Power equipment health state assessment method and system based on artificial intelligence

Abstract

The invention discloses a power equipment health state assessment method and system based on artificial intelligence. The method comprises the steps of power equipment health state data acquisition, marking and power equipment health state assessment. According to the method, a learnable spectrum modulation vector is adopted, so that the robustness to a complex environment is improved; through Lie group symmetry operation, the model can still keep the consistency of feature extraction at different installation angles; small-amplitude random disturbance is introduced, so that the exploration capability on abnormal samples and noise data is enhanced; by monitoring the training state in real time, the training efficiency is improved, and overfitting of a specific fault mode is avoided; the fault identification capability of the model under geometric distortion and complex working conditions is improved; by selectively pruning redundant channels, the model structure is more compact, the calculation cost is lower, and meanwhile, the high classification precision is kept.

Description

Technical Field

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

[0002] With the continuous expansion and intelligent development of the power system, the health status monitoring of power equipment has become particularly important. During the long-term operation of the equipment, affected by mechanical wear, environmental factors, and changes in operating load, faults such as bearing wear, rotor imbalance, looseness, and overheating may occur. If not detected and processed in time, it may lead to a decline in equipment performance, an increase in energy consumption, and even serious safety accidents. Therefore, accurate and efficient power equipment health assessment technology is crucial for ensuring the stability of the power grid and the normal operation of industrial systems.

[0003] In the prior art, 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 the spectral changes under different working conditions; mainly relying on data augmentation (such as rotation, flipping, etc.) to improve the robustness of the model to geometric transformations, they cannot fundamentally solve the signal distortion caused by changes in the installation position or angle of sensors; traditional gradient descent methods are prone to falling into local optima, resulting in overfitting or training stagnation; mostly using fixed learning rates or preset strategies to adjust the learning rate, it is 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 ability when dealing with complex fault modes of power equipment; traditional pruning methods are mainly based on the weight size or sensitivity, and may lose key information. Summary of the Invention

[0004] (I) Technical Problems to be Solved

[0005] In order to solve the above technical problems, the present invention provides a method and system for evaluating the health status of power equipment based on artificial intelligence.

[0006] (II) Technical Solutions

[0007] In order to solve the above existing technical problems and achieve the invention purpose, the present invention is realized through the following technical solutions:

[0008] A method for evaluating the health status of power equipment based on artificial intelligence includes the following steps:

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

[0010] S2: Label the acquired vibration signal data, where the labeling is implemented based on the operating status and fault type of the device;

[0011] S3: Evaluate the health status of power equipment, including evaluating the health status of power equipment based on a convolutional neural network and training 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 features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; applying perturbations to the gradient field; updating the weights of convolutional kernels based on a dynamic adjustment factor; pruning based on the importance of weights; dynamically adjusting the learning rate based on the real-time performance during training; and calculating the loss function by fusing classification error, feature comparison consistency, and dynamic sparse regularization.

[0012] Further, the learning and modulating the spectral components of features based on adaptive spectral transformation adjusts the amplitude of each frequency component through element-wise multiplication.

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

[0014] Further, the updating the weights of convolutional kernels based on a dynamic adjustment factor dynamically focuses on regions with larger losses by backpropagating the perturbed gradients.

[0015] Further, the calculation method of the dynamic adjustment factor is expressed as:

[0016]

[0017] where is the adjustment coefficient of the dynamic adjustment factor; is the gradient of the loss function of the convolutional neural network with respect to the weights of the convolutional kernel of the th layer, is the loss sensitivity, is the value of the loss function.

[0018] Further, the calculation method of pruning based on the importance of weights is to calculate the sum of the absolute values of all convolutional kernel weights of this layer and compare it with a preset threshold, and represent retention or pruning through a binary variable.

[0019] Further, the expression of the loss function of the convolutional neural network is as follows:

[0020]

[0021] where is the classification loss, is the contrast consistency loss, is the dynamic regularization loss; is the classification loss weight; is the contrast consistency loss weight; is the dynamic regularization loss weight.

