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; applying perturbations to the gradient field; updating the convolution kernel weights based on a dynamic adjustment factor; pruning based on weight importance; dynamically adjusting the learning rate based on real-time performance during training; and calculating the loss function by integrating classification error, feature comparison consistency, and dynamic sparse regularization.

[0027] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which program instructions of an artificial intelligence-based power equipment health status assessment method are stored. The program instructions of the artificial intelligence-based power equipment health status assessment method can be executed by one or more processors to implement the steps of the artificial intelligence-based power equipment health status assessment method as described above.

[0028] (3) Beneficial effects

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The present invention adopts a learnable spectrum modulation vector to enhance key frequency bands and suppress noise through frequency domain modulation, thereby improving robustness to complex environments.

[0031] (2) The present invention uses Lie group symmetry operations to adaptively correct signal morphology changes during network forward propagation, so that the model can maintain consistency in feature extraction at different installation angles.

[0032] (3) The present invention introduces small random perturbations when calculating the gradient, which enhances the ability to explore abnormal samples and noisy data and makes the model more generalizable.

[0033] (4) The present invention monitors the training status in real time and adaptively adjusts the learning rate to improve the training efficiency and avoid overfitting of specific fault modes.

[0034] (5) The present invention integrates classification error, contrast consistency loss and dynamic sparse regularization to improve the model's fault recognition ability under geometric distortion and complex working conditions.

[0035] (6) The present invention calculates the global importance of weights and selectively prunes redundant channels, making the model structure more compact and reducing computational cost while maintaining high classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 This is a flow chart of a method for evaluating the health status of power equipment based on artificial intelligence according to an embodiment of the present application;

[0038] Figure 2 The noise robustness analysis results of adaptive spectrum transformation according to the method of the embodiment of the present application are compared with those of the prior art;

[0039] Figure 3 This is a comparison of the geometric perturbation adaptability of the method according to the embodiment of the present application and the Lie group symmetry operation layer of the prior art. DETAILED DESCRIPTION

[0040] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0041] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0042] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0043] See also Figure 1 ,A method for assessing the health status of power equipment based on artificial intelligence includes the following steps:

[0044] S1: Acquisition of health status data of power equipment;

[0045] Vibration monitoring signals are collected during the actual operation of power equipment, including but not limited to transformers, generators, motors and other important power facilities. During the daily operation and maintenance of the equipment and specialized inspections, vibration signals are obtained through high-precision vibration sensors installed at specific locations on the equipment. These sensors can sense tiny vibration changes in the equipment in real time during operation and output raw vibration data in the form of electrical signals.

[0046] To comprehensively capture the vibration characteristics of the equipment under different operating conditions and potential fault states, sensors are placed at multiple key locations on the equipment, including near the bearing seat of the motor and the stator and rotor of the generator, to obtain vibration information from different angles. At the same time, the acquisition process covers the normal operating state of the equipment as well as various artificially simulated or actual fault states, such as vibration signals under different fault modes such as wear, imbalance, looseness, and overheating of equipment components, to enrich the diversity of the data and ensure that the model can learn comprehensive feature representations.

[0047] S2: Label the acquired vibration signal data to realize the supervised training of the health status assessment model of the power equipment. The labeling work is carried out based on the operating status and fault type of the equipment. When the acquisition equipment is in normal operation, the acquired vibration signal is labeled as the "normal" category; and when the equipment has a specific failure mode, such as the aforementioned component wear, imbalance and other faults, the corresponding vibration signal is accurately labeled as the corresponding "fault" category, such as "bearing wear fault" and "rotor imbalance fault".

