Transformer mechanical fault diagnosis method and device based on vibration signal detection, and storage medium

Abstract

The invention discloses a transformer mechanical fault diagnosis method and device based on vibration signal detection and a storage medium, and belongs to the technical field of deep learning, and the method comprises the steps: obtaining a to-be-diagnosed vibration signal matrix of the surface of a box body in the operation process of a transformer; inputting the to-be-diagnosed vibration signal matrix into a pre-trained fault diagnosis model to obtain a fault diagnosis result; the fault diagnosis model training method comprises the steps of obtaining a training sample set; wherein the training sample set comprises a historical vibration signal matrix of the surface of the transformer box body and a fault category label corresponding to the historical vibration signal matrix; the training sample set is input into a pre-constructed fault diagnosis model, a trained fault diagnosis model is obtained, the fault diagnosis model integrates a multi-channel self-adaptive weight mechanism, a time sequence segmentation attention mechanism and a time-space interaction mechanism with a liquid neural network to enhance feature expression, and the fault diagnosis model is obtained. And the fault diagnosis accuracy is remarkably improved, and meanwhile, excellent noise robustness and engineering deployment feasibility are achieved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, and in particular relates to a method, device and storage medium for diagnosing mechanical faults in transformers based on vibration signal detection. Background Technology

[0002] As a core piece of equipment in the power system, the condition of the windings and core of the power transformer directly determines the safety of the power grid operation. Winding and core faults are among the main types of power equipment faults. Winding loosening faults, in particular, can lead to serious consequences such as insulation wear, posing a significant threat to the stable operation of the power system. Transformer vibration signals originate from internal electromagnetic and mechanical coupling, and their characteristics dynamically change with the equipment's health status. During normal operation, vibration mainly originates from the magnetostrictive effect of the silicon steel sheets in the core and the electromagnetic force of the windings. When a fault occurs, the vibration signal exhibits characteristic time-frequency changes. These changes are implicit in multi-channel vibration data and possess significant nonlinear and spatiotemporal coupling characteristics, making it difficult for traditional analysis methods to fully capture their dynamic evolution. The evolution of mechanical structural faults or damage exhibits nonlinear characteristics. Vibration signals are affected by various factors during propagation, resulting in attenuation and phase shifts. Their dynamic characteristics undergo a complex and variable evolution process as the mechanical state changes. Traditional linear methods can only capture the stationary features of signals and cannot effectively characterize the weak time-frequency coupling features of early faults. Existing deep learning models also have limitations. Some models are not capable of modeling the gradual evolution of vibration signals, and some models face state drift problems when processing long time series data, making it impossible to accurately model the cross-cycle evolution mode of faults.

[0003] Traditional monitoring methods have significant limitations: physical modeling requires simplification of complex operating conditions, affecting practical application; manual feature engineering relies on expert experience to extract feature indicators, making it difficult to effectively identify early fault characteristics. While deep learning technology has made progress in recent years, some models suffer from difficulties in capturing non-stationary signals, low noise robustness, long-term reliance on difficult-to-capture data, and overfitting with small samples. Liquid neural networks can avoid these problems; however, they are inherently susceptible to redundancy, struggle to capture dynamic correlations across long time scales, and are difficult to correlate with spatiotemporal data. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device, and storage medium for diagnosing transformer mechanical faults based on vibration signal detection. It integrates a multi-channel adaptive weighting mechanism, a time-series segmented attention mechanism, a spatiotemporal interaction mechanism, and a liquid neural network. Through a continuous-time dynamic system, it performs refined analysis of winding vibration signals to achieve accurate early fault identification.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for diagnosing mechanical faults in transformers based on vibration signal detection, comprising:

[0007] Obtain the matrix of vibration signals to be diagnosed on the surface of the transformer enclosure during operation;

[0008] The vibration signal matrix to be diagnosed is input into a pre-trained fault diagnosis model to obtain the fault diagnosis result;

[0009] The training method for the fault diagnosis model includes:

[0010] Obtain a training sample set; wherein, the training sample set includes a historical vibration signal matrix of the transformer tank surface and its corresponding fault category labels;

[0011] The training sample set is input into a pre-built fault diagnosis model to obtain a trained fault diagnosis model. The fault diagnosis model enhances feature expression by fusing a multi-channel adaptive weight mechanism, a temporal segmented attention mechanism, and a spatiotemporal interaction mechanism with a liquid neural network.

[0012] Optionally, the method for obtaining the historical vibration signal matrix of the transformer tank surface includes:

[0013] Perform a short-circuit test on the transformer;

[0014] Multi-channel vibration signals are acquired by monitoring multiple vibration sensor points pre-set on the surface of the transformer housing;

[0015] The multi-channel vibration signals are preprocessed to obtain an effective historical vibration signal matrix.

[0016] Optionally, the preprocessing of the multi-channel vibration signal to obtain an effective historical vibration signal matrix includes: performing IQR outlier filtering, linear detrending, 1-200Hz bandpass filtering, and adaptive window smoothing on the multi-channel vibration signal to obtain an effective multi-channel vibration signal that has removed instantaneous high amplitude, baseline drift, and noise frequency bands.

[0017] Optionally, the fault diagnosis model includes an input layer, a feature extraction module, a feature enhancement module, and an output module connected in series.

[0018] The feature extraction module includes a cascaded multi-channel adaptive weight module and a feature extractor;

[0019] The feature enhancement module includes a temporal segmentation attention module, a liquid neuron dynamic modeling module, a spatiotemporal interaction layer, and a feature fusion layer, which are connected in series.

[0020] The output module includes a cascaded augmentation classifier and an output layer.

[0021] Optionally, the step of inputting the training sample set into a pre-constructed fault diagnosis model to obtain a trained fault diagnosis model includes:

[0022] The input layer receives and shapes the historical vibration signal matrix of the input fault diagnosis model to obtain standard input features.

[0023] The feature extraction module extracts features from the standard input features to obtain a high-dimensional feature vector;

[0024] The feature enhancement module performs deep feature extraction on the high-dimensional feature vector, and then fuses and enhances the extracted features to obtain fused and enhanced features.

[0025] The fault diagnosis result is obtained by the output module based on the fusion-enhanced feature mapping.

[0026] The parameters of the fault diagnosis model are adjusted based on the loss between the fault detection results and the corresponding true fault category labels to obtain a trained fault diagnosis model.

[0027] Optionally, the step of extracting features from the standard input features using the feature extraction module to obtain a high-dimensional feature vector includes:

[0028] The physical features of each channel of the standard input feature are extracted by the multi-channel adaptive module, and the value score of each channel is calculated based on the physical features of each channel. The input features after multi-channel fusion weighting are obtained based on the value scores of each channel.

[0029] The input features after multi-channel fusion and weighting are mapped to high-dimensional features using a feature extractor.

[0030] Optionally, the step of performing deep feature extraction on the high-dimensional feature vector through the feature enhancement module, and fusing and enhancing the extracted features to obtain an enhanced feature vector includes:

[0031] The high-dimensional feature vector is segmented temporally by the temporal segmentation attention module to obtain multiple segment vectors. Intra-segment self-attention and inter-segment similarity are calculated for each segment vector to obtain segment attention feature vectors that integrate intra-segment and inter-segment dual attention.

[0032] A neural ODE system is constructed using a liquid neuron dynamic modeling module. The ODE hidden trajectory is obtained by integrating the segment attention feature vector through the neural ODE system.

[0033] By performing parallel spatiotemporal attention weighting and gated fusion on the hidden trajectories of ODE through a spatiotemporal interaction layer, joint spatiotemporal features are obtained.

[0034] The feature fusion layer fuses the joint spatiotemporal features with the high-dimensional feature vector output by the feature extractor to obtain fused enhanced features.

[0035] Optionally, obtaining the fault diagnosis result through the output module based on the fusion-enhanced feature mapping includes:

[0036] The fused enhanced features are processed by an enhanced classifier to calculate the fault category confidence, and the score components of each fault category label are obtained.

[0037] The output layer calculates the probability based on the score components of each fault category label to obtain the predicted fault category label.

[0038] Secondly, the present invention provides a transformer mechanical fault diagnosis device based on vibration signal detection, comprising:

[0039] Diagnostic signal acquisition module: used to acquire the matrix of diagnostic vibration signals on the surface of the transformer enclosure during operation;

[0040] Fault diagnosis result acquisition module: used to input the vibration signal matrix to be diagnosed into a pre-trained fault diagnosis model to obtain fault diagnosis results;

[0041] The training method for the fault diagnosis model includes:

[0042] Obtain a training sample set; wherein, the training sample set includes a historical vibration signal matrix of the transformer tank surface and its corresponding fault category labels;

[0043] The training sample set is input into a pre-built fault diagnosis model to obtain a trained fault diagnosis model. The fault diagnosis model enhances feature expression by fusing a multi-channel adaptive weight mechanism, a temporal segmented attention mechanism, and a spatiotemporal interaction mechanism with a liquid neural network.

[0044] Thirdly, the present invention provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the transformer mechanical fault diagnosis method based on vibration signal detection as described in any of the first aspects.

[0045] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By integrating a multi-channel adaptive weighting mechanism, a temporal segmentation attention mechanism, a spatiotemporal interaction mechanism, and a liquid neural network, continuous-time dynamic mapping of discrete vibration signals is realized, thereby enabling refined analysis of winding vibration signals and providing a high-precision real-time diagnostic solution for transformer mechanical faults. This not only overcomes the dependence of traditional methods on manual feature engineering but also solves the limitations of existing deep models in long sequence processing, multi-source information fusion, and spatiotemporal modeling. It significantly improves the accuracy of fault diagnosis while also possessing excellent noise robustness and engineering deployment feasibility. The multi-channel adaptive weighting module achieves dynamic screening of vibration data by evaluating the physical characteristics of each channel signal. The temporal segmentation attention module divides long sequences into multiple physical period segments based on the fundamental frequency characteristics of the power system and reduces computational complexity through local and cross-segment correlation analysis. The liquid neural network dynamic modeling module couples the spatiotemporal interaction layer to fuse multi-scale temporal features and spatial correlation information, and improves feature representation capabilities through a dynamic integration mechanism, solving the defects of traditional LNNs in multi-channel redundancy suppression, long sequence dynamic modeling, and spatial correlation representation. Attached Figure Description

[0046] Figure 1 The diagram shown is a flowchart of a transformer mechanical fault diagnosis method based on vibration signal detection in one embodiment of the present invention.

[0047] Figure 2 The diagram shown is an architecture diagram of a fault diagnosis model in one embodiment of the present invention;

[0048] Figure 3 The diagram shown is a schematic diagram of the confusion matrix of a fault diagnosis model in one embodiment of the present invention;

[0049] Figure 4 The diagram shown is a schematic representation of the performance indicators of a fault diagnosis model in one embodiment of the present invention.

[0050] Figure 5 The figure shown is a schematic diagram of the PR curve of the fault diagnosis model in one embodiment of the present invention;

[0051] Figure 6 The figure shown is a schematic diagram of the accuracy curve of the fault diagnosis model in one embodiment of the present invention;

[0052] Figure 7 The diagram shown illustrates the noise robustness of a fault diagnosis model in one embodiment of the present invention.

[0053] Figure 8 The diagram shown illustrates the computational efficiency of a fault diagnosis model in one embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a transformer mechanical fault diagnosis method based on vibration signal detection. First, multi-channel vibration signals from the transformer tank surface are collected using a short-circuit implementation scheme. After multi-level preprocessing, these signals are input into a Spatial-Temporal Dynamic Weighted Liquid Neural Network (STDW-LNN) fault diagnosis model. The multi-channel adaptive weight module in the fault diagnosis model extracts the physical statistical features of each channel, dynamically calculates the channel value weights, and fuses them to generate a high information density signal. Then, a temporal segmentation attention module is used to divide the signal into segments based on the power fundamental frequency cycle. Long-sequence feature extraction is achieved through local attention and cross-segment physical similarity correlation. Next, the neural ODE system of the liquid neuron dynamic modeling module couples the spatiotemporal interaction layer to achieve continuous evolution modeling of fault features. Finally, the fault diagnosis result is output. This invention solves the shortcomings of traditional LNNs in multi-channel redundancy suppression, long-sequence dynamic modeling, and spatial correlation representation, providing a high-precision real-time diagnosis solution for transformer mechanical faults. The specific implementation steps are as follows:

[0057] Step S1: Multi-channel vibration signal acquisition

[0058] First, multiple vibration sensors are arranged on the surface of the transformer tank according to the spatial location of the windings / core. A three-phase short circuit is implemented on the low-voltage winding of the transformer, and a gradually increasing power frequency voltage is applied to the high-voltage winding until the low-voltage winding current reaches the rated value. During the short-circuit implementation, multiple vibration signals are simultaneously recorded using vibration sensors to obtain the original data matrix.

[0059] Set fault category labels, including winding nut loose, nut axial looseness 30%, nut axial looseness 50%, core radial looseness 50%, core radial looseness, and normal.

[0060] This embodiment describes the setting of mechanical faults in the windings, core, and internal structure fasteners of an oil-immersed power transformer of model S13-M-500 / 10 from Zhengzhou Xianghe Group Electrical Equipment Co., Ltd. The transformer connection group number is Dyn11, the rated voltage on the high-voltage side is 10.5kV, and the rated current is 23.9A.

[0061] Vibration signals were acquired using a piezoelectric vibration accelerometer (model 1A212E) and a dynamic signal acquisition instrument (DH5922D), with three measuring points placed on the surface of the transformer tank. Considering the frequency range of the transformer vibration signal, the sampling frequency was set to 10kHz during the test.

[0062] After the transformer body is vertically lifted from the oil tank using a bridge crane, the tightness of the winding core is adjusted using a digital torque wrench. For a loosening fault scenario, axial looseness of 30% and 50% is set respectively. At this point, the torque values ​​of the winding clamping bolts should be 25.76 N·m and 18.4 N·m. The torque values ​​of the winding clamping bolts are adjusted step-by-step to the corresponding values, with each level of torque reduction not exceeding 10% of the initial torque value, allowing the bolts to gradually adapt to the stress state. After each adjustment, the transformer is allowed to stand for 10 minutes to release structural stress and stabilize the stress distribution of the pressure plate. A fastener detachment fault is simulated by the detachment of the axially tightened nut of the winding.

[0063] For the core loosening fault setting, the core preload is set to 50% and 0% of the rated value. At this time, the torque value of the core clamp bolt should be 19.36 N·m and 0, corresponding to the fault states of 50% core loosening and complete core loosening. The core preload is adjusted by diagonal alternating tightening and loosening method, and the torque value of the clamp bolt is gradually reduced to the corresponding value. Each torque reduction value does not exceed 10% of the initial torque value to avoid sudden changes in the core preload and local gaps in the lamination. Similarly, after each adjustment, it is allowed to stand for 10 minutes to make the final set torque value more stable and improve the reliability of the experimental data.

[0064] Step S2: Sample set construction

[0065] Multi-level preprocessing is performed on each vibration signal in multiple sets of original data matrices, sequentially as follows:

[0066] IQR (Interquartile Range) outlier removal: Removes outliers that exceed a preset range;

[0067] Linear detrending: Eliminates temperature drift or sensor zero drift;

[0068] 1-200 Hz bandpass filter: preserves the main frequency band of mechanical vibration of the iron core / winding;

[0069] Adaptive window smoothing: The vibration signal in the main frequency band of mechanical vibration is smoothed to obtain an effective data matrix, reducing the amount of subsequent model calculations.

[0070] By combining each set of valid data matrices with the corresponding real fault category labels, a sample set of mechanical faults in transformer windings is generated.

[0071] Step S3: Construct the STDW-LNN fault diagnosis model

[0072] like Figure 3 As shown, the STDW-LNN fault diagnosis model includes an input layer, a feature extraction module, a feature enhancement module, and an output module connected in series. The feature extraction module includes a multi-channel adaptive weight module and a feature extractor connected in series; the feature enhancement module includes a temporal segmentation attention module, a liquid neuron dynamic modeling module, a spatiotemporal interaction layer, and a feature fusion layer connected in series; the output module includes an enhanced classifier and an output layer connected in series. The multi-channel adaptive weight module generally includes a feature channel processing part, a weight learning part, and a feature fusion part. The feature channel processing part consists of independent processing units for each channel, used to process the features of each input channel individually; the weight learning part typically includes a feature compression component, a feature excitation component, and a weight generation component. The feature compression component compresses the features of each channel into a global description, the feature excitation component captures the dependencies between channels through nonlinear transformations, and the weight generation component outputs the weights corresponding to each channel; the feature fusion part combines the features of each channel with their corresponding weights through operations such as multiplication or addition to form the final fused features. The temporal segmentation attention module comprises a temporal segmentation unit, a local feature extraction unit, a global association learning unit, a weight generation unit, and a feature integration unit. The temporal segmentation unit is responsible for dividing the input temporal data into continuous segments; the local feature extraction unit processes each segment to obtain local features; the global association learning unit captures the relationships between different segments; the weight generation unit generates corresponding attention weights based on the importance of each segment; and the feature integration unit aggregates the weighted segment features into overall temporal features.

[0073] Step S4: Fault Diagnosis Model Training

[0074] A hierarchical temporal partitioning method was used to divide the sample set, with a training set:validation set:test set ratio of 7:2:1. The training set data was input into the STDW-LNN fault diagnosis model, and the model was trained using cross-entropy loss and the Adam optimizer. After multiple rounds of training, the model performance was calculated using the validation set data, and the model hyperparameters were adjusted. If the performance did not improve after multiple consecutive validations, the iteration was stopped. Finally, the final performance index of the model was calculated using the test set data. The specific steps of each training round are as follows:

[0075] The input layer receives and shapes the effective data matrix of the input fault diagnosis model to obtain standard input features.

[0076] The multi-channel adaptive module extracts the standard deviation of each channel. kurtosis Peak Composition of physical feature vectors And calculate the value score of each channel using learnable parameters. ,in, Indicates learnable parameters, The bias is represented by the input feature after being normalized by softmax and weighted by the main channel to obtain the input feature after multi-channel fusion weighting; the feature extractor performs high-dimensional feature mapping on the input feature after multi-channel fusion weighting.

[0077] The calculation of the physical characteristics includes: standard deviation Used to characterize the intensity of signal fluctuations, calculated using the formula... ;in, This indicates the number of signal data points involved in the calculation. Represents a single data value in the signal. This represents the average of all data.

[0078] Kudo Used to capture impact characteristics, calculation formula ;in, Indicates the expression Take the average. Represents a single data value in the signal. This represents the average value of all signal data. Standard deviation; Peak value Used to detect extreme amplitudes, the calculation formula is as follows: ; among them, in time The signal value at a given point describes how the signal changes over time.

[0079] The temporal segmentation attention module performs temporal segmentation on the high-dimensional feature vector according to the power fundamental frequency cycle, with segment length... ,in, Indicates the sampling frequency. Indicates the power fundamental frequency. The segmentation factor is represented, resulting in multiple segment vectors. Each segment vector is then subjected to intra-segment scaling dot product self-attention and inter-segment physical feature signature cosine similarity correlation to obtