Transformer working condition identification method and system based on multi-point vibration and improved MobileNet

Abstract

The invention discloses a transformer working condition identification method and system based on multi-point vibration and improved MobileNet, and belongs to the technical field of transformer detection, and the method comprises the steps: applying synchronous vibration excitation to a plurality of preset point positions of a transformer, and collecting a multi-dimensional vibration sequence; extracting frequency band energy characteristics of the multi-dimensional vibration sequence, and constructing a three-dimensional space-time matrix; inputting the three-dimensional space-time matrix into an improved MobileNetV3 network to obtain transformer working condition probability distribution; outputting a transformer working condition identification result according to the working condition probability distribution; wherein the MobileNetV3 network is improved by inserting a mixed cavity convolution block and a three-dimensional pooling layer and introducing a convolution block attention module, and the problem that a traditional single-measuring-point vibration analysis method is insufficient in sensitivity under complex working conditions is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and particularly to a method and system for identifying transformer operating conditions based on multi-point vibration and improved MobileNet. Background Art

[0002] In recent years, in the current industrial monitoring field, vibration analysis, as an important non-destructive testing technology, has been widely used in the fault diagnosis and health management of mechanical equipment. The traditional single-point vibration analysis method mainly installs sensors at specific positions of the equipment to collect the vibration signals at these points, and then analyzes the operating state of the equipment. This method relies on means such as spectrum analysis and time-domain analysis of vibration signals, and can identify abnormal vibration modes of the equipment, such as fault characteristics like imbalance, looseness, and wear. The single-point vibration analysis technology has the advantages of relatively low cost and simple implementation, so it has been widely used in many simple or standardized equipment monitoring tasks.

[0003] However, with the increasing complexity of modern industrial equipment, the traditional single-point vibration analysis method has gradually exposed the defect of insufficient sensitivity. Mechanical equipment under complex working conditions often has multiple vibration sources and complex vibration transmission paths, and the data collected by a single point often cannot comprehensively reflect the overall vibration state of the equipment. In addition, due to factors such as environmental noise, sensor accuracy limitations, and attenuation during signal transmission, single-point data is easily interfered with, resulting in inaccurate analysis results. Especially in the early fault stage, the vibration signals are weak and complex, and single-point analysis often fails to capture these subtle vibration changes, thus delaying the timely discovery and handling of faults. Therefore, the application of the traditional single-point vibration analysis method under complex working conditions has been severely restricted, and there is an urgent need for a more sensitive and comprehensive vibration analysis technology to meet the monitoring requirements of modern industrial equipment. Summary of the Invention

[0004] The technical problem solved by the present invention is: how to solve the problem of insufficient sensitivity of the traditional single-point vibration analysis method under complex working conditions.

[0005] To solve the above technical problem, the present invention provides the following technical solutions:

[0006] In the first aspect, the present invention provides a method for identifying transformer operating conditions based on multi-point vibration and improved MobileNet, including:

[0007] Applying synchronous vibration excitation to multiple preset points of the transformer and collecting multi-dimensional vibration sequences;

[0008] Extracting the band energy characteristics of the multi-dimensional vibration sequences and constructing a three-dimensional spatio-temporal matrix;

[0009] Input the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network to obtain the probability distribution of the transformer operating conditions;

[0010] Output the recognition result of the transformer operating conditions according to the probability distribution of the operating conditions;

[0011] Among them, the MobileNetV3 network is improved by inserting hybrid dilated convolution blocks and three-dimensional pooling layers and introducing a convolutional block attention module.

[0012] Furthermore, apply synchronous vibration excitation to multiple preset points of the transformer and collect multi-dimensional vibration sequences, including:

[0013] Arrange multiple groups of piezoelectric and inertial sensor arrays on the side wall of the transformer oil tank and the radiator area;

[0014] Apply a mixed excitation signal of pulse excitation and sweep excitation with a frequency band in the range of 0 - 5 kHz and a frequency increasing exponentially from 10 Hz to 10 kHz through a multi-channel synchronous signal generator to form a multi-physical field coupled vibration environment;

[0015] Use a distributed acquisition system to synchronously obtain the six-axis vibration signals of each sensor node in the sensor array and construct a multi-dimensional vibration sequence including displacement, velocity, and acceleration.

[0016] Furthermore, adopt wavelet packet transform to extract the frequency band energy characteristics of the multi-dimensional vibration sequence and construct a three-dimensional spatio-temporal matrix, including:

[0017] Use the improved variational mode decomposition algorithm to process the multi-dimensional vibration sequence and separate each modal component;

[0018] Adopt wavelet packet transform to divide the frequency bands of each modal component and calculate the energy entropy within each frequency band as the feature vector;

[0019] Based on the tensor decomposition theory, reorganize the feature vectors in the three-dimensional coordinate system of "number of sensors × time window length × frequency band energy" to construct a three-dimensional spatio-temporal matrix.