[0048] S3: Power Equipment Health Assessment

[0049] A convolutional neural network is used to determine the health status of power equipment. In one embodiment, the structure of the convolutional neural network is shown in the following table:

[0050] Layer Type Parameter settings Output shape Input layer Signal length T=1024 (1024,1) Adaptive convolution block 1 Conv1D(filters=64,kernel_size=5) (1024,64) Max Pooling 1 Pool_size=2 (512,64) Adaptive convolution block 2 Conv1D(filters=128,kernel_size=3) (512,128) Max Pooling 2 Pool_size=2 (256,128) Adaptive convolution block 3 Conv1D(filters=256,kernel_size=3) (256,256) Global Average Pooling (GAP) - (256,) Fully connected projection head FC(128) (128,) Classification output layer FC(C)+Softmax (C,)

[0051] The convolutional neural network training steps are as follows:

[0052] S31. Normalize the vibration monitoring signals of power equipment to eliminate the large amplitude differences in the original signals and improve the input data quality and stability of the convolutional neural network.

[0053] S32, the initialization stage sets the convolutional neural network structure and initializes the convolution kernel weights to avoid the gradient instability that may occur in the early iterations of the convolutional neural network, which is expressed as:

[0054]

[0055] Where, is the initial convolution kernel weight; is the convolution kernel weight scaling factor, which represents the coefficient for adjusting the initialization amplitude. Preferably, Set to 0.8; is the He normal initialization function, For convolutional neural network Number of neurons in a layer; is a positive integer representing the number of layers of the convolutional neural network.

[0056] The He normal initialization function is a weight initialization method for a neural network. Its main purpose is to prevent the signal from attenuating or amplifying during forward or backward propagation in a deep network. It automatically adjusts the distribution of the initial convolution kernel weights according to the number of neurons in the current layer, so that the randomly generated weights have an appropriate amplitude. Since the collected data often has noise and amplitude differences, this initialization method can make the network more stable when processing signal features, thereby helping to more accurately capture subtle changes in vibration patterns in the early stages of convolutional neural network training.

[0057] S33, learning and modulating the spectral components of the features based on adaptive spectral transformation;

[0058] The acquired health status data of the power equipment is input into the convolutional neural network. The vibration signal of the power equipment contains key fault characteristics in both the time domain and the frequency domain. Traditional convolution operations rely on fixed frequency domain filters (such as Gabor filters) or short-time Fourier transforms, and the frequency band selection is fixed. Fixed filters cannot adapt to the noise frequency band offset or resonance frequency changes caused by changes in equipment operating conditions, and the frequency domain feature extraction has poor robustness.

[0059] The present invention adopts adaptive spectrum transformation in the forward propagation process to automatically learn and modulate the spectral components of the features, enhance the model's ability to capture frequency domain information, and improve its robustness to complex noise environments. The calculation method is expressed as:

[0060]

[0061] Where, For convolutional neural network The output of the layer is converted into a feature representation in the frequency domain; Fast Fourier transform is used to convert vibration signals from time domain to frequency domain; Characterization will Converting to the frequency domain yields , For convolutional neural network The output of the layer; is the frequency domain feature after modulation; For the learnable convolutional neural network The spectrum modulation vector of the layer adjusts the amplitude of each frequency component through element-by-element multiplication. For example, for a bearing fault in a power equipment with a resonant frequency of 2kHz, the spectrum modulation vector learning enhances the frequency band near 2kHz (modulation coefficient close to 1) and suppresses high-frequency noise (>5kHz, modulation coefficient close to 0); is the element-wise product; is the inverse Fourier transform; It is the time domain feature restored by inverse Fourier transform.

[0062] The spectrum modulation vector adaptively enhances key frequency bands (such as the resonant frequency of power equipment) and suppresses noise frequency bands, solving the performance degradation problem of traditional fixed filters caused by changes in equipment operating conditions. Its learning method is as follows:

[0063]

[0064] Where, is the Sigmoid activation function, which constrains the modulation amplitude to interval; It is a global average pooling operation that extracts global statistics of frequency domain features to capture the average energy distribution of each frequency component in the batch samples; is the weight parameter of the preset fully connected layer that can be learned, is the frequency domain feature dimension, represents the parameter dimension space, The bias parameters of the fully connected layer can be learned by default.

[0065] The preset fully connected layer is a fully connected neural network layer, which functions to establish a nonlinear mapping relationship between frequency energy and modulation coefficient.

