Power equipment inspection optimization method and device based on deep learning

Abstract

The invention provides an inspection optimization method and device for power equipment based on deep learning. The method comprises the following steps: acquiring an image signal and a time sequence signal acquired for power equipment; carrying out denoising processing on the time sequence signal by adopting an adaptive wavelet denoising method, and carrying out denoising processing on the image signal by adopting a first convolutional neural network; performing abnormal value detection on the denoised time sequence signals and the high-level image features by adopting an abnormal value detection algorithm based on dynamic density perception; reconstructing abnormal values in the image signal and the time sequence signal based on a classification abnormal value reconstruction algorithm to obtain a corrected time sequence signal and image signal; therefore, the quality of data collected in the inspection process of the power equipment is improved, high-quality time sequence signals and image signals are input into the hybrid intelligent model, the equipment state prediction result of the power equipment is output, and the inspection accuracy of the power equipment is further improved.

Description

Technical Field

[0001] The present invention relates to the field of equipment monitoring technology, and in particular to a patrol optimization method and device for power equipment based on deep learning. Background Art

[0002] Power equipment is a core component of the power system, and its proper operation directly impacts the stability of the power supply. Power equipment not only provides electricity for homes and industries but also plays a vital role in critical sectors such as transportation, communications, and healthcare. However, due to the complexity of power equipment, any failure or anomaly can cause widespread power outages, disrupting business operations, causing economic losses, and potentially endangering public safety. Inspections can promptly identify and correct potential faults, such as equipment overheating, aging, and mechanical damage, playing a crucial role in ensuring the continuity and stability of the power supply.

[0003] In recent years, power inspection technology has made significant progress, with a wide range of monitoring methods being widely used. These methods, equipped with a variety of sensors such as high-definition cameras, infrared equipment, lidar, temperature and humidity sensors, and voltage sensors, enable comprehensive monitoring of transmission lines and equipment. These sensors capture data from different dimensions. Visual image data includes images of equipment exteriors and infrared thermal imaging. Time series data includes environmental monitoring data such as temperature, humidity, and wind speed, as well as electrical data such as voltage and current.

[0004] However, these monitoring methods are susceptible to environmental factors during operation, such as wind speed, temperature, and humidity fluctuations, which can affect the quality of collected data. For example, in harsh conditions like strong winds and heavy rainfall, monitoring equipment can malfunction, leading to interruptions in data collection or degradation of data quality. This not only impacts inspection tasks but can also lead to potential missed faults or misdiagnosis. Summary of the Invention

[0005] The embodiments of the present invention provide a deep learning-based inspection optimization method and device for power equipment to solve the problem of poor inspection accuracy of existing power equipment.

[0006] In a first aspect, an embodiment of the present invention provides a method for optimizing inspection of power equipment based on deep learning, comprising:

[0007] Acquire image signals and time sequence signals collected from the power equipment; the image signals include an appearance image of the power equipment and infrared thermal imaging data; the time sequence signals include environmental data and electrical data of the power equipment;

[0008] Performing denoising processing on the time series signal using an adaptive wavelet denoising method, performing denoising processing on the image signal using a first convolutional neural network, and outputting high-level image features of the image signal;

[0009] An outlier detection algorithm based on dynamic density perception is used to perform outlier detection on the denoised time series signal and the high-level image features respectively;

[0010] reconstructing the outliers in the image signal and the time series signal based on a classification outlier reconstruction algorithm to obtain a corrected time series signal and image signal;

[0011] The corrected timing signal and image signal are input into a hybrid intelligent model to output a prediction result of the equipment status of the power equipment. The hybrid intelligent model is constructed based on a deep learning algorithm.

[0012] In a second aspect, an embodiment of the present invention provides an inspection optimization device for power equipment based on deep learning, comprising:

[0013] A signal acquisition module is used to acquire image signals and time sequence signals collected from the power equipment; the image signals include the appearance image and infrared thermal imaging data of the power equipment; the time sequence signals include the environmental data and electrical data of the power equipment;

[0014] a denoising module, configured to perform denoising processing on the time series signal using an adaptive wavelet denoising method, perform denoising processing on the image signal using a first convolutional neural network, and output high-level image features of the image signal;

[0015] An outlier detection module, configured to perform outlier detection on the denoised time series signal and the high-level image features using an outlier detection algorithm based on dynamic density perception;

[0016] an outlier reconstruction module, configured to reconstruct outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain a corrected time series signal and image signal;

[0017] The device status prediction module is used to input the corrected timing signal and image signal into the hybrid intelligent model and output the device status prediction result of the power equipment. The hybrid intelligent model is constructed based on the deep learning algorithm.

[0018] An embodiment of the present invention provides an inspection optimization method and device for power equipment based on deep learning. The method first obtains image signals and time series signals collected from the power equipment; then adopts an adaptive wavelet denoising method to denoise the time series signal, and adopts a first convolutional neural network to denoise the image signal; adopts an outlier detection algorithm based on dynamic density perception to perform outlier detection on the denoised time series signal and the high-level image feature respectively; reconstructs the outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain corrected time series signals and image signals; improves the quality of data collected during the inspection of the power equipment, and finally inputs the high-quality time series signals and image signals into a hybrid intelligent model to output the equipment status prediction result of the power equipment. The hybrid intelligent model is constructed based on a deep learning algorithm, and the model has high accuracy, thereby further improving the inspection accuracy of the power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is a flowchart of an implementation method for optimizing inspection of power equipment based on deep learning provided by an embodiment of the present invention;

[0021] Figure 2 Schematic diagram of the structure of the hybrid intelligent model provided by an embodiment of the present invention;

[0022] Figure 3 1 is a schematic structural diagram of a deep learning-based inspection and optimization device for power equipment provided by an embodiment of the present invention;

[0023] Figure 4 is a schematic diagram of a terminal provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In the following description, specific details such as particular system structures and techniques are provided for purposes of illustration, not limitation, to facilitate a thorough understanding of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

[0025] In the description of this application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0027] In addition, the “plurality” mentioned in the embodiments of the present application should be interpreted as two or more.

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clear, specific embodiments will be described below with reference to the accompanying drawings.