[0022] Furthermore, the classification loss is calculated using weighted cross - entropy; the contrast consistency loss generates geometric enhanced views of the same signal using Lie group transformation; the dynamic regularization loss combines pruning strategies with gradient perturbation to impose sparse constraints and parameter smoothing constraints.

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

[0024] A power equipment health status data acquisition module, which is used to acquire vibration monitoring signals of power equipment during actual operation;

[0025] A power equipment data annotation module, which is used to annotate data based on the operating status and fault types of the equipment;

[0026] A power equipment health status assessment module, which is used to build 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 features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; applying perturbations to the gradient field; updating the convolutional kernel weights based on a dynamic adjustment factor; pruning based on the importance of weights; dynamically adjusting the learning rate based on the real - time performance during training; fusing classification error, feature contrast consistency, and dynamic sparse regularization to calculate the loss function.

[0027] In addition, to achieve the above - mentioned purpose, the present invention also provides a computer - readable storage medium, on which program instructions for an artificial - intelligence - based power equipment health status assessment method are stored, and the program instructions for the artificial - intelligence - based power equipment health status assessment 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] (III) Beneficial effects

[0029] Compared with the prior art, the beneficial effects of the present invention are:

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

[0031] (2) Through Lie group symmetry operations, the present invention adaptively corrects the signal form changes during the forward propagation of the network, enabling the model to maintain the consistency of feature extraction under different installation angles.

[0032] (3) When calculating the gradient, the present invention introduces small random perturbations to enhance the exploration ability for abnormal samples and noise data, making the model more generalizable.

[0033] (4) By real-time monitoring of the training status, the present invention adaptively adjusts the learning rate, improves the training efficiency, and avoids overfitting to specific fault modes.

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

[0035] (6) By calculating the global importance of weights, the present invention selectively prunes redundant channels, making the model structure more compact, reducing the computational cost, and maintaining high classification accuracy. Description of the Drawings

[0036] The drawings described herein are used to provide a further understanding of the present application and form 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 of the present application. In the drawings:

[0037] Figure 1 is a schematic flow diagram 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 is a comparison of the adaptive spectrum transformation noise robustness analysis results between the method according to an embodiment of the present application and the prior art;

[0039] Figure 3 is a comparison of the geometric perturbation adaptation ability between the method according to an embodiment of the present application and the Lie group symmetry operation layer of the prior art. Detailed Embodiments

[0040] The following describes in detail the embodiments of the present disclosure with reference to the drawings.

[0041] The following specific examples illustrate the embodiments of the present disclosure. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content 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. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0042] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present disclosure in a schematic manner. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in its actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0043] See Figure 1 , a method for evaluating 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] Collect vibration monitoring signals of power equipment during actual operation. The power equipment includes, but is not limited to, various important power facilities such as transformers, generators, and motors. During the daily operation and maintenance of the equipment and in the special detection process, vibration signals are obtained through high-precision vibration sensors installed at specific positions of the equipment. These sensors can real-time sense the minute vibration changes during equipment operation and output the original vibration data in the form of electrical signals.

[0046] In order to comprehensively capture the vibration characteristics of the equipment under different working conditions and potential fault states, sensors are arranged at multiple key positions of the equipment, including parts such as the bearing seat of the motor and the vicinity of the stator and rotor of the generator, to obtain vibration information from different angles. At the same time, the acquisition process covers the normal operation state of the equipment and various fault states artificially simulated or actually occurred, such as vibration signals under different fault modes such as equipment component wear, imbalance, looseness, and overheating, 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 achieve supervised training of the power equipment health status assessment model. The labeling work is based on the operating status and fault types of the equipment. When the acquisition device is in normal operation, the acquired vibration signal is labeled as the "normal" category; when the device exhibits specific fault modes, such as the aforementioned component wear, imbalance, etc., the corresponding vibration signal is accurately labeled as the corresponding "fault" category, such as "bearing wear fault", "rotor imbalance fault", etc.