[0066] Furthermore, the weight parameters and bias parameters of the preset fully connected layer are dynamically adjusted through back propagation.

[0067] S34, processing intermediate features based on Lie group symmetry operations;

[0068] The vibration signals of power equipment may be distorted when collected at different installation positions or angles. During the forward propagation of the convolutional neural network, the present invention not only performs conventional convolution, pooling, and nonlinear activation operations, but also processes intermediate features through Lie group symmetry operations, thereby maintaining high invariance under geometric transformations and improving the model's robustness to distortion changes (such as rotation or reflection). It can be expressed as:

[0069]

[0070] Where, For convolutional neural network Layer output, For convolutional neural network The convolution kernel weights of the layer, is the Lie group symmetry operation function, For convolutional neural network The output of the layer, For convolutional neural network The bias of the layer, is the Sigmoid activation function, is the convolution operation.

[0071] Furthermore, in conventional convolutional neural networks, data enhancement and other methods are usually used to improve the robustness of the model to geometric transformations, such as rotation, flipping and other operations to expand training data. However, these methods cannot fundamentally guarantee that the model can maintain consistent feature extraction when facing signals collected at different installation positions or angles. There is a lack of an effective processing mechanism for geometric transformations during the forward propagation of the network. The present invention performs a specific geometric transformation on the input features through a Lie group symmetry operation function. The Lie group symmetry operation function is intended to maintain the invariance of the intermediate features after geometric transformations (such as rotation or reflection). By applying an operation representing the geometric transformation and its corresponding inverse operation before and after the feature input, the feature correction is achieved to compensate for the signal distortion caused by different installation angles or positions of the power equipment. When facing data after changes such as rotation and reflection, the model can still extract consistent features, thereby improving the overall robustness and recognition accuracy. The calculation method is expressed as:

[0072]

[0073] Where, This is a Lie group symmetry operation function. When the sensor installation angle deviation causes the signal waveform phase shift (for example, when the sensor is rotated 5°, the time domain waveform is delayed by 2ms), the Lie group operation corrects the characteristic phase to maintain the time alignment of the fault characteristics (such as the impact pulse). is a Lie group element, representing a geometric transformation such as rotation or reflection. Preferably, The purpose of using Lie group elements to represent geometric transformations is to accurately simulate the morphological differences of power equipment vibration signals caused by slight changes in installation angles during acquisition. By using group elements representing small-angle rotations, the network can fine-tune the signal without destroying the original feature structure, thereby stably extracting key features in multi-angle acquisition scenarios and ensuring the model's adaptability and robustness to signal distortion. for The inverse operation of .

[0074] S35, applying a disturbance to the gradient field;

[0075] The vibration signal of the power equipment may be mixed with random high-amplitude noise or abnormal pulses. When updating the network weights, the traditional gradient descent method directly updates according to the calculated gradient without considering the addition of disturbance terms. When processing random high-amplitude noise or abnormal pulses mixed with the vibration signal of the power equipment, it is easy for the model to fall into the local optimal solution, resulting in stagnation or overfitting of the training process, and reducing the generalization performance of the model. The present invention applies perturbation to the gradient field before backpropagation. Specifically, after calculating the gradient of the loss function with respect to the convolution kernel weight, the gradient is corrected by adding a random noise term related to the disturbance intensity. That is, on the basis of the original gradient, a small-amplitude random perturbation is added to expand the parameter search space during training, help the model escape the local optimal solution, and enhance the ability to explore abnormal samples and noisy data, thereby improving the generalization performance of the entire network. The calculation method is expressed as:

[0076]

[0077] Where, The perturbed gradient retains the key information of the original gradient in indicating the direction of loss decrease. At the same time, by using random noise, it prevents the model from being overly dependent on the local optimal solution. This makes the network have a certain degree of jump when updating the weights, thereby avoiding stagnation or overfitting of the training process due to noise or abnormal data, and improving the recognition and robustness of complex failure modes. is the loss function for the convolutional neural network The gradient of the convolution kernel weight of the layer, is the loss function of the convolutional neural network; is the disturbance intensity coefficient, preferably, Set to 0.1; is random noise.