[0029] See also Figure 1 , which shows an implementation flow chart of the inspection optimization method for power equipment based on deep learning provided by an embodiment of the present invention, and is detailed as follows:

[0030] S101: Acquire image signals and time sequence signals collected from power equipment; the image signals include an appearance image of the power equipment and infrared thermal imaging data; the time sequence signals include environmental data and electrical data of the power equipment.

[0031] This embodiment can be implemented by a terminal device. The terminal device first collects multi-source data from power equipment using various types of sensors. This includes time-series signals and image signals, where the time-series signals include electrical and environmental data. Electrical data includes voltage and current. The terminal can obtain voltage on power equipment such as transformers and circuit breakers through voltage transformers, and current through current transformers. Environmental data, including temperature, humidity, and wind speed, is collected using temperature sensors, humidity sensors, and anemometers deployed around the power equipment, respectively. This data is used to assess the impact of the external environment on equipment operation.

[0032] The image signal includes image data and infrared thermal imaging data. In this embodiment, multiple high-resolution cameras can be installed at different positions and angles on power towers and fixed monitoring devices to collect equipment appearance images and infrared thermal imaging data for subsequent detection of possible physical damage, corrosion, and overheating areas on the equipment surface.

[0033] S102: Using an adaptive wavelet denoising method to denoise the time series signal, using a first convolutional neural network to denoise the image signal, and outputting high-level image features of the image signal.

[0034] To ensure both data quality and the real-time nature of power inspections, this embodiment employs different denoising methods for image data and time series data. Time series data is denoised using an adaptive wavelet denoising algorithm, while image data is denoised using a convolutional neural network.

[0035] S103: Using an outlier detection algorithm based on dynamic density perception to perform outlier detection on the denoised time series signal and the high-level image features respectively.

[0036] S104: reconstructing the outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain a corrected time series signal and image signal.

[0037] S105: Inputting the corrected timing signal and image signal into a hybrid intelligent model, and outputting a prediction result of the equipment state of the power equipment, wherein the hybrid intelligent model is constructed based on a deep learning algorithm.

[0038] As can be seen from the above examples, this embodiment addresses the data quality limitations of existing inspection systems due to environmental factors. It proposes multi-source data fusion and intelligent preprocessing technology. By using multi-angle and multi-dimensional sensors to collect electrical parameters, environmental parameters, and visual image data, it provides comprehensive information on the status of power equipment. To address potential environmental interference during data acquisition, the present invention introduces an adaptive multimodal denoising strategy, using dynamic wavelet denoising for time series data and convolutional neural network denoising for image data to reduce the interference of environmental noise on the data. Secondly, to maintain high-quality data input under harsh environmental conditions, an outlier detection algorithm based on dynamic density perception and a classified outlier reconstruction algorithm are designed to identify and process abnormal data. Through these technical means, the present system effectively reduces the interference of environmental factors on data quality, ensuring high-quality and comprehensive data input even in harsh environmental conditions, providing reliable basic data for subsequent deep learning models. Finally, to address the lack of accuracy of power inspection fault diagnosis systems in complex power equipment environments, this embodiment constructs a deep learning hybrid intelligent model that combines convolutional neural networks, long short-term memory networks, and temporal convolutional networks. This model comprehensively extracts spatial and temporal features from multimodal data and concatenates all these features to form a joint feature vector. This feature vector, incorporating spatial and temporal information from multiple sources, delivers more accurate equipment status assessment and fault classification results. By applying this deep learning model, the system's fault diagnosis capabilities in complex scenarios are significantly improved, ensuring accurate inspections.

[0039] In one possible implementation, the specific implementation process of S102 includes:

[0040] Performing Fourier transform on the time series signal to obtain a frequency spectrum of the time series signal;

[0041] According to the formula Determining the number of layers of wavelet decomposition, and performing wavelet decomposition on the time series signal based on the layers to obtain an approximate part and a detail part;

[0042] Using soft thresholding Eliminating high-frequency noise in the detail part, performing wavelet reconstruction on the detail part and the approximate part after the high-frequency noise is eliminated, and converting the spectrum after wavelet reconstruction into the time domain to obtain a denoised time series signal;

[0043] Among them, L represents the level of wavelet decomposition, T represents the signal sampling length, k represents the compensation coefficient, round() represents the rounding function, and f main represents the main frequency component of the spectrum, α represents the first preset value, σ f represents the standard deviation of the spectrum; D i' represents the soft threshold, D i represents the i-th component of the detail part, σ represents the noise standard deviation, and N represents the signal length of the time series signal.

[0044] This embodiment proposes a wavelet denoising algorithm for time series data that dynamically adjusts the number of wavelet decomposition layers based on data characteristics. Specifically, the spectral characteristics of the time series data are analyzed, using a fast Fourier transform algorithm to calculate the primary frequency components and energy distribution of the time series signal. The number of wavelet decomposition layers is dynamically selected based on these characteristics. For signals with a high number of high-frequency components or dispersed frequency components, more decomposition layers are used to capture high-frequency noise; for signals with a high number of low-frequency components, fewer decomposition layers are used to reduce the computational burden.

[0045] Specifically, use Determine the number of layers of wavelet decomposition, where σ f It represents the standard deviation of the spectrum and represents the spectrum width of the signal. α is used to ensure that the calculation of is reasonable and adapts to the actual scenario.

[0046] After performing wavelet decomposition on the time series signal, the approximate part A can be obtained L and details D1, D2, ..., D i ,...D L For the details, a soft threshold is used to eliminate high-frequency noise. Finally, the processed details and the approximate part are reconstructed by wavelet to obtain the denoised signal.

[0047] The collected image data is denoised by training a lightweight convolutional neural network. This network consists of multiple convolutional and deconvolutional layers. During the encoding phase, multiple convolutional and pooling layers are used to gradually extract multi-scale features of the image. During the decoding phase, the extracted features are decoded through deconvolutional layers, gradually restoring them to the same resolution as the input, resulting in a high-quality denoised image output.

[0048] In one possible implementation, to optimize the denoising effect, this embodiment proposes a hybrid loss function consisting of two parts: mean square error and perceptual loss. The loss function of the first convolutional neural network is:

[0049] Among them, β1 and β2 represent weight parameters, Y i Represents the i-th image pixel value in the original image, Y i ' represents the pixel value of the i-th image after denoising; φ j Represents the j-th layer feature extraction result in the VGG network, n represents the number of image pixels, and m represents the number of feature channels.