[0048] S3: Power equipment health status assessment

[0049] Use a convolutional neural network for 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 setting 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 training steps of the convolutional neural network are as follows:

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

[0053] S32. Set the structure of the convolutional neural network in the initialization stage and initialize the weights of the convolutional kernels to avoid the phenomenon of unstable gradients that may occur during the early iteration of the convolutional neural network, expressed as:

[0054]

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

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

[0057] S33. Learn and modulate the spectral components of features based on adaptive spectral transformation;

[0058] Input the obtained power equipment health status data into a convolutional neural network. The vibration signals of power equipment contain key fault features in both the time domain and the frequency domain. Traditional convolutional operations rely on fixed-frequency filters (such as Gabor filters) or short-time Fourier transforms, with fixed frequency band selection. Fixed filters cannot adapt to the noise frequency band shift or resonance frequency change caused by the change of equipment working conditions, resulting in poor robustness in frequency domain feature extraction.

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

[0060]

[0061] In the formula, is the feature representation obtained by converting the output of the th layer of the convolutional neural network into the frequency domain; is the fast Fourier transform, which converts the vibration signal from the time domain to the frequency domain; represents converting to the frequency domain to obtain , is the output of the th layer of the convolutional neural network; is the modulated frequency domain feature; is the spectral modulation vector of the th layer of the learnable convolutional neural network, which adjusts the amplitude of each frequency component through element-wise multiplication. For example, for a bearing fault of a certain power equipment, the resonance frequency is 2 kHz, and the spectral modulation vector learns to enhance the frequency band near 2 kHz (the modulation coefficient is close to 1) and suppress high-frequency noise (>5 kHz, the modulation coefficient is close to 0); is the element-wise multiplication; is the inverse Fourier transform; is the time domain feature restored by the inverse Fourier transform.

[0062] The spectral modulation vector adaptively enhances the key frequency band (such as the resonance frequency of power equipment) and suppresses the noise frequency band, solving the problem of performance degradation caused by the change of equipment working conditions of traditional fixed filters. Its learning method is:

[0063]

[0064] In the formula, is the Sigmoid activation function, which constrains the modulation amplitude in the interval; is a global average pooling operation that extracts the global statistics of the 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 dimension of the frequency-domain features, represents the parameter dimension space, is the bias parameter of the preset fully connected layer that can be learned.

[0065] The preset fully connected layer is a fully connected neural network layer, and its function is to establish a non-linear mapping relationship between the frequency energy and the modulation coefficient.

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

[0067] S34. Process the intermediate features based on Lie group symmetry operations;

[0068] When the vibration signals of power equipment are collected at different installation positions or different angles, the features may be distorted. In the forward propagation process of the convolutional neural network of the present invention, not only conventional convolution, pooling, and non-linear activation operations are performed, but also the intermediate features are processed through Lie group symmetry operations, so as to maintain high invariance under geometric transformations and improve the robustness of the model to distorted (rotation or reflection, etc.) changes, which is expressed as:

[0069]

[0070] In the formula, is the output of the th layer of the convolutional neural network, is the convolutional kernel weight of the th layer of the convolutional neural network, is the Lie group symmetry operation function, is the output of the th layer of the convolutional neural network, is the bias of the th layer of the convolutional neural network, is the Sigmoid activation function, is the convolution operation.

[0071] Furthermore, in a conventional convolutional neural network, methods such as data augmentation are usually used to improve the robustness of the model to geometric transformations, such as rotation, flipping, etc. to expand the training data. However, these methods cannot fundamentally ensure 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 mechanism for processing geometric transformations in the forward propagation process of the network. In the present invention, a Lie group symmetry operation function is used to perform specific geometric transformations on the input features. The Lie group symmetry operation function is designed to maintain the invariance of the intermediate features after geometric transformations (such as rotation or reflection). By applying an operation representing geometric transformation and its corresponding inverse operation before and after feature input respectively, feature correction is achieved, compensating for signal distortion caused by different installation angles or positions of power equipment, so that the model can still extract consistent features when facing data after changes such as rotation and reflection, thereby improving the overall robustness and recognition accuracy. The calculation method is expressed as:

[0072]

[0073] In the formula, is the Lie group symmetry operation function. When the phase shift of the signal waveform is caused by the deviation of the sensor installation angle (for example, when rotating by 5°, the time-domain waveform is delayed by 2 ms), the Lie group operation corrects the feature phase to maintain the time alignment of fault features (such as impact pulses); is a Lie group element, representing geometric transformations such as rotation or reflection. Preferably, is set as the group element representing small-angle rotation. The purpose of using Lie group elements to represent geometric transformations is to accurately simulate the morphological differences that may be caused by small changes in the installation angle during the acquisition of power equipment vibration signals. By using the group element representing small-angle rotation, the network can fine-tune the signal without destroying the original feature structure, so that key features can still be stably extracted in the scenario of multi-angle acquisition, ensuring the adaptability and robustness of the model to signal distortion; is 's inverse operation.

[0074] S35. Apply a perturbation to the gradient field;

[0075] Random high-amplitude noise or abnormal pulses may be mixed in the vibration signals of power equipment. When the traditional gradient descent method updates the network weights, it directly updates according to the calculated gradient without considering adding a perturbation term. When dealing with the random high-amplitude noise or abnormal pulses mixed in the vibration signals of power equipment, it is easy for the model to fall into a local optimal solution, resulting in the stagnation or overfitting of the training process, and reducing the generalization performance of the model. In the present invention, a perturbation is applied to the gradient field and then backpropagation is performed. Specifically, after calculating the gradient of the loss function with respect to the convolutional kernel weights, a random noise term related to the perturbation intensity is added to correct the gradient. That is, on the basis of the original gradient, a small-amplitude random perturbation is additionally added to expand the parameter search space during training, help the model jump out of the local optimum, and enhance the exploration ability for abnormal samples and noisy data, thereby improving the generalization performance of the entire network. The calculation method is expressed as:

[0076]

[0077] In the formula, is the perturbed gradient, which not only retains the key information of the original gradient in indicating the loss descent direction, but also prevents the model from relying too much on the local optimal solution by using random noise, making the network have a certain jump when updating the weights, so as to avoid the stagnation or overfitting of the training process due to noise or abnormal data, and improve the recognition and robustness of complex fault patterns; is the gradient of the loss function with respect to the convolutional kernel weights of the th layer of the convolutional neural network, is the loss function of the convolutional neural network; is the perturbation intensity coefficient. Preferably, is set to 0.1; is the random noise.

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

[0079]

[0080] In the formula, is the sign function, functions to determine the positive and negative of the gradient; is Random noise uniformly distributed within the interval. Preferably, is set in the range of .

[0081] S36. Update the weights of the convolutional kernel based on the dynamic adjustment factor;

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

[0083]

[0084] In the formula, is the weight of the convolutional kernel of the th layer of the updated convolutional neural network, serving as the weight of the convolutional kernel of the th layer of the convolutional neural network for the next iteration; is the learning rate of the convolutional neural network, is the dynamic adjustment factor. Preferably, is set to 0.001.

[0085] Furthermore, during the training process, different fault modes may cause the loss values in some regions to be abnormally high. Conventional convolutional neural networks usually adopt a fixed learning rate or some simple learning rate adjustment strategies, such as step decay, etc., to control the amplitude of weight update. However, for the complex and variable fault modes in the vibration signals of power equipment, these methods cannot dynamically focus on the regions with larger losses, resulting in a slower convergence speed of the model in the key regions and making it difficult to quickly capture the key features and improve the model performance. In order to make the convolutional neural network pay more attention to these high-loss regions, accelerate convergence and enhance the classification discrimination ability, the present invention uses a dynamic adjustment factor to control the loss sensitivity and compensate for the additional attention to the regions with larger losses. Based on the sensitivity of the current layer weights to the loss, when the loss is high, the gradient update amplitude of this region will increase accordingly, thereby prompting the network to make more significant parameter adjustments in these key regions. When facing the complex and variable fault modes in the health state assessment of power equipment, it can capture the key features faster and improve the model performance. The calculation method is expressed as:

[0086]

[0087] In the formula, is the adjustment coefficient of the dynamic adjustment factor. Preferably, is set to 0.1; is the gradient of the loss function of the convolutional neural network with respect to the weight of the convolutional kernel of the th layer, is the loss sensitivity, is the loss function value, preferably, is set to 0.01.