[0078] Furthermore, in order to combine the gradient direction information with the perturbation and control the noise range, the sign of the gradient is calculated, that is, whether the update direction of each weight is positive or negative. This sign information is then multiplied by a uniformly distributed random number within a preset range (for example, from -0.1 to 0.1). This ensures that the random noise not only has a corresponding effect along the positive and negative directions of the gradient, but also its amplitude is limited to the set range. This not only retains the main information of the gradient, but also avoids the training instability problem caused by excessive perturbations, ensuring the smoothness and efficiency of the overall training process. The calculation method of random noise is expressed as:

[0079]

[0080] Where, is a symbolic function, The role of is to determine the positive or negative gradient; for Random noise uniformly distributed within the interval. Preferably, The range is set to .

[0081] S36, updating the convolution kernel weight based on the dynamic adjustment factor;

[0082] When the fault modes in the vibration signals of power equipment are complex and changeable, it is difficult to converge quickly. By backpropagating the perturbed gradient and dynamically focusing on the areas with larger losses, the discrimination between different fault types can be improved and the convergence speed can be accelerated. It can be expressed as:

[0083]

[0084] Where, The updated convolutional neural network The convolution kernel weight of the layer is used as the convolutional neural network of the next iteration. The convolution kernel weights of the layer; is the learning rate of the convolutional neural network, is a dynamic adjustment factor. Preferably, Set to 0.001.

[0085] Furthermore, since different fault modes may cause abnormally high loss values in some areas during training, conventional convolutional neural networks usually use a fixed learning rate or some simple learning rate adjustment strategies, such as step-by-step attenuation, to control the amplitude of weight updates. However, for the complex and changeable fault modes in the vibration signals of power equipment, these methods cannot dynamically focus on areas with larger losses, resulting in a slow convergence speed of the model in key areas, making it difficult to quickly capture key features and improve model performance. In order to make the convolutional neural network pay more attention to these high-loss areas, accelerate convergence and improve classification and differentiation capabilities, the present invention uses a dynamic adjustment factor to control the loss sensitivity and compensate for the additional attention to the areas with larger losses. Based on the sensitivity of the current layer weight to the loss, when the loss is high, the gradient update amplitude of the area will increase accordingly, thereby prompting the network to make more significant parameter adjustments in these key areas. When faced with complex and changeable fault modes in the health status assessment of power equipment, key features can be captured faster and model performance can be improved. The calculation method is expressed as:

[0086]

[0087] Where, is the adjustment coefficient of the dynamic adjustment factor, preferably, Set to 0.1; The loss function of the convolutional neural network is The gradient of the convolution kernel weight of the layer, is the loss sensitivity, is the loss function value, preferably, Set to 0.01.

[0088] S37, pruning based on weight importance;

[0089] Common neural network pruning methods include weight-based pruning and sensitivity-based pruning, which prune unimportant weights or channels by setting certain thresholds or rules. However, when applied to convolutional neural networks, these methods may not fully consider the characteristics of power equipment vibration signals and the complexity of the network structure, resulting in the pruned network structure being not concise and efficient enough, or having limited effect in improving generalization ability. In response to the large number of redundant or invalid feature channels that may appear in convolutional neural networks, the present invention prunes after evaluating the importance of weights, avoiding the computational burden and overfitting problems caused by excessive model size, simplifying the convolutional neural network structure, and improving generalization ability. The calculation method is expressed as:

[0090]

[0091] Where, It is a binary variable, indicating whether the current layer is pruned; For the The sum of the absolute values of the layer weights; is the number of convolutional neural network layer convolution kernel weights, For convolutional neural network The convolution kernel layer weights; is the pruning threshold, preferably, Set to 0.5.