[0050] In one possible implementation, the specific implementation process of S103 includes:

[0051] Determine the local density of each time series data point in the denoised time series signal;

[0052] Based on the formula Identify outliers in the denoised time series signal;

[0053] Among them, Anomaly(x i ) represents the abnormal sign of the i-th time series data point. When Anomaly(x i )=1, it means that the i-th time series data point is an abnormal value; when Anomaly(x i )=0, it means that the i-th time series data point is a normal value; ρ(x i ) represents the local density of the i-th time series data point;

[0054]

[0055] Among them, θ(x i ) represents the dynamic density threshold of the i-th time series data point, β3 and γ represent weights, and ε(x i ) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th time series data point, δ(x i ) represents the density change factor of the i-th time series data point; d(x i ,x j ) represents the Euclidean distance from the i-th time series data point to the j-th time series data point; h-NN(x i ) represents the h nearest neighbors of the i-th time series data point, α1 represents the second preset value, and median() represents the median function.

[0056] In this embodiment, an outlier detection algorithm based on dynamic density sensing is proposed for denoised time series data. By introducing a dynamic density sensing mechanism to adapt to density changes in different data regions, the accuracy of outlier detection is improved, thereby effectively screening out abnormal data.

[0057] In order to cope with the change of data density and adapt to the local characteristics of data distribution, the algorithm introduces the adaptive neighborhood radius ε(x i ), where the median is used instead of the mean to calculate the neighborhood radius to reduce the impact of outliers or isolated points on the radius calculation. In order to measure the density fluctuation around the time series data points, the density change factor δ(x i ). If the density fluctuation around a certain time series data point is large, then δ(x i ) value is high, the time series data point is more likely to be an outlier. Therefore, δ(x i ) can help distinguish areas with significant local density changes and improve the sensitivity of outlier detection.

[0058] Based on the above-mentioned density perception mechanism, this embodiment combines the average distance within the neighborhood with the density change factor to obtain a dynamic density threshold to adapt to the changes in the local density of the data distribution, thereby enhancing the algorithm's ability to identify outliers in sparse and dense areas.

[0059] In a possible implementation, the specific implementation process of S103 further includes:

[0060] Performing dimensionality reduction processing on the high-level image features to obtain an image signal after dimensionality reduction processing;

[0061] Based on the formula determining outliers in the denoised image signal;

[0062] Among them, Anomaly(y i ) represents the abnormal sign of the i-th image data point. When Anomaly(y i )=1, it means that the i-th image data point is an abnormal value; when Anomaly(y i )=0, it means that the i-th image data point is a normal value; ρ(y i ) represents the local density of the i-th image data point;

[0063]

[0064] Among them, θ(y i ) represents the dynamic density threshold of the i-th image data point, α2, α3, α4 represent weights, ε(y i ) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th image data point, δ(y i ) represents the density change factor of the i-th image data point; d(y i ,y j ) represents the Euclidean distance from the i-th image data point to the j-th image data point; h-NN(x i ) represents the h nearest neighbors of the i-th image data point, α1 represents the second preset value, and median() represents the median function; μ i represents the mean of the local area where the i-th image data point is located, σ i represents the variance of the local area where the i-th image data point is located; median() represents the median function, and S represents the total number of image feature extraction layers under multi-scale.

[0065] For the denoised image data, the first convolutional neural network in S102 is used to extract high-level image features of the image signal, the principal component analysis method is used to reduce the dimension of the extracted high-level image features, and then the density perception mechanism is used to detect outliers.

[0066] Unlike time series data, the density change factor of image data is calculated by combining the multi-scale local features of the image, including gradient, texture, etc. The dynamic density threshold is calculated and adjusted by combining statistical features such as local variance and texture complexity of the local area of the image. The specific definition is as follows: θ(y i )=α1·μ i +α2·σ i +α3·δ i Among them, μ i is the mean of the local area where the image data point is located, which is used to describe the average brightness information of the local area, σ i It is the variance of the local area where the image data point is located, which is used to describe the texture complexity of the area.

[0067] After determining the dynamic density threshold, the local density of the image data points is calculated, and then the formula is used Identify outliers in the image signal.

[0068] In one possible implementation, the specific implementation process of S104 includes:

[0069] Obtaining an average density of all image data points in the image signal, and using the average density of the image data points as an image density threshold;

[0070] dividing the outliers in the image data points into isolated outliers and intra-cluster outliers according to the image density threshold;

[0071] For each isolated outlier in the image data points, a local polynomial fitting algorithm is used to fit the image data points in the neighborhood of the isolated outlier to obtain a first local interpolation model;

[0072] Reconstructing the isolated outlier using the first local interpolation model;

[0073] For each in-cluster outlier in the image data point, the formula is used Reconstruct the outliers in the cluster;

[0074] Among them, y i ' represents the i-th image data point after reconstruction, y i represents the abnormal point in the i-th cluster, α5 represents the smoothing factor, represents the gradient operator, ρ i represents the density mean of the neighborhood data points of the outlier in the i-th cluster, ρ threshold-1 Indicates the image density threshold.

[0075] For the detected outliers, this embodiment uses a classification outlier reconstruction algorithm to restore the real data information near the outliers. The image density threshold ρ is set according to the average density of the image data set corresponding to the image signal. threshold ,According to the image density threshold, the detected outliers are divided into isolated outliers and,intra-cluster outliers.,Among them, isolated outliers are located in sparser data areas,,and intra-cluster outliers are located in denser data areas.,Because of the large local density changes, they are detected as,outliers.

[0076] For isolated outliers in image data, local interpolation is used for reconstruction. Local polynomial fitting is used to fit the data points in the neighborhood to obtain a local interpolation model. The specific model is: f(y) = a0 + a1y + a2y 2 +…+a n y n Among them, the coefficients a0, a1, ..., an are obtained by fitting the least square method of the image data points in the neighborhood. The outlier y in the image data is interpolated by the interpolation model. i Reconstruct and get new data points: y i '=f(y i ). For the outliers within the cluster of image data, in order to avoid over-smoothing in the image detail area, the gradient information of the image is considered in the correction process. The specific formula is: To ensure the rationality of the correction results, the system uses an adaptive feedback mechanism to reapply the above-mentioned outlier detection algorithm based on dynamic density perception to the corrected image data points, and dynamically adjusts the correction strategy to ensure that the corrected data points achieve the expected quality and consistency.