[0088] S37. Pruning based on weight importance;

[0089] Common neural network pruning methods include pruning based on weight magnitude, pruning based on sensitivity, etc. By setting certain thresholds or rules to prune unimportant weights or channels, but when these methods are applied to convolutional neural networks, they may not fully consider the characteristics of power equipment vibration signals and the complexity of the network structure, resulting in an inefficient network structure after pruning or limited effect in improving generalization ability. For the large number of redundant or invalid feature channels that may appear in convolutional neural networks, the present invention prunes after evaluating weight importance, avoiding the computational burden and overfitting problems caused by an overly large model scale, simplifying the convolutional neural network structure, and improving generalization ability. The calculation method is expressed as:

[0090]

[0091] In the formula, is a binary variable, representing whether the current layer is pruned; is the sum of the absolute values of the weights of the th layer; is the number of weights of the convolutional kernel of the convolutional neural network layer, is the th convolutional kernel of the convolutional neural network at the th weight; is the pruning threshold, preferably, is set to 0.5.

[0092] Furthermore, by calculating the sum of the absolute values of all the weights of the convolutional kernels of this layer and comparing it with a pre-set threshold, when the sum is lower than the threshold, the binary variable takes "1", indicating that this channel does not have sufficient importance and can thus be pruned; otherwise, it takes "0" to indicate retention. The binary variable acts as a switch throughout the pruning process, determining which channels are pruned, thereby making the network structure more concise and efficient while improving generalization ability.

[0093] S38. Dynamically adjusting the learning rate based on the real-time performance during training;

[0094] When the number of samples of the vibration signal of power equipment is limited and the distribution is variable, it is easy to fall into training stagnation. Conventional learning rate adjustment methods are mostly based on preset strategies, such as decaying at fixed intervals, manual adjustment based on the performance of the validation set, etc. These methods can adapt to the changes during training to a certain extent. However, for the situation where the number of samples of the vibration signal of power equipment is limited and the distribution is variable, these methods may not be able to dynamically adjust the learning rate according to the training state in a timely and accurate manner, easily leading to the model falling into training stagnation or overfitting to specific fault patterns, 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 the model overfitting to specific fault patterns and maintain the flexibility and adaptability of the model. The calculation method is expressed as:

[0095]

[0096] In the formula, is the learning rate of the updated convolutional neural network and serves as the learning rate for the next iteration of the convolutional neural network; is the decay rate. Preferably, is set to 0.01; is the current training epoch.

[0097] S39. Calculate the loss function by fusing classification error, feature comparison consistency, and dynamic sparse regularization;

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

[0099]

[0100] In the formula, is the classification loss, is the comparison consistency loss, is the dynamic regularization loss; is the classification loss weight. Preferably, is set to 0.6; is the comparison consistency loss weight. Preferably, is set to 0.2; is the dynamic regularization loss weight. Preferably, is 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 fault patterns, which is expressed as:

[0102]

[0103] In the formula, is the number of batch samples, representing the quantity of power equipment vibration signal data input into the convolutional neural network in the current batch; is a positive integer, is the number of fault categories, is a positive integer; is the weight coefficient of the th category, and the calculation method is , is the th category's occurrence frequency in the training set. The weight coefficient of low - frequency fault categories is larger, strengthening the model's attention to them; is the true label (one - hot encoding format) of the th sample, is the probability of the th class output by Softmax. is the output of the convolutional neural network for the th sample at the th layer, is the total number of layers of the convolutional neural network.

[0104] The contrastive consistency loss uses Lie group transformation to generate geometric enhanced views of the same signal. By minimizing the difference between the original features and the enhanced features, it constrains the similarity in the feature space, forcing the model to ignore irrelevant geometric transformations and improve the robustness to installation angle changes. The calculation method is expressed as:

[0105]

[0106] In the formula, is the feature projection head (fully - connected layer), which maps features to a low - dimensional space; is the Lie group symmetry operation function, is the cosine similarity function, is the output of the convolutional neural network for the th sample at the L - th layer.

[0107] The dynamic regularization loss combines pruning strategies with gradient perturbation, imposing sparse constraints (implemented based on the L1 norm) and parameter smoothing constraints (based on the L2 norm). The calculation method is expressed as:

[0108]

[0109] In the formula, is a binary variable, representing whether pruning is performed on the current layer. When pruning is activated ( ), an L1 sparse penalty is applied; is the L1 norm, is the L2 norm; is the adjustment coefficient of the dynamic adjustment factor. Preferably, is set to 0.1.

[0110] S310. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0111] When using the trained convolutional neural network for power equipment health status evaluation and classification, first, the on-site power equipment vibration monitoring signal needs to be preprocessed by normalization 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 eliminate the possible amplitude differences in the original signal;

[0112] Furthermore, input the preprocessed vibration signal into the trained convolutional neural network. The network will automatically perform the forward propagation process, successively passing through operations such as convolutional layers, pooling layers, and fully connected layers, etc., to extract and transform the features of the input signal;

[0113] Furthermore, the classification output layer of the network uses the Softmax activation function to map the extracted features to each category of the power equipment health status, output the probability values of each health status category. According to the magnitude of the output probability values, the current most likely health status of the power equipment can be determined (take the corresponding category with the largest probability value), realizing the accurate evaluation and classification of its health status.

[0114] To verify the advantages of the present invention, the following experiments are carried out for verification:

[0115] To verify the effectiveness of the adaptive spectrum transformation module in a complex noise environment, by comparing with the traditional fixed filter scheme and the benchmark model without introducing the spectrum modulation mechanism, this experiment simulates the equipment fault recognition scenarios under different intensities of noise interference. The experimental results are as Figure 2 shown. It can be seen that the technical solution adopting 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 bands according to the frequency domain characteristics of the input signal, while effectively suppressing the high-frequency noise components, overcoming the defect of the rigid frequency response characteristics of the fixed filter when the working conditions change, and significantly improving the robustness of the model to noise interference.

[0116] Regarding the geometric perturbation adaptation ability of the Lie group symmetry operation layer, by constructing a signal distortion test set with different rotation angles, the performance differences between the benchmark model without Lie group operations and the present technology are compared and analyzed. As Figure 3 shown, the experimental results show that the present invention can still maintain a high classification accuracy in the presence of significant geometric perturbations, and the performance decay amplitude is much smaller than that of traditional methods. This indicates that the Lie group symmetry operation enables the network to automatically correct the signal morphology changes caused by the sensor installation angle deviation by establishing an algebraic representation of geometric transformations, effectively improving the invariance of feature representation to geometric perturbations and solving the problem of feature distortion caused by traditional methods ignoring the signal spatial relationship.

[0117] In this embodiment, the key frequency bands are enhanced and the noise is suppressed through frequency domain modulation to improve the robustness to complex environments; through the Lie group symmetry operation, the signal morphology changes are adaptively corrected during the forward propagation of the network, so that the model can still maintain the consistency of feature extraction at different installation angles; a small random perturbation is introduced when calculating the gradient to enhance the exploration ability for abnormal samples and noise data, making the model more generalizable; by real-time monitoring the training status, the learning rate is adaptively adjusted to improve the training efficiency and avoid overfitting to specific fault modes; by fusing the classification error, contrast consistency loss and dynamic sparse regularization, the fault recognition ability of the model under geometric distortion and complex working conditions is improved; by calculating the global importance of weights, redundant channels are selectively pruned to make the model structure more compact, the computational cost lower, and at the same time maintain a high classification accuracy.