[0092] Furthermore, by calculating the sum of the absolute values of all convolution kernel weights in this layer and comparing it with a preset threshold, when the sum is lower than the threshold, the binary variable takes "1" to indicate that the channel is not important enough and can be pruned; otherwise, it takes "0" to indicate that it is retained. The binary variable acts as a switch in the entire pruning process, determining which channels are pruned, thereby making the network structure more concise and efficient, while improving the generalization ability.

[0093] S38, dynamically adjust the learning rate based on real-time performance during training;

[0094] When the number of power equipment vibration signal samples is limited and the distribution is variable, training is prone to stagnation. Conventional learning rate adjustment methods are mostly based on pre-set strategies, such as fixed interval decay and manual adjustment based on validation set performance. These methods can adapt to changes in the training process to a certain extent. However, when the number of power equipment vibration signal samples is limited and the distribution is variable, these methods may not be able to dynamically adjust the learning rate according to the training status in a timely and accurate manner, which can easily cause the model to fall into training stagnation or overfit to specific fault modes, affecting the flexibility and adaptability of the model. The present invention dynamically adjusts the learning rate based on the real-time performance during training to avoid overfitting of the model to specific fault modes and maintain the flexibility and adaptability of the model. The calculation method is expressed as:

[0095]

[0096] Where, is the learning rate of the updated convolutional neural network, which is used as the learning rate of the next iteration of the convolutional neural network; is the decay rate, preferably, Set to 0.01; is the current training round number.

[0097] S39, fusion classification error, feature comparison consistency and dynamic sparse regularization to calculate the loss function;

[0098] Conventional convolutional neural network loss functions usually only focus on a single objective, such as classification loss, or simply perform a weighted summation of several loss functions. However, the weight setting may not be reasonable enough. When dealing with problems such as geometric distortion of power equipment vibration signals and complex fault modes, it is impossible to fully take into account the generalization ability and fault discrimination of the model under complex working conditions, resulting in limited performance of the model in practical applications. The present invention calculates the loss function of the convolutional neural network by integrating classification error, feature comparison consistency and dynamic sparse regularization, thereby improving the generalization ability and fault discrimination of the model under complex working conditions. The specific formula is as follows:

[0099]

[0100] Where, is the classification loss, To compare the consistency loss, is the dynamic regularization loss; is the classification loss weight, preferably, Set to 0.6; To compare the consistency loss weights, preferably, Set to 0.2; is the dynamic regularization loss weight, preferably, Set to 0.2.

[0101] The classification loss is calculated using weighted cross entropy to address the problem of uneven sample distribution and highlight the importance of rare failure modes, expressed as:

[0102]

[0103] Where, is the batch sample number, representing the number of power equipment vibration signal data input to the convolutional neural network in the current batch; is a positive integer, is the number of fault categories, is a positive integer; For the The weight coefficient of each category is calculated as follows: , For the The frequency of occurrence of each category in the training set. The low-frequency fault category has a larger weight coefficient, which strengthens the model's attention to it. For the The true label of each sample (one-hot encoding format), is the first output of Softmax Class probability. The convolutional neural network is The first sample The output of the layer, is the total number of layers in the convolutional neural network.

[0104] The contrast consistency loss uses Lie group transformation to generate a geometrically enhanced view of the same signal. By minimizing the difference between the original features and the enhanced features, it constrains the similarity in the feature space, forces the model to ignore irrelevant geometric transformations, and improves robustness to changes in installation angles. The calculation method is expressed as:

[0105]

[0106] Where, The feature projection head (fully connected layer) maps the features to a low-dimensional space; is the Lie group symmetry operation function, is the cosine similarity function, The convolutional neural network is The output of the Lth layer of samples.

[0107] The dynamic regularization loss combines pruning strategies with gradient perturbations, imposes sparse constraints (based on the L1 norm) and parameter smoothness constraints (based on the L2 norm), and is calculated as follows:

[0108]

[0109] Where, Is a binary variable that represents whether the current layer is pruned. When pruning is activated ( ), impose L1 sparsity penalty; is the L1 norm, is the L2 norm; is the adjustment coefficient of the dynamic adjustment factor, preferably, Set to 0.1.

[0110] S310, repeat the above steps until a preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0111] When using a trained convolutional neural network to assess and classify the health status of power equipment, the vibration monitoring signals of on-site power equipment must first be normalized and preprocessed in the same way as the training data to ensure that the format and quality of the input data are consistent with those during training and to eliminate any amplitude differences in the original signal.

[0112] Furthermore, the preprocessed vibration signal is input into the trained convolutional neural network. The network automatically performs the forward propagation process, sequentially passing through various convolutional layers, pooling layers, and fully connected layers to extract features and transform the input signal.

[0113] Furthermore, the classification output layer of the network maps the extracted features to various categories of the health status of the power equipment through the Softmax activation function, and outputs the probability value of each health status category. According to the size of the output probability value, the most likely current health status of the power equipment can be determined (the corresponding category with the largest probability value is taken), thereby achieving accurate assessment and classification of its health status.

[0114] In order to verify the advantages of the present invention, the following experiments were performed:

[0115] To verify the effectiveness of the adaptive spectrum conversion module in complex noise environments, this experiment simulates equipment fault identification scenarios under noise interference of different intensities by comparing it with the traditional fixed filter solution and the benchmark model without the introduction of spectrum modulation mechanism. The experimental results are as follows: Figure 2 As shown in the figure, it can be seen that the technical solution using the dynamic spectrum modulation mechanism can still maintain stable recognition performance in a strong noise environment, and its accuracy is significantly better than the traditional method, indicating that the learnable spectrum modulation vector can automatically enhance the key fault frequency band according to the frequency domain characteristics of the input signal, while effectively suppressing the high-frequency noise component, overcoming the defect of the fixed filter's rigid frequency response characteristics when the working conditions change, and significantly improving the model's robustness to noise interference.

[0116] In terms of the geometric perturbation adaptability of the Lie group symmetry operation layer, a signal distortion test set with different rotation angles is constructed to compare and analyze the performance differences between the benchmark model that does not use Lie group operations and this technology. Figure 3 As shown in the figure, the experimental results show that the present invention can still maintain a high classification accuracy in the presence of significant geometric disturbances, and its performance degradation is much smaller than that of traditional methods. This shows that Lie group symmetry operations, by establishing an algebraic representation of geometric transformations, enable the network to automatically correct signal morphology changes caused by sensor installation angle deviations, effectively improving the invariance of feature representation to geometric disturbances, and solving the problem of feature distortion caused by traditional methods ignoring signal spatial relationships.

[0117] In this embodiment, frequency domain modulation is used to enhance key frequency bands and suppress noise, thereby improving robustness to complex environments. Lie group symmetry operations are used to adaptively correct signal morphology changes during network forward propagation, so that the model can maintain consistency in feature extraction under different installation angles. Small random perturbations are introduced when calculating gradients to enhance the ability to explore abnormal samples and noise data, making the model more generalizable. Real-time monitoring of the training status and adaptive adjustment of the learning rate improve training efficiency and avoid overfitting of specific fault modes. The model's fault recognition ability under geometric distortion and complex working conditions is improved by fusing classification errors, contrast consistency loss, and dynamic sparse regularization. The global importance of weights is calculated and redundant channels are selectively pruned to make the model structure more compact and the computational cost lower while maintaining high classification accuracy.

[0118] The embodiment of the present invention further provides an artificial intelligence-based power equipment health status assessment system, comprising:

[0119] The power equipment health status data acquisition module is used to obtain the vibration monitoring signal of the power equipment during actual operation;

[0120] Power equipment data annotation module, which is used to annotate data based on the equipment's operating status and fault type;

[0121] The power equipment health status assessment module is used to construct and train a convolutional neural network. The training process includes initializing the weights of the neural network based on the He normal initialization function; learning and modulating the spectral components of the features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; applying perturbations to the gradient field; updating the convolution kernel weights based on a dynamic adjustment factor; pruning based on weight importance; dynamically adjusting the learning rate based on real-time performance during training; and calculating the loss function by integrating classification error, feature comparison consistency, and dynamic sparse regularization.