[0077] In a possible implementation, the specific implementation process of S104 further includes:

[0078] Obtaining an average density of all time series data points in the time series signal, and using the average density of the time series data points as an image density threshold;

[0079] dividing the outliers in the time series data points into isolated outliers and intra-cluster outliers according to the image density threshold;

[0080] For each isolated outlier in the time series data points, a local polynomial fitting algorithm is used to fit the time series data points in the neighborhood of the isolated outlier to obtain a second local interpolation model;

[0081] Reconstructing the isolated outlier using the second local interpolation model;

[0082] For each in-cluster outlier in the time series data point, the formula x i '=x i -α6·(ρi -ρ threshold-2 ) reconstruct the outliers in the cluster;

[0083] Among them, x i ' represents the i-th time series data point after reconstruction, x i represents the outlier in the i-th cluster, α6 represents the smoothing factor, and ρ i represents the density mean of the outlier neighborhood data points of the outlier in the i-th cluster, ρ threshold-2 Indicates the time series density threshold.

[0084] For isolated outliers in time series data, local interpolation is used to reconstruct them. Local polynomial fitting is used to fit the data points in the neighborhood to obtain a local interpolation model. The specific model is: f(x) = a0 + a1x + a2x 2 +…+a n x n Among them, the coefficients a0, a1, ..., an are obtained by fitting the least square method of the data points in the neighborhood. i Reconstruct and get new data points: x i '=f(x i ). For outliers within the cluster of time series data, smooth correction is performed in combination with local density information. The specific formula is as follows: i '=x i -α6·(ρ i -ρ threshold ). Among them, ρ i is the density mean of the outlier neighborhood data points. i is an abnormal data point, x i ' is the corrected time series data point, and α6 is the smoothing factor used to control the correction amplitude.

[0085] In a possible implementation, after S104, the method provided in this embodiment further includes:

[0086] The time series signals collected by different sensors are synchronized based on the multi-dimensional dynamic time warping (MD-DTW) algorithm with the introduction of error compensation mechanism.

[0087] Specifically, after obtaining the reconstructed time series data, considering that there are many sensors in the power inspection system and there is nonlinear time offset due to the different characteristics of the sensors and environmental factors, this embodiment proposes a multi-dimensional dynamic time warping (MD-DTW) algorithm that introduces an error compensation mechanism to synchronize the time series data collected by different sensors. The multi-dimensional dynamic time warping (MD-DTW) algorithm is a method for measuring and aligning the similarity of time series in multiple dimensions, and is widely used in speech recognition, data mining, bioinformatics and other fields. The MD-DTW algorithm that introduces the error compensation mechanism is based on this, and further considers the errors that may occur in the time series matching process, and compensates for these errors in a certain way to improve the accuracy and reliability of the algorithm.

[0088] The MD-DTW algorithm can handle nonlinear time deformation and align multiple different types of sensor time series data to ensure optimal alignment in multidimensional space. The MD-DTW cumulative distance matrix D(i,j) is obtained by calculating the sum of the absolute differences in all dimensions. The specific formula is: Among them, A(i,k) and B(j,k) represent the values of the kth dimension at the i-th time point and the j-th time point respectively, and K represents the total dimension of all time series data. In order to compensate for the error caused by complex nonlinear time offset, the error compensation term ε(i,j) is introduced and expressed as Among them, |ΔA(i)-ΔB(j)| is the compensation for the sensor time step difference, which is mainly used to correct the step size deviation caused by different sensor sampling frequencies or environmental noise. This compensates for accumulated time drift, primarily over long periods of time, ensuring a more stable alignment path over long time spans. Therefore, the cumulative distance matrix D'(i, j) after error compensation can be expressed as: D'(i, j) = D(i, j) + ε(i, j). The final alignment path P is found from the cumulative matrix D'(i, j) using a backtracking algorithm: P = argminD'(i, j).

[0089] The aligned time series is generated according to the optimal path P. After time alignment, the aligned multi-source data are integrated into a comprehensive dataset in a unified format through weighted data fusion, ensuring that information from different data sources is effectively integrated and improving data quality.

[0090] In one possible implementation, the hybrid intelligent model includes a convolutional neural network module, a long short-term memory network module, a temporal convolutional network module, and a feature fusion module; the specific implementation process of S105 includes:

[0091] The corrected time series signal is input into the Long Short-Term Memory (LSTM) module, which outputs a comprehensive time series feature vector.

[0092] The corrected image signal is input into the Convolutional Neural Network (CNN) module to output a high-level image feature vector;

[0093] Inputting the comprehensive temporal feature vector into the temporal convolutional network (TCN) module to output a high-level temporal feature vector;

[0094] The high-level image feature vector and the high-level time series feature vector are input into the feature fusion module, and the device state prediction result of the power device is output.

[0095] In this embodiment, a hybrid intelligent model is constructed by combining convolutional neural networks, long short-term memory networks and temporal convolutional networks to achieve parallel processing and feature fusion of image data and time series data. Specifically, the hybrid intelligent model includes three modules, among which the CNN module and the LSTM / TCN module are in parallel, processing data from the spatial dimension and the temporal dimension respectively. Finally, through the feature fusion module, the modules work together to achieve comprehensive processing and prediction of the data.

[0096] Specifically, the hybrid intelligent model processes data in both spatial and temporal dimensions through parallel processing. Specifically, the CNN module processes high-resolution visual image data, capturing spatial features in the image layer by layer through multi-layer convolution operations, such as abnormal conditions such as cracks and mechanical damage on the device surface. Finally, it outputs a high-level feature vector (FCNN) reflecting the device status. The LSTM module focuses on processing time series data such as voltage, current, temperature, and humidity. It captures long-term dependencies in the data through a multi-layer LSTM network, ultimately extracting and outputting comprehensive time series features (FLSTM). Based on the time series features extracted by the LSTM module, the TCN module further processes these features, capturing data dependencies and regularities over longer time spans through multi-layer dilated convolution operations, enabling the entire system to conduct a deeper analysis of complex temporal patterns. Ultimately, it outputs high-level time series features (FTCN), enhancing the feature representation of time series data.