[0020] Furthermore, introduce a modal number adaptive selection mechanism, and determine the optimal modal number of the variational mode decomposition algorithm based on the complexity of the practical signal of the multi-dimensional vibration sequence through the Bayesian information criterion to improve the variational mode decomposition algorithm, and obtain the improved variational mode decomposition algorithm.

[0021] Furthermore, determine the time window length through the dynamic programming algorithm, including:

[0022] Define the window length at the current moment as the state variable;

[0023] Define the window length increment at the current moment as a decision variable, where the window length increment represents the change in the window length from the current moment to the next moment;

[0024] Based on the state variable, decision variable, and the stationarity and feature separability of the multi-dimensional vibration sequence within the window, establish a state transition equation to describe the transition from the current state variable and current decision variable to the next state variable;

[0025] Define a cost function according to the computational complexity of the state variable and the loss of feature separability to evaluate the quality of each state variable;

[0026] Construct a Bellman equation based on the cost function, state transition equation, and discount factor to solve the optimal decision sequence;

[0027] Starting from the final moment, calculate the optimal decision and optimal cost of each state variable in reverse, and iteratively solve the Bellman equation backward to obtain the optimal decision sequence;

[0028] Based on the optimal decision sequence, calculate the actual window length sequence forward and determine the optimal time window length as the time window length.

[0029] Furthermore, the state transition equation is expressed as:

[0030]

[0031] In the formula, L t+1 represents the change in the window length at the next moment t + 1, L t+1 = L t + ΔL t ∈[L min , L max , L t represents the window length at the current moment, L t ∈[L min , L max , t represents the current moment, L min represents the preset minimum window, L max represents the preset maximum window length, ΔL t represents the window length increment at the current moment, represents the variance of the multi-dimensional vibration sequence within the window, represents the preset variance threshold, J t represents the ratio of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the window, J th represents the preset ratio threshold of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the window, if represents "if", and represents "and", otherwise represents "otherwise";

[0032] The cost function is expressed as:

[0033] C(L t ,ΔL t )=α·Comp(L t )+β·Loss(L t );

[0034] In the formula, C(L t ,ΔL t ) represents the cost function, α represents the weight coefficient of the computational complexity Comp(L t ) at the current moment when the window length is L t ), β represents the weight coefficient of the feature separability loss Loss(L t ) at the current moment when the window length is L t ).

[0035] Furthermore, the Bellman equation is expressed as:

[0036]

[0037] In the formula, V(L t ) represents the optimal cost function value at the current moment when the window length is L t , V(L t+1 ) represents the optimal cost function value at the current moment when the window length is L t+1 , γ represents the discount factor, C(L t ,ΔL t ) represents the cost function, represents taking the minimum value of the window length increment ΔL t at the current moment, and t represents the current moment.

[0038] Furthermore, the MobileNetV3 network is improved by inserting hybrid dilated convolution blocks and 3D pooling layers and introducing a convolutional block attention module, including:

[0039] Insert hybrid dilated convolution blocks after the initial convolutional layer, and cycle through dilation rate configurations using convolutional kernels with dilation rates of 1, 2, and 3 to maintain the number of parameters and expand the receptive field to the 7×7 equivalent range;

[0040] Insert a 3D pooling layer after the depthwise separable convolutional layer to perform downsampling only along the time dimension and retain the original spatial dimension resolution;

[0041] Introduce a convolutional block attention module in the bottleneck layer to perform dual-channel and spatial attention weighting on the spatio-temporal feature map.

[0042] Further, it also includes connecting a fully connected layer after the improved MobileNetV3 network, using the fully connected layer to classify the operating conditions of the high-dimensional features output by the improved MobileNetV3 network, and outputting the operating condition probability distribution through the Softmax function.

[0043] In a second aspect, the present invention provides a transformer operating condition recognition system based on multi-point vibration and improved MobileNet, including the following modules:

[0044] A vibration excitation module for applying synchronous vibration excitation to multiple preset points of the transformer;

[0045] A signal acquisition module for acquiring multi-dimensional vibration sequences;

[0046] A feature extraction module for extracting the band energy features of the multi-dimensional vibration sequence and constructing a three-dimensional spatio-temporal matrix;

[0047] A network inference module for inputting the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network to obtain the transformer operating condition probability distribution;

[0048] A decision output module for outputting the transformer operating condition recognition result according to the operating condition probability distribution;

[0049] Among them, the MobileNetV3 network is improved by inserting a hybrid dilated convolution block and a three-dimensional pooling layer and introducing a convolutional block attention module.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. The present invention collects multi-dimensional vibration sequences through synchronous vibration excitation, constructs a three-dimensional spatio-temporal matrix based on the band energy features, can efficiently extract key information from complex vibration signals, improves the MobileNetV3 network by inserting a hybrid dilated convolution block and a three-dimensional pooling layer, enhances the learning ability of the MobileNetV3 network for spatio-temporal features, and at the same time introduces a convolutional block attention module to further improve the accuracy and robustness of feature expression. It not only improves the accuracy and reliability of transformer operating condition recognition, but also has stronger generalization ability and higher recognition efficiency compared with traditional methods, providing strong technical support for the state monitoring and fault diagnosis of transformers. The present invention solves the problem of insufficient sensitivity of traditional single-point vibration analysis methods under complex operating conditions.