[0118] The embodiment of the present invention also proposes an artificial intelligence-based power equipment health status evaluation system, including:

[0119] A power equipment health status data acquisition module, which is used to acquire the vibration monitoring signals of the power equipment during actual operation;

[0120] A power equipment data annotation module, which is used to annotate the data based on the operating status and fault types of the equipment;

[0121] A power equipment health status evaluation module, which 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 the adaptive spectral transformation; processing the intermediate features based on the Lie group symmetry operation; applying perturbations to the gradient field; updating the convolutional kernel weights based on the dynamic adjustment factor; pruning based on the weight importance; dynamically adjusting the learning rate based on the real-time performance during the training process; fusing the classification error, feature contrast consistency and dynamic sparse regularization to calculate the loss function.

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

[0123] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An artificial intelligence-based method for evaluating the health status of power equipment, characterized in that, It includes the following steps: S1: Obtaining the health status data of power equipment, including collecting vibration monitoring signals of power equipment during actual operation; S2: Annotating the obtained vibration signal data, which is implemented based on the operating status and fault type of the equipment; S3: Evaluating the health status of power equipment, including evaluating the health status of power equipment based on a convolutional neural network and training 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 features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; Applying perturbations to the gradient field; updating the convolutional kernel weights based on a dynamic adjustment factor; Pruning based on the importance of weights; Dynamically adjusting the learning rate based on the real-time performance during the training process; calculating the loss function by fusing classification error, feature contrast consistency, and dynamic sparse regularization.

2. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, wherein The learning and modulating the spectral components of features based on adaptive spectral transformation adjusts the amplitude of each frequency component through element-wise multiplication.

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

4. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, wherein The updating of the convolutional kernel weights based on a dynamic adjustment factor dynamically focuses on regions with larger losses by backpropagating the perturbed gradients.

5. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 4, wherein, The dynamic adjustment factor The calculation method is expressed as: ; where, is the adjustment coefficient of the dynamic adjustment factor; is the gradient of the loss function of the convolutional neural network with respect to the convolutional kernel weights of the th layer, is the loss sensitivity, is the loss function value.

6. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, characterized in that, The pruning calculation method based on the importance of weights calculates the sum of the absolute values of all convolutional kernel weights in this layer, compares it with a preset threshold, and represents retention or pruning through binary variables.

7. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 1, wherein, The expression of the loss function L of the convolutional neural network is as follows: ; where, is the classification loss, is the contrast consistency loss, is the dynamic regularization loss; is the classification loss weight; is the contrast consistency loss weight; is the dynamic regularization loss weight.

8. The method for evaluating the health status of power equipment based on artificial intelligence according to claim 7, wherein The classification loss is calculated using weighted cross-entropy; the contrast consistency loss generates geometric enhanced views of the same signal using Lie group transformation; the dynamic regularization loss combines the pruning strategy and gradient perturbation to impose sparse constraints and parameter smoothing constraints.

9. A system based on the method for evaluating the health status of power equipment based on artificial intelligence according to any one of claims 1-8, which includes: A module for obtaining the health status data of power equipment, which is used to obtain vibration monitoring signals of power equipment during actual operation; A module for annotating power equipment data, which is used to annotate data based on the operating status and fault type of the equipment; A module for evaluating the health status of power equipment, which is used to construct a convolutional neural network and train it; 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 features based on adaptive spectral transformation; processing intermediate features based on Lie group symmetry operations; Applying perturbations to the gradient field; Updating the convolutional kernel weights based on a dynamic adjustment factor; Pruning based on the importance of weights; Dynamically adjusting the learning rate based on the real-time performance during the training process; Calculating the loss function by fusing classification error, feature contrast consistency, and dynamic sparse regularization.

10. A computer-readable storage medium, characterized in that, Program instructions for an artificial intelligence-based power equipment health status assessment method are stored on the computer-readable storage medium, and the program instructions for the artificial intelligence-based power equipment health status assessment 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 in any one of claims 1-8.

Images