[0097] In the feature fusion module, the image features (FCNN) extracted by the CNN module are combined with the time series features (FTCN) extracted by the LSTM and TCN modules to form a joint feature vector. This feature vector integrates the spatial and temporal information of multi-source data, ultimately outputting highly accurate predictions of device status and fault classification, ensuring the system's effective fault detection and prevention capabilities in complex power inspection environments.

[0098] In a possible implementation, after S105, the method provided in this embodiment further includes:

[0099] Acquiring an actual device state of the electric device, and determining a current prediction error based on the actual device state of the electric device and a device state prediction result;

[0100] According to the formula determining a dynamic error threshold, and determining that a detection anomaly exists in the current device state prediction result when the absolute value of the current prediction error is greater than the dynamic error threshold;

[0101] When there is a detection anomaly in the current device state prediction result, the parameters in the hybrid intelligent model are updated based on the adaptive gradient boosting Bayesian optimization algorithm;

[0102] Among them, T d (t) represents the dynamic error threshold, μ(t) represents the mean value of the prediction error in the time window t, and i Represents the prediction error for time i, t represents the length of the time window, σ t represents the standard deviation of the prediction error within the time window t; η represents the first preset coefficient.

[0103] After the model training phase, a real-time feedback mechanism is established. Based on the comparison between actual monitoring data and model prediction results, an adaptive Bayesian optimization algorithm is introduced to adjust model parameters and continuously optimize system performance.

[0104] Specifically, the image and time series signals of the power equipment are monitored in real time. The image and time series signals processed through S101 to S104 are then input into the hybrid intelligent model to obtain the equipment status prediction results. The error between the actual equipment status and the predicted equipment status is then calculated, i.e., the prediction error. The specific formula is as follows: Where i represents the time point, is the prediction error, Y actual is the actual device status, Y predicted Predict the results for the equipment status.

[0105] Specifically, the device status may be an operating status of the power equipment such as faulty or normal, or may be operating data of another type / time of the power equipment predicted based on input time series data and image data.

[0106] Taking into account the dynamic changes in environmental conditions during actual power inspections, this embodiment introduces a dynamic error threshold adjustment mechanism. By calculating the mean and variance of the prediction error within a short time window, the error threshold is dynamically adjusted to more sensitively reflect data volatility and cope with complex power inspection environments. The dynamic error threshold is defined as T d (t)=μ(t)+k·σ ε (t). μ t Represents the overall error level of the system prediction within time t, σ t Indicates the degree of error fluctuation. In the case of large error fluctuations, the dynamic error threshold is correspondingly higher to avoid triggering unnecessary abnormal alarms due to accidental large errors. In the case of small error fluctuations, the dynamic error threshold is lower, thereby detecting abnormalities more sensitively. By setting this dynamic error threshold, it is determined whether there is a monitoring abnormal error. The monitoring abnormality flag is defined as: Among them, when Anomaly=1, it means monitoring is abnormal, and when Anomaly=0, it means monitoring is normal.

[0107] When the current device status prediction results detect anomalies, this embodiment proposes a dynamic adjustment mechanism based on an adaptive gradient-boosted Bayesian optimization algorithm. If an abnormal error is detected in the previous step, that is, when Anomaly = 1, this algorithm is applied to adjust the model parameters and generate the optimal hybrid intelligent model in real time. The specific operation is as follows:

[0108] A surrogate model is constructed using Gaussian process regression to simulate the distribution of the hyperparameter space.

[0109] The optimization process begins with a randomly initialized hyperparameter combination. Based on the surrogate model, an adaptive expected improvement strategy is introduced, incorporating gradient information to dynamically adjust the search direction to avoid falling into local optima. Combining dynamic perception of surrogate model uncertainty with gradient information, the expected improvement (AEI(X)) is calculated to select the next hyperparameter combination. The specific formula is:

[0110]

[0111] Where X = [x1, x2, …, xn] is the vector combination of hyperparameters, and σ(X) is the variance estimate of the surrogate model at the hyperparameter combination X, which represents the uncertainty of the model's prediction for this hyperparameter combination. It is the gradient modulus of the surrogate model at the hyperparameter combination X, indicating the intensity of change in the predicted value in the local area. It is used to balance the contribution of uncertainty and gradient information to the expected improvement value AEI.

[0112] Finally, the selected hyperparameter combination is evaluated and the surrogate model is updated. The above process is repeated until a satisfactory parameter combination is found.

[0113] The final parameters are applied to the hybrid intelligent model to ensure continuous optimization of the model performance to meet the real-time response and adjustment needs of the power inspection system.

[0114] When the prediction error of the hybrid intelligent model reaches a preset error value, or the number of iterations of parameter update reaches a preset number, the parameter update process of the hybrid intelligent model is stopped.

[0115] It can be seen from the above embodiments that in order to solve the limitations of the existing inspection system in data quality due to environmental factors, this embodiment proposes innovative multi-source data fusion and intelligent preprocessing technology to improve the clarity and consistency of inspection data and ensure the reliability and efficiency of the system in various complex environmental scenarios.

[0116] To address the problem of insufficient accuracy of fault diagnosis systems for power inspections in complex power equipment environments, this embodiment constructs a hybrid intelligent model that combines convolutional neural networks, long short-term memory networks, and temporal convolutional networks. This model comprehensively extracts the spatial and temporal features of multimodal data and concatenates all extracted features to form a joint feature vector. This feature vector integrates the spatial and temporal information of multi-source data and can output more accurate equipment status assessment and fault classification results. Through the application of this deep learning model, the system's fault diagnosis capabilities in complex scenarios are significantly improved, ensuring the accuracy of inspections.

[0117] In order to solve the problem of low real-time performance of the inspection and monitoring system, this embodiment establishes a dynamic feedback mechanism based on adaptive gradient enhanced Bayesian optimization. During the power inspection process, due to the rapid changes in equipment status and environmental conditions, the system needs to be able to respond quickly to reduce inspection errors. To this end, this embodiment designs a dynamic error threshold adjustment mechanism, which dynamically adjusts the error threshold by the mean and variance of the prediction error within a short time window to ensure that the system can sensitively monitor the impact of changes in environmental conditions on the equipment status; in addition, an adaptive Bayesian optimization algorithm is proposed. When the sensor detects an abnormal situation, the algorithm will adjust the model parameters in real time to generate the optimal model to ensure that the system can quickly respond to abnormal conditions in the power grid and continuously optimize performance, thereby improving the real-time and reliability of the power inspection process.

[0118] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0119] The following are device embodiments of the present invention. For details not fully described therein, reference may be made to the corresponding method embodiments described above.

[0120] Figure 3 The following is a schematic diagram of the structure of a deep learning-based inspection and optimization device for power equipment according to an embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown, which are described in detail as follows:

[0121] like Figure 3 As shown, the inspection optimization device 100 for power equipment based on deep learning includes:

[0122] The signal acquisition module 110 is used to acquire image signals and time sequence signals collected from the power equipment; the image signals include the appearance image and infrared thermal imaging data of the power equipment; the time sequence signals include the environmental data and electrical data of the power equipment;

[0123] a denoising module 120 for denoising the time series signal using an adaptive wavelet denoising method, denoising the image signal using a first convolutional neural network, and outputting high-level image features of the image signal;

[0124] An outlier detection module 130 is configured to perform outlier detection on the denoised time series signal and the high-level image features using an outlier detection algorithm based on dynamic density perception;

[0125] An outlier reconstruction module 140 is configured to reconstruct outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain a corrected time series signal and image signal;

[0126] The device state prediction module 150 is used to input the corrected timing signal and image signal into the hybrid intelligent model and output the device state prediction result of the power equipment. The hybrid intelligent model is constructed based on a deep learning algorithm.

[0127] In one possible implementation, the denoising module 120 includes:

[0128] Performing Fourier transform on the time series signal to obtain a frequency spectrum of the time series signal;

[0129] According to the formula Determining the number of layers of wavelet decomposition, and performing wavelet decomposition on the time series signal based on the layers to obtain an approximate part and a detail part;

[0130] Using soft thresholding Eliminating high-frequency noise in the detail part, performing wavelet reconstruction on the detail part and the approximate part after the high-frequency noise is eliminated, and converting the spectrum after wavelet reconstruction into the time domain to obtain a denoised time series signal;

[0131] Among them, L represents the level of wavelet decomposition, T represents the signal sampling length, k represents the compensation coefficient, round() represents the rounding function, and f main represents the main frequency component of the spectrum, α represents the first preset value, σ f represents the standard deviation of the spectrum; D i ' represents the soft threshold, D i represents the i-th component of the detail part, σ represents the noise standard deviation, and N represents the signal length of the time series signal.

[0132] In one possible implementation, the loss function of the first convolutional neural network is:

[0133]

[0134] Among them, β1 and β2 represent weight parameters, Y i Represents the i-th image pixel value in the original image, Y i ' represents the pixel value of the i-th image after denoising; φ j Represents the j-th layer feature extraction result in the VGG network, n represents the number of image pixels, and m represents the number of feature channels.

[0135] In one possible implementation, the outlier detection module 130 includes:

[0136] Determine the local density of each time series data point in the denoised time series signal;

[0137] Based on the formula Identify outliers in the denoised time series signal;

[0138] Among them, Anomaly(x i ) represents the abnormal sign of the i-th time series data point. When Anomaly(x i )=1, it means that the i-th time series data point is an abnormal value; when Anomaly(x i )=0, it means that the i-th time series data point is a normal value; ρ(x i ) represents the local density of the i-th time series data point;

[0139]

[0140] Among them, θ(x i ) represents the dynamic density threshold of the i-th time series data point, β3 and γ represent weights, and ε(x i) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th time series data point, δ(x i ) represents the density change factor of the i-th time series data point; d(x i ,x j ) represents the Euclidean distance from the i-th time series data point to the j-th time series data point; h-NN(x i ) represents the h nearest neighbors of the i-th time series data point, α1 represents the second preset value, and median() represents the median function.

[0141] In one possible implementation, the outlier detection module 130 further includes:

[0142] Performing dimensionality reduction processing on the high-level image features to obtain an image signal after dimensionality reduction processing;

[0143] Based on the formula determining outliers in the denoised image signal;

[0144] Among them, Anomaly(y i ) represents the abnormal sign of the i-th image data point. When Anomaly(y i )=1, it means that the i-th image data point is an abnormal value; when Anomaly(y i )=0, it means that the i-th image data point is a normal value; ρ(y i ) represents the local density of the i-th image data point;

[0145]

[0146] Among them, θ(y i ) represents the dynamic density threshold of the i-th image data point, α2, α3, α4 represent weights, ε(y i ) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th image data point, δ(y i ) represents the density change factor of the i-th image data point; d(y i ,y j ) represents the Euclidean distance from the i-th image data point to the j-th image data point; h-NN(x i ) represents the h nearest neighbors of the i-th image data point, α1 represents the second preset value, and median() represents the median function; μ i represents the mean of the local area where the i-th image data point is located, σ i represents the variance of the local area where the i-th image data point is located; median() represents the median function, and S represents the total number of image feature extraction layers under multi-scale.

[0147] In one possible implementation, the outlier reconstruction module 140 includes:

[0148] Obtaining an average density of all image data points in the image signal, and using the average density of the image data points as an image density threshold;

[0149] dividing the outliers in the image data points into isolated outliers and intra-cluster outliers according to the image density threshold;

[0150] For each isolated outlier in the image data points, a local polynomial fitting algorithm is used to fit the image data points in the neighborhood of the isolated outlier to obtain a first local interpolation model;

[0151] Reconstructing the isolated outlier using the first local interpolation model;

[0152] For each in-cluster outlier in the image data point, the formula is used Reconstruct the outliers in the cluster;

[0153] Among them, y i ' represents the i-th image data point after reconstruction, y i represents the abnormal point in the i-th cluster, α5 represents the smoothing factor, represents the gradient operator, ρ i represents the density mean of the neighborhood data points of the outlier in the i-th cluster, ρ threshold-1 Indicates the image density threshold.