[0052] 2. The present invention arranges multiple groups of sensor arrays on the side wall of the transformer oil tank and in the radiator area, and applies a hybrid excitation signal to form a multi-physical-field coupled vibration environment, which can comprehensively capture the complex vibration characteristics of the transformer under different working conditions. By using a distributed acquisition system to synchronously obtain multi-dimensional vibration sequences, combining wavelet packet transform and variational mode decomposition algorithms to extract frequency band energy characteristics, and constructing a three-dimensional spatio-temporal matrix, the accuracy and reliability of working condition identification are significantly improved. This method that combines multi-physical-field information and advanced signal processing technology provides strong support for the accurate identification of transformer working conditions.

[0053] 3. The present invention realizes the adaptive extraction and optimization of multi-dimensional vibration sequence features by introducing a modal number adaptive selection mechanism and a dynamic programming algorithm to determine the optimal time window length. It also determines the optimal modal number of the variational mode decomposition algorithm through the Bayesian information criterion, improving the accuracy and efficiency of feature extraction. At the same time, the dynamic programming algorithm reversely iteratively solves the Bellman equation according to the stationarity and feature separability of the multi-dimensional vibration sequence within the window to obtain the optimal decision sequence, thereby determining the optimal time window length, enabling the present invention to maintain high recognition performance under different working conditions and environments.

[0054] 4. The present invention expands the receptive field through a hybrid dilated convolution block, improving the ability to capture global information; uses a three-dimensional pooling layer for downsampling along the time dimension while retaining the original resolution in the spatial dimension, which helps to maintain the spatial structure information of the features; uses a convolutional block attention module to perform dual channel and spatial attention weighting on the spatio-temporal feature map, further enhancing the feature expression ability. The improved MobileNetV3 network not only improves the efficiency of working condition identification but also enhances the robustness of the model to complex environments and noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a schematic diagram of the basic process of a transformer working condition identification method based on multi-point vibration and improved MobileNet provided by an embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of a transformer working condition identification system based on multi-point vibration and improved MobileNet provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention is provided in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0058] Embodiment 1

[0059] As Figure 1An embodiment of the present invention is shown. From a principle perspective, this embodiment introduces a transformer condition recognition method based on multi-point vibration and improved MobileNet, including:

[0060] Step 1: Apply synchronous vibration excitation to multiple preset points of the transformer and collect multi-dimensional vibration sequences.

[0061] In the present invention, an external excitation device is used to apply synchronous vibration signals to preset points of the transformer, such as the oil tank, winding, iron core, etc. A multi-dimensional vibration response sequence is collected by using a vibration sensor array. There is a strong coupling relationship between the multi-dimensional vibration response sequence and the internal mechanical state of the transformer, such as winding tightness, iron core looseness, insulation deterioration, etc., and it is not affected by electromagnetic interference.

[0062] The external excitation can amplify the weak vibration signals caused by potential faults. The synchronously collected multi-dimensional vibration sequences contain spatial distribution characteristics, providing rich information for condition recognition, avoiding the high risk and limitations of traditional electrical monitoring methods. The multi-dimensional vibration sequences collected in the present invention directly reflect the internal mechanical state of the transformer, providing basic data for subsequent feature extraction and enhancing the reliability and safety of monitoring.

[0063] Step 2: Extract the band energy characteristics of the multi-dimensional vibration sequence and construct a three-dimensional spatio-temporal matrix.

[0064] In the present invention, by extracting the band energy characteristics of the multi-dimensional vibration sequence and analyzing the energy distribution in different frequency bands, such as the low-frequency band below 100Hz reflects mechanical looseness, and the high-frequency band above 1kHz corresponds to partial discharge, highlighting the fault characteristics.

[0065] Reorganize the band energy characteristics of multiple sensors and multiple time periods according to the "time × space × frequency band" dimension to form a three-dimensional spatio-temporal matrix. The constructed three-dimensional spatio-temporal matrix integrates multi-sensor data, retains spatio-temporal correlation information, enhances the feature expression ability, improves the identification ability of complex working conditions, and provides a structured input for the improved MobileNetV3 network.

[0066] Step 3: Improve the MobileNetV3 network by inserting hybrid dilated convolution blocks and three-dimensional pooling layers and introducing a convolutional block attention module.