[0122] In addition, an embodiment of the present invention also proposes a computer-readable storage medium, on which program instructions of an artificial intelligence-based power equipment health status assessment method are stored. The artificial intelligence-based power equipment health status assessment program instructions can be executed by one or more processors to implement the steps of the artificial intelligence-based power equipment health status assessment method as described above.

[0123] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for evaluating the health status of power equipment based on artificial intelligence, characterized in that: The steps include: S1: Acquisition of power equipment health status data, including collecting vibration monitoring signals of power equipment during actual operation; S2: Label the acquired vibration signal data based on the equipment's operating status and fault type; S3: Power equipment health status assessment, including power equipment health status based on convolutional neural networks and network training; The training process includes initializing the weights of the neural network based on the He normal initialization function; learning and modulating the spectral components of the features based on adaptive spectral transformation; and processing intermediate features based on Lie group symmetry operations. Apply perturbations to the gradient field; update the convolution kernel weights based on dynamic adjustment factors; Pruning based on weight importance; Dynamically adjust the learning rate based on real-time performance during training; calculate the loss function by integrating classification error, feature comparison consistency, and dynamic sparse regularization; The convolution kernel weight update based on the dynamic adjustment factor dynamically focuses on the area with larger loss by backpropagating the perturbed gradient; The dynamic regulatory factor The calculation method is expressed as: ； Where, is the adjustment coefficient of the dynamic adjustment factor; The loss function of the convolutional neural network is The gradient of the convolution kernel weight of the layer, is the loss sensitivity, is the loss function value.

2. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, characterized in that: The spectrum components based on the adaptive spectrum transformation learning and modulation features adjust the amplitude of each frequency component through element-by-element multiplication.

3. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, characterized in that: The processing of intermediate features based on Lie group symmetry operations includes performing specific geometric transformations on input features through Lie group symmetry operation functions, and correcting features by applying an operation representing the geometric transformation and its corresponding inverse operation before and after feature input.

4. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, characterized in that: The weight importance-based pruning calculation method calculates the sum of the absolute values of all convolution kernel weights in the current layer and compares it with a preset threshold, and uses a binary variable to indicate whether to retain or prune.

5. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, characterized in that: The loss function L of the convolutional neural network is expressed as follows: ； Where, is the classification loss, To compare the consistency loss, is the dynamic regularization loss; is the classification loss weight; is the comparison consistency loss weight; is the dynamic regularization loss weight.

6. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 5, characterized in that: The classification loss is calculated using weighted cross entropy; the contrast consistency loss uses Lie group transformation to generate geometrically enhanced views of the same signal; the dynamic regularization loss combines pruning strategies with gradient perturbations to impose sparsity constraints and parameter smoothing constraints.

7. A system based on the artificial intelligence-based power equipment health status assessment method according to any one of claims 1 to 6, comprising: The power equipment health status data acquisition module is used to obtain the vibration monitoring signal of the power equipment during actual operation; Power equipment data annotation module, which is used to annotate data based on the equipment's operating status and fault type; Power equipment health status assessment module, which is used to build and train convolutional neural networks; The training process includes initializing the weights of the neural network based on the He normal initialization function; learning and modulating the spectral components of the features based on adaptive spectral transformation; and processing intermediate features based on Lie group symmetry operations. Applying perturbations to the gradient field; Update convolution kernel weights based on dynamic adjustment factors; Pruning based on weight importance; Dynamically adjust the learning rate based on real-time performance during training; The loss function is calculated by integrating classification error, feature comparison consistency and dynamic sparsity regularization.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions for an artificial intelligence-based method for assessing the health status of electric power equipment. The program instructions for an artificial intelligence-based method for assessing the health status of electric power equipment can be executed by one or more processors to implement the steps of the artificial intelligence-based method for assessing the health status of electric power equipment as described in one of claims 1-6.

Images