[0154] In one possible implementation, the outlier reconstruction module 140 further includes:

[0155] Obtaining an average density of all time series data points in the time series signal, and using the average density of the time series data points as an image density threshold;

[0156] dividing the outliers in the time series data points into isolated outliers and intra-cluster outliers according to the image density threshold;

[0157] For each isolated outlier in the time series data points, a local polynomial fitting algorithm is used to fit the time series data points in the neighborhood of the isolated outlier to obtain a second local interpolation model;

[0158] Reconstructing the isolated outlier using the second local interpolation model;

[0159] For each in-cluster outlier in the time series data point, the formula x i '=x i -α6·(ρ i -ρ threshold-2 ) reconstruct the outliers in the cluster;

[0160] Among them, x i ' represents the i-th time series data point after reconstruction, x i represents the outlier in the i-th cluster, α6 represents the smoothing factor, and ρ i represents the density mean of the outlier neighborhood data points of the outlier in the i-th cluster, ρ threshold-2 Indicates the time series density threshold.

[0161] In one possible embodiment, the hybrid intelligent model includes a convolutional neural network module, a long short-term memory network module, a temporal convolutional network module, and a feature fusion module; the device state prediction module 150 includes:

[0162] The corrected time series signal is input into the long short-term memory network module, and the comprehensive time series feature vector is output;

[0163] The corrected image signal is input into the convolutional neural network module, which outputs a high-level image feature vector;

[0164] Inputting the comprehensive time series feature vector into the temporal convolutional network module and outputting a high-level time series feature vector;

[0165] The high-level image feature vector and the high-level time series feature vector are input into the feature fusion module, and the device state prediction result of the power device is output.

[0166] In one possible implementation, the deep learning-based inspection optimization device 100 for power equipment further includes a parameter training module for:

[0167] Acquiring an actual device state of the electric device, and determining a current prediction error based on the actual device state of the electric device and a device state prediction result;

[0168] According to the formula determining a dynamic error threshold, and determining that a detection anomaly exists in the current device state prediction result when the absolute value of the current prediction error is greater than the dynamic error threshold;

[0169] When there is a detection anomaly in the current device state prediction result, the parameters in the hybrid intelligent model are updated based on the adaptive gradient boosting Bayesian optimization algorithm;

[0170] Among them, T d (t) represents the dynamic error threshold, μ(t) represents the mean value of the prediction error in the time window t, and i Represents the prediction error for time i, t represents the length of the time window, σ t represents the standard deviation of the prediction error within the time window t; η represents the first preset coefficient.

[0171] Figure 4 Schematic diagram of a terminal provided by an embodiment of the present invention. Figure 4 As shown, the terminal 4 of this embodiment includes: a processor 40 and a memory 41. The memory 41 is used to store a computer program 42, and the processor 40 is used to call and run the computer program 42 stored in the memory 41 to perform the steps in the above-mentioned embodiments of the inspection optimization method for power equipment based on deep learning, such as Figure 1 Alternatively, the processor 40 is used to call and run the computer program 42 stored in the memory 41 to implement the functions of each module / unit in the above-mentioned device embodiments, such as Figure 3 The functions of the modules 110 to 150 are shown.

[0172] Exemplarily, the computer program 42 may be divided into one or more modules / units, which are stored in the memory 41 and executed by the processor 40 to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program 42 in the terminal 4. For example, the computer program 42 may be divided into Figure 4 Modules 110 to 150 are shown.

[0173] The terminal 4 can be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal 4 can include, but is not limited to, a processor 40 and a memory 41. Those skilled in the art will understand that Figure 4 It is only an example of terminal 4 and does not constitute a limitation on terminal 4. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal may also include input and output devices, network access devices, buses, etc.

[0174] The processor 40 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0175] The memory 41 may be an internal storage unit of the terminal 4, such as a hard disk or memory of the terminal 4. The memory 41 may also be an external storage device of the terminal 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, etc. equipped on the terminal 4. Furthermore, the memory 41 may include both an internal storage unit of the terminal 4 and an external storage device. The memory 41 is used to store the computer program and other programs and data required by the terminal. The memory 41 may also be used to temporarily store data that has been output or is about to be output.

[0176] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0177] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0178] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0179] In the embodiments provided by the present invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms.

[0180] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0181] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0182] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, it can implement the steps of the above-mentioned embodiments of the inspection optimization method for power equipment based on deep learning. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable media does not include electrical carrier signals and telecommunication signals.

[0183] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A patrol optimization method for power equipment based on deep learning, characterized in that: include: Acquire image signals and time series signals collected from power equipment; the image signal includes an image of the equipment appearance and infrared thermal imaging data of the power equipment; the time series signal includes environmental data and electrical data of the power equipment; use an adaptive wavelet denoising method to denoise the time series signal, use a first convolutional neural network to denoise the image signal, and output high-level image features of the image signal; use an outlier detection algorithm based on dynamic density perception to perform outlier detection on the denoised time series signal and the high-level image features respectively; reconstruct the outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain corrected time series signals and image signals; input the corrected time series signals and image signals into a hybrid intelligent model, and output the equipment status prediction result of the power equipment, wherein the hybrid intelligent model is constructed based on a deep learning algorithm.

2. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: The denoising process of the time series signal using the adaptive wavelet denoising method includes: Performing Fourier transform on the time series signal to obtain a frequency spectrum of the time series signal; According to the formula Determining the number of layers of wavelet decomposition, and performing wavelet decomposition on the time series signal based on the layers to obtain an approximate part and a detail part; Using soft thresholding Eliminating high-frequency noise in the detail part, performing wavelet reconstruction on the detail part and the approximate part after the high-frequency noise is eliminated, and converting the spectrum after wavelet reconstruction into the time domain to obtain a denoised time series signal; Among them, L represents the level of wavelet decomposition, T represents the signal sampling length, k represents the compensation coefficient, round() represents the rounding function, and f main represents the main frequency component of the spectrum, α represents the first preset value, σ f represents the standard deviation of the spectrum; D i ' represents the soft threshold, D i represents the i-th component of the detail part, σ represents the noise standard deviation, and N represents the signal length of the time series signal.

3. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: The loss function of the first convolutional neural network is: Among them, β1 and β2 represent weight parameters, Y i Represents the original i-th image pixel value, Y i ' represents the pixel value of the i-th image after denoising; φ j Represents the j-th layer feature extraction result in the VGG network, n represents the number of image pixels, and m represents the number of feature channels.

4. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: An outlier detection algorithm based on dynamic density perception is used to detect outliers in the denoised time series signal, including: Determine the local density of each time series data point in the denoised time series signal; Based on the formula Identify outliers in the denoised time series signal; Among them, Anomaly(x i ) represents the abnormal sign of the i-th time series data point. When Anomaly(x i )=1, it means that the i-th time series data point is an abnormal value; when Anomaly(x i )=0, it means that the i-th time series data point is a normal value; ρ(x i ) represents the local density of the i-th time series data point; Among them, θ(x i ) represents the dynamic density threshold of the i-th time series data point, β3 and γ represent weights, and ε(x i ) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th time series data point, δ(x i ) represents the density change factor of the i-th time series data point; d(x i ,x j ) represents the Euclidean distance from the i-th time series data point to the j-th time series data point; h-NN(x i ) represents the h nearest neighbors of the i-th time series data point, α1 represents the second preset value, and median() represents the median function.

5. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: An outlier detection algorithm based on dynamic density perception is used to detect outliers on the high-level image features, including: Performing dimensionality reduction processing on the high-level image features to obtain an image signal after dimensionality reduction processing; Based on the formula determining outliers in the denoised image signal; Among them, Anomaly(y i ) represents the abnormal sign of the i-th image data point. When Anomaly(y i )=1, it means that the i-th image data point is an abnormal value; when Anomaly(y i )=0, it means that the i-th image data point is a normal value; ρ(y i ) represents the local density of the i-th image data point; Among them, θ(y i ) represents the dynamic density threshold of the i-th image data point, α2, α3, α4 represent weights, ε(y i ) represents the adaptive neighborhood radius of the jth nearest neighbor of the i-th image data point, δ(y i ) represents the density change factor of the i-th image data point; d(y i ,y j ) represents the Euclidean distance from the i-th image data point to the j-th image data point; h-NN(x i ) represents the h nearest neighbors of the i-th image data point, α1 represents the second preset value, and median() represents the median function; μ i represents the mean of the local area where the i-th image data point is located, σ i represents the variance of the local area where the i-th image data point is located; median() represents the median function, and S represents the total number of image feature extraction layers under multi-scale.

6. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: The method of reconstructing the outliers in the image signal based on the classification outlier reconstruction algorithm to obtain a corrected image signal includes: Obtaining an average density of all image data points in the image signal, and using the average density of the image data points as an image density threshold; dividing the outliers in the image data points into isolated outliers and intra-cluster outliers according to the image density threshold; For each isolated outlier in the image data points, a local polynomial fitting algorithm is used to fit the image data points in the neighborhood of the isolated outlier to obtain a first local interpolation model; Reconstructing the isolated outlier using the first local interpolation model; For each in-cluster outlier in the image data point, the formula is used Reconstruct the outliers in the cluster; Among them, y i ' represents the i-th image data point after reconstruction, y i represents the abnormal point in the i-th cluster, α5 represents the smoothing factor, represents the gradient operator, ρ i represents the density mean of the neighborhood data points of the outlier in the i-th cluster, ρ threshold-1 Indicates the image density threshold.

7. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: The method of reconstructing the outliers in the time series signal based on the classification outlier reconstruction algorithm to obtain a corrected time series signal includes: Obtaining an average density of all time series data points in the time series signal, and using the average density of the time series data points as an image density threshold; dividing the outliers in the time series data points into isolated outliers and intra-cluster outliers according to the image density threshold; For each isolated outlier in the time series data points, a local polynomial fitting algorithm is used to fit the time series data points in the neighborhood of the isolated outlier to obtain a second local interpolation model; Reconstructing the isolated outlier using the second local interpolation model; For each in-cluster outlier in the time series data point, the formula x i '=x i -α6·(ρ i -ρ threshold-2 ) reconstruct the outliers in the cluster; Among them, x i ' represents the i-th time series data point after reconstruction, x i represents the outlier in the i-th cluster, α6 represents the smoothing factor, and ρ i represents the density mean of the outlier neighborhood data points of the outlier in the i-th cluster, ρ threshold-2 Indicates the time series density threshold.

8. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: The hybrid intelligent model includes a convolutional neural network module, a long short-term memory network module, a temporal convolutional network module and a feature fusion module; The step of inputting the corrected time series signal and image signal into the hybrid intelligent model and outputting the equipment state prediction result of the power equipment includes: The corrected time series signal is input into the long short-term memory network module, and the comprehensive time series feature vector is output; The corrected image signal is input into the convolutional neural network module, which outputs a high-level image feature vector; Inputting the comprehensive time series feature vector into the temporal convolutional network module and outputting a high-level time series feature vector; The high-level image feature vector and the high-level time series feature vector are input into the feature fusion module, and the device state prediction result of the power device is output.

9. The inspection optimization method for power equipment based on deep learning according to claim 1, characterized in that: After inputting the corrected time series signal and image signal into the hybrid intelligent model and outputting the device state prediction result of the power device, the method further includes: Acquiring an actual device state of the electric device, and determining a current prediction error based on the actual device state of the electric device and a device state prediction result; According to the formula determining a dynamic error threshold, and determining that a detection anomaly exists in the current device state prediction result when the absolute value of the current prediction error is greater than the dynamic error threshold; When there is a detection anomaly in the current device state prediction result, the parameters in the hybrid intelligent model are updated based on the adaptive gradient boosting Bayesian optimization algorithm; Among them, T d (t) represents the dynamic error threshold, μ(t) represents the mean value of the prediction error in the time window t, and i Represents the prediction error for time i, t represents the length of the time window, σ t represents the standard deviation of the prediction error within the time window t; η represents the first preset coefficient.

10. A patrol optimization device for power equipment based on deep learning, characterized in that: include: A signal acquisition module is used to acquire image signals and time sequence signals collected from the power equipment; the image signals include the appearance image and infrared thermal imaging data of the power equipment; the time sequence signals include the environmental data and electrical data of the power equipment; a denoising module, configured to perform denoising processing on the time series signal using an adaptive wavelet denoising method, perform denoising processing on the image signal using a first convolutional neural network, and output high-level image features of the image signal; An outlier detection module, configured to perform outlier detection on the denoised time series signal and the high-level image features using an outlier detection algorithm based on dynamic density perception; an outlier reconstruction module, configured to reconstruct outliers in the image signal and the time series signal based on a classified outlier reconstruction algorithm to obtain a corrected time series signal and image signal; The device status prediction module is used to input the corrected timing signal and image signal into the hybrid intelligent model and output the device status prediction result of the power equipment. The hybrid intelligent model is constructed based on the deep learning algorithm.

Images