[0067] In the present invention, through hybrid dilated convolution blocks, convolutional kernels with different dilation rates are connected in parallel. On the premise of keeping the number of parameters unchanged, the receptive field is expanded to capture multi-scale spatio-temporal features.

[0068] The present invention adopts adaptive three-dimensional average pooling to replace traditional two-dimensional pooling, directly downsampling the three-dimensional spatio-temporal matrix, avoiding information loss. At the same time, through the three-dimensional pooling layer, it can also adapt to the input of the three-dimensional spatio-temporal matrix, reducing the dimension and retaining key information.

[0069] Through the channel attention and spatial attention mechanisms of the convolutional block attention module, the present invention adaptively enhances important features. Among them, the channel attention generates channel weights through global average / max pooling to strengthen the frequency band features strongly related to the working conditions. The spatial attention uses a 7×7 convolution to generate a spatial weight map, focusing on the key sensor positions and time windows, improving the extraction efficiency of the improved MobileNetV3 network for spatio-temporal features. The dilated convolution and the attention mechanism work together to enhance the identification accuracy for complex working conditions and are applicable to resource-constrained industrial scenarios.

[0070] Step 4: Input the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network to obtain the probability distribution of transformer working conditions.

[0071] The present invention inputs the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network. After passing through the global average pooling layer and then connecting to a fully connected layer, finally, the probability distributions of various working conditions of the transformer, such as normal, partial discharge, winding deformation, etc., are output through the Softmax function. The probability distribution of transformer working conditions reflects the confidence of the improved MobileNetV3 network in judging transformer working conditions, facilitating subsequent decision-making. The Softmax layer supports multi-classification tasks, adapts to the diversity of transformer working conditions, outputs the probability distributions of various working conditions through the Softmax layer, quantifies the uncertainty of transformer working conditions, provides a basis for decision-making, and supports real-time status monitoring and early warning.

[0072] Step 5: Output the recognition result of the transformer working condition according to the probability distribution of the working condition.

[0073] Based on the probability distribution, the present invention selects the working condition category with the highest probability as the recognition result, which can be directly applied to equipment maintenance decision-making, such as fault early warning, load adjustment, etc., improving the operation and maintenance efficiency and power grid stability.

[0074] The present invention adopts the "maximum probability principle" to select the working condition category corresponding to the highest value in the probability distribution as the recognition result. For example, if the probability of the "winding deformation" category reaches 85%, it is determined that the current working condition is winding deformation. Combining with a threshold can achieve early fault warning, such as triggering an alarm when the probability exceeds 70%.

[0075] Through the combination of vibration monitoring and deep learning, the present invention realizes non-invasive and high-precision identification of transformer working conditions, with both technological innovation and engineering practicability.

[0076] Embodiment 2

[0077] With the same inventive concept as in Embodiment 1, this embodiment introduces the implementation steps of the transformer working condition identification method based on multi-point vibration and improved MobileNet, including:

[0078] Step 1: Apply synchronous vibration excitation to multiple preset points of the transformer and collect multi-dimensional vibration sequences.

[0079] In some embodiments, applying synchronous vibration excitation to multiple preset points of the transformer and collecting multi-dimensional vibration sequences includes:

[0080] Step 1.1: Arrange multiple groups of piezoelectric and inertial sensor arrays on the side wall of the transformer oil tank and the radiator area.

[0081] In this embodiment, different types of sensors are used to monitor the vibration characteristics of the transformer. Among them, piezoelectric sensors work based on the piezoelectric effect, can directly convert mechanical vibration into charge signals, and are particularly sensitive to high-frequency vibrations above 1 kHz. Therefore, this type of sensor is very suitable for monitoring transient shocks caused by partial discharges such as inter-turn short circuits in windings. Inertial sensors, on the other hand, use MEMS accelerometers to evaluate vibration acceleration by measuring the displacement of the mass block, are more sensitive to low-frequency vibrations below 500 Hz, and are suitable for capturing periodic vibrations caused by magnetostriction of the iron core.

[0082] In this embodiment, the piezoelectric sensors and the inertial sensor arrays jointly cover a frequency band range of 0 - 5 kHz, enabling the piezoelectric sensors and the inertial sensors to match the different fault characteristics of the low-frequency vibrations of the transformer windings and the high-frequency vibrations of the insulation structure.

[0083] In addition, this embodiment also optimizes the design of the radiator area and uses it as the main channel for oil flow. The vibration signals of the radiator contain oil flow pulsation information and local temperature gradient information. By reasonably arranging the sensors, the distribution of the hot spots in the windings is indirectly monitored. To achieve three-dimensional spatial positioning of the vibration source, multiple groups of sensor arrays are arranged on the side wall and the radiator, and the positioning error is controlled within 5 cm.

[0084] Step 1.2: Apply a mixed excitation signal of pulse excitation and swept-frequency excitation with a frequency band in the range of 0 - 5 kHz and a frequency increasing exponentially from 10 Hz to 10 kHz through a multi-channel synchronous signal generator to form a multi-physical-field coupled vibration environment.

[0085] In this embodiment, the pulse excitation uses a Gaussian-modulated pulse train with an adjustable center frequency. For example, a 5 MHz pulse is used to excite the resonance of the winding. At the same time, the frequency of the swept-frequency excitation increases gradually from 10 Hz to 10 kHz according to an exponential law, and the growth factor for each step is 1.2. Such a swept-frequency range covers the inherent frequency range of the transformer, which is mainly distributed between 800 Hz and 3.5 kHz. To simulate the magnetic-mechanical coupling effect during the actual operation of the transformer, an electromagnetic bias magnetic field of 0.1 T is applied synchronously.

[0086] In this embodiment, the pulse excitation is mainly used to stimulate the transient signal of partial discharge, while the swept-frequency excitation is used to amplify the harmonic components generated by winding looseness. By comparing the correlation between the excitation and the response in the frequency domain, this embodiment can identify the phenomenon of core looseness caused by frequency response deviation and the winding deformation problem caused by the mutation of harmonic components. Step 1.3: Use the distributed acquisition system to synchronously obtain the six-axis vibration signals of each sensor node in the sensor array, and construct a multi-dimensional vibration sequence including displacement, velocity, and acceleration.

[0087] This embodiment adopts a distributed acquisition system, which realizes nanosecond-level clock synchronization through the Precision Time Protocol technology and supports expansion to 100 or more sensor nodes. The system uses a capacitive sensor with a resolution of 0.1 micrometer to synchronously collect three-axis displacement data. At the same time, three-axis acceleration information is collected through a MEMS sensor with a range of ±50g. Among them, the sampling rate is not less than 50kHz. To optimize data transmission, the system adopts the time-division multiplexing TDM technology to encapsulate multi-node data into a unified frame structure, thereby reducing the transmission bandwidth requirement and achieving a data compression ratio of 12:1. In addition, the system supports the vibration mode decomposition of six-axis data, can distinguish the translational vibration caused by the overall displacement of the oil tank from the torsional vibration caused by winding deformation, and calculates the velocity / displacement through acceleration integration to realize real-time quantification of vibration energy.

[0088] Step 2: Extract the frequency band energy characteristics of the multi-dimensional vibration sequence and construct a three-dimensional spatio-temporal matrix.

[0089] In some embodiments, wavelet packet transform is used to extract the frequency band energy characteristics of the multi-dimensional vibration sequence and construct a three-dimensional spatio-temporal matrix, including:

[0090] Step 2.1: Process the multi-dimensional vibration sequence using an improved variational mode decomposition algorithm and separate each modal component.

[0091] In this embodiment, a modal number adaptive selection mechanism is introduced. Based on the practical signal complexity of the multi-dimensional vibration sequence, the optimal modal number of the variational mode decomposition algorithm is determined through the Bayesian information criterion to improve the variational mode decomposition algorithm, and an improved variational mode decomposition algorithm is obtained.

[0092] Step 2.2: Use wavelet packet transform to divide the frequency bands of each modal component and calculate the energy entropy within each frequency band as the feature vector.

[0093] Wavelet packet transform can capture weak characteristic signals such as the 2.5kHz harmonic caused by winding deformation and the 1.6kHz resonance generated by core looseness. At the same time, the energy entropy has strong robustness to local noise and can automatically select the optimal decomposition layer according to the specific characteristics of the signal.

[0094] In this embodiment, each modal component is recursively decomposed by a multi-layer filter bank, and each layer divides the frequency band into 2 sub-frequency bands. Taking 4-layer decomposition as an example, the 0-5 kHz frequency band can be divided into 16 uniform sub-frequency bands, such as 0-312.5 Hz, 312.5-625 Hz, …, 4.6875-5 kHz.

[0095] In this embodiment, the energy entropy is used to characterize the complexity of the energy distribution within the frequency band. Among them, the calculation formula of the energy entropy is expressed as:

[0096]

[0097] In the formula, H i represents the energy entropy within the i-th frequency band, N represents the total number of modal components, and p ij represents the proportion of the j-th modal component within the i-th frequency band, and log represents the logarithmic function.

[0098] Step 2.3: Based on the tensor decomposition theory, the eigenvectors are reorganized in a three-dimensional coordinate system of "number of sensors × time window length × frequency band energy" to construct a three-dimensional spatio-temporal matrix.

[0099] In this embodiment, the CP (CANDECOMP / PARAFAC) decomposition algorithm is used to reorganize the eigenvectors into a three-dimensional tensor X ∈ R I×J×K , where I represents the number of sensors, such as 6 on the fuel tank side wall + 4 on the heat sink, a total of 10; J represents the time window length, such as 200 ms after dynamic optimization, and K represents the number of frequency band energy channels, such as 16.

[0100] In this embodiment, spatio-temporal correlation features are extracted through tensor slicing operations to locate the position of the fault source, such as abnormal vibration in the heat sink area corresponding to local overheating.

[0101] In this embodiment, the time window length is determined by the dynamic programming algorithm, including:

[0102] Step 2.3.1: Define the window length at the current moment as the state variable.

[0103] Step 2.3.2: Define the window length increment at the current moment as the decision variable, where the window length increment represents the change in the window length from the current moment to the next moment.

[0104] Step 2.3.3: According to the state variable, the decision variable, and the stationarity and feature separability vibration of the multi-dimensional vibration sequence within the window, establish a state transition equation for describing the transfer from the current state variable and the current decision variable to the next state variable.

[0105] In this embodiment, the state transition equation is expressed as:

[0106]

[0107] Wherein, L t+1 represents the change amount of the window length at the next moment t+1, L t+1 =L t +ΔL t ∈[L min ,L max , L t represents the window length at the current moment, L t ∈[L min ,L max , t represents the current moment, L min represents the preset minimum window, L max represents the preset maximum window length, ΔL t represents the window length increment at the current moment, represents the variance of the multi-dimensional vibration sequence within the window, represents the preset variance threshold, J t represents the ratio of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the window, J th represents the ratio threshold of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the preset window, if represents "if", and represents "and", otherwise represents "otherwise".

[0108] Step 2.3.4: Define a cost function according to the computational complexity and feature separability loss of the state variables to evaluate the quality of each state variable.

[0109] In this embodiment, the cost function is expressed as:

[0110] C(L t ,ΔL t ) = α·Comp(L t ) + β·Loss(L t );

[0111] Wherein, C(L t ,ΔL t ) represents the cost function, α represents the weight coefficient of the computational complexity Comp(L t ) when the window length at the current moment is L t , β represents the weight coefficient of the feature separability loss Loss(L t ) when the window length at the current moment is L t .

[0112] Step 2.3.5: Construct a Bellman equation according to the cost function, state transition equation and discount factor to solve the optimal decision sequence.

[0113] In this embodiment, the Bellman equation is expressed as:

[0114]

[0115] In the formula, V(L t ) represents the optimal cost function value at the current moment when the window length is L t ; V(L t+1 ) represents the optimal cost function value at the current moment when the window length is L t+1 ; γ represents the discount factor; C(L t , ΔL t ) represents the cost function, represents taking the minimum value of the window length increment ΔL t at the current moment; t represents the current moment.

[0116] Step 2.3.6: Starting from the final moment, calculate the optimal decision and optimal cost of each state variable in reverse, and solve the Bellman equation by reverse iteration to obtain the optimal decision sequence.

[0117] Step 2.3.7: Based on the optimal decision sequence, calculate the actual window length sequence forward, and determine the optimal time window length as the time window length.

[0118] In this embodiment, when the working condition changes suddenly, such as a sudden increase in load, the window length can be quickly adjusted, for example, shortened from 200 ms to 100 ms, and the response delay < 200 ms. Compared with the fixed window, the accuracy of working condition recognition is improved by using the dynamic programming algorithm to determine the time window length.

[0119] Step 3: Improve the MobileNetV3 network by inserting a hybrid dilated convolution block and a three-dimensional pooling layer and introducing a convolutional block attention module.

[0120] In some embodiments, improving the MobileNetV3 network by inserting a hybrid dilated convolution block and a three-dimensional pooling layer and introducing a convolutional block attention module includes:

[0121] Step 3.1: Insert a hybrid dilated convolution block after the initial convolutional layer, and cycle the dilation rate configuration using convolutional kernels with dilation rates of 1, 2, and 3 respectively to keep the number of parameters and expand the receptive field to an equivalent range.

[0122] In this embodiment, the hybrid dilated convolution block effectively expands the receptive field through sparse sampling technology. When the dilation rate is set to 2, the equivalent receptive field of a 3×3 convolution kernel can be expanded to 5×5. To avoid the grid effect that may be caused by a single dilation rate, this embodiment cyclically uses convolution operations with dilation rates of 1, 2, and 3 while keeping the number of convolution kernels unchanged. Only by adjusting the dilation rate can the overall number of parameters be ensured to be consistent with that of the standard convolutional layer. In this way, the hybrid dilated convolution block can further expand the equivalent receptive field to 35×35, directly capture the long-range dependencies in the vibration signal, and achieve the fusion of multi-scale features. Specifically, the convolution operation with a dilation rate of 1 can retain the local details of the signal, while the convolution operation with a dilation rate of 3 can extract the macroscopic trends of the signal. The use of the hybrid dilated convolution block can effectively improve the detection rate of fault features.

[0123] In addition, in terms of computational efficiency, the hybrid dilated convolution block also shows significant advantages. Compared with the traditional stacked standard convolutional layers, the use of the hybrid dilated convolution block can increase the inference speed by 15%, thus improving the operation efficiency while ensuring the detection accuracy.

[0124] Step 3.2: Insert a three-dimensional pooling layer after the depthwise separable convolutional layer, and perform downsampling only along the time dimension while retaining the original resolution in the spatial dimension.

[0125] This embodiment uses a 3D maximum pooling kernel to compress only the time dimension and completely retain the spatial resolution of the sensor. By retaining the original data in the spatial dimension, the problem of spatial information loss that may occur in the traditional 2D pooling process is successfully avoided. The compression of the time dimension not only effectively reduces the interference of high-frequency noise but also improves the signal-to-noise ratio. At the same time, since the data in the spatial dimension is completely retained, the positioning accuracy of faults is significantly improved, and the positioning error is strictly controlled within 5 centimeters.

[0126] In addition, compared with full-dimensional pooling, the application of the three-dimensional pooling layer significantly reduces the computational amount of subsequent layers and can operate more efficiently, thus meeting the real-time requirements of online monitoring.

[0127] In summary, the 3D maximum pooling kernel adopted in this embodiment retains the key spatial information, reduces the noise interference by compressing the time dimension, and optimizes the computational efficiency, providing strong support for online monitoring.

[0128] Step 3.3: Introduce a convolutional block attention module in the bottleneck layer to perform dual attention weighting on the spatio-temporal feature map in terms of channels and space.

[0129] In some embodiments, it further includes connecting a fully-connected layer after the improved MobileNetV3 network, using the fully-connected layer to classify the operating conditions of the high-dimensional features output by the improved MobileNetV3 network, and outputting the operating condition probability distribution through the Softmax function.

[0130] Step 4: Input the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network to obtain the transformer operating condition probability distribution.

[0131] Step 5: Output the transformer operating condition recognition result according to the operating condition probability distribution.

[0132] Embodiment 3

[0133] Referring to Figure 2 , an embodiment of the present invention, with the same inventive concept as Embodiment 1 or 2, provides a transformer operating condition recognition system based on multi-point vibration and improved MobileNet, including:

[0134] A vibration excitation module for applying synchronous vibration excitation to multiple preset points of the transformer;

[0135] A signal acquisition module for collecting multi-dimensional vibration sequences;

[0136] A feature extraction module for extracting the band energy features of the multi-dimensional vibration sequence and constructing a three-dimensional spatio-temporal matrix;

[0137] A network inference module for inputting the three-dimensional spatio-temporal matrix into the improved MobileNetV3 network to obtain the transformer operating condition probability distribution;

[0138] A decision output module for outputting the transformer operating condition recognition result according to the operating condition probability distribution;

[0139] Among them, the MobileNetV3 network is improved by inserting hybrid dilated convolution blocks and three-dimensional pooling layers and introducing a convolutional block attention module.

[0140] The specific function implementation of the above modules refers to the relevant content in the methods of Embodiment 1 or 2 and will not be elaborated here.

[0141] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Among them, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or a plurality of processes and / or Figure 1 boxes or a plurality of boxes.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A transformer condition recognition method based on multi-point vibration and improved MobileNet, characterized in that Including: Applying synchronous vibration excitation to multiple preset points of the transformer and collecting multi-dimensional vibration sequences; Extracting the band energy characteristics of the multi-dimensional vibration sequences and constructing a three-dimensional spatio-temporal matrix; Inputting the three-dimensional spatio-temporal matrix into an improved MobileNetV3 network to obtain the transformer operating condition probability distribution; Outputting the transformer operating condition recognition result according to the operating condition probability distribution; Among them, the MobileNetV3 network is improved by inserting hybrid dilated convolution blocks and three-dimensional pooling layers and introducing a convolutional block attention module.

2. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 1, characterized in that, Applying synchronous vibration excitation to multiple preset points of the transformer and collecting multi-dimensional vibration sequences, including: Arranging multiple groups of piezoelectric and inertial sensor arrays on the side wall of the transformer oil tank and the radiator area; Applying a mixed excitation signal of pulse excitation and swept-frequency excitation with a frequency band in the range of 0 - 5 kHz and a frequency increasing exponentially from 10 Hz to 10 kHz through a multi-channel synchronous signal generator to form a multi-physical-field coupled vibration environment; Using a distributed acquisition system to synchronously obtain the six-axis vibration signals of each sensor node in the sensor array and constructing a multi-dimensional vibration sequence including displacement, velocity, and acceleration.

3. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 2, wherein, Using wavelet packet transform to extract the band energy characteristics of the multi-dimensional vibration sequences and constructing a three-dimensional spatio-temporal matrix, including: Processing the multi-dimensional vibration sequences using an improved variational mode decomposition algorithm and separating each modal component; Using wavelet packet transform to divide the frequency bands of each modal component and calculating the energy entropy within each frequency band as a feature vector; Recombining the feature vectors according to the three-dimensional coordinate system of "number of sensors × time window length × band energy" based on tensor decomposition theory to construct a three-dimensional spatio-temporal matrix.

4. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 3, characterized in that, Introducing a modal number adaptive selection mechanism, and determining the optimal modal number of the variational mode decomposition algorithm based on the complexity of the practical signal of the multi-dimensional vibration sequence through the Bayesian information criterion to improve the variational mode decomposition algorithm, obtaining the improved variational mode decomposition algorithm.

5. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 3, characterized in that, Determining the time window length through a dynamic programming algorithm, including: Defining the window length at the current moment as a state variable; Defining the window length increment at the current moment as a decision variable, where the window length increment represents the change in the window length from the current moment to the next moment; Establishing a state transition equation according to the state variable, the decision variable, and the stationarity and feature separability vibration of the multi-dimensional vibration sequence within the window, which is used to describe the transfer from the current state variable and the current decision variable to the next state variable; Defining a cost function according to the computational complexity and feature separability loss of the state variable, which is used to evaluate the quality of each state variable; Constructing a Bellman equation according to the cost function, the state transition equation, and the discount factor, which is used to solve the optimal decision sequence; Backward calculating the optimal decision and optimal cost of each state variable starting from the final moment, and iteratively solving the Bellman equation in reverse to obtain the optimal decision sequence; Forward calculating the actual window length sequence based on the optimal decision sequence and determining the optimal time window length as the time window length.

6. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 1, wherein, The state transition equation is expressed as: where L t+1 represents the change in window length at the next moment t + 1, L t+1 = L t + ΔL t ∈ [L min , L max , L t represents the window length at the current moment, L t ∈ [L min , L max , t represents the current moment, L min represents the preset minimum window, L max represents the preset maximum window length, ΔL t represents the window length increment at the current moment, represents the variance of the multi-dimensional vibration sequence within the window, represents the preset variance threshold, J t represents the ratio of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the window, J th represents the preset ratio threshold of the between-class scatter matrix to the within-class scatter matrix of the multi-dimensional vibration sequence within the window, if represents "if", and represents "and", otherwise represents "otherwise"; The cost function is expressed as: C(L t , ΔL t ) = α·Comp(L t ) + β·Loss(L t ); where C(L t , ΔL t ) represents the cost function, α represents the weight coefficient of the computational complexity Comp(L t ) at the current time when the window length is L t , and β represents the weight coefficient of the feature separability loss Loss(L t ) at the current time when the window length is L t ).

7. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 6, characterized in that, The Bellman equation is expressed as: Where, V(L t ) represents the optimal cost function value when the window length at the current moment is L t , V(L t+1 ) represents the optimal cost function value when the window length at the current moment is L t+1 , γ represents the discount factor, C(L t , ΔL t ) represents the cost function, represents taking the minimum value of the window length increment ΔL t at the current moment, and t represents the current moment.

8. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 1, characterized in that Improve the MobileNetV3 network by inserting a hybrid dilated convolution block and a 3D pooling layer and introducing a convolutional block attention module, including: Insert a hybrid dilated convolution block after the initial convolutional layer, and cycle the dilation rate configuration using convolutional kernels with dilation rates of 1, 2, and 3 respectively, maintaining the number of parameters and expanding the receptive field to the 7×7 equivalent range; Insert a 3D pooling layer after the depthwise separable convolutional layer, and perform downsampling only along the time dimension, retaining the original resolution of the spatial dimension; Introduce a convolutional block attention module in the bottleneck layer to perform dual channel and spatial attention weighting on the spatio-temporal feature map.

9. The transformer condition recognition method based on multi-point vibration and improved MobileNet according to claim 8, wherein, It also includes connecting a fully connected layer after the improved MobileNetV3 network, using the fully connected layer to perform working condition classification on the high-dimensional features output by the improved MobileNetV3 network, and outputting the working condition probability distribution through the Softmax function.

10. The transformer working condition recognition system based on multi-point vibration and improved MobileNet according to claim 1, characterized in that, Include the following modules: A vibration excitation module for applying synchronous vibration excitation to multiple preset points of the transformer; A signal acquisition module for acquiring multi-dimensional vibration sequences; A feature extraction module for extracting the band energy features of the multi-dimensional vibration sequences and constructing a 3D spatio-temporal matrix; A network inference module for inputting the 3D spatio-temporal matrix into the improved MobileNetV3 network to obtain the transformer working condition probability distribution; A decision output module for outputting the transformer working condition recognition result according to the working condition probability distribution; Among them, the MobileNetV3 network is improved by inserting a hybrid dilated convolution block and a 3D pooling layer and introducing a convolutional block attention module.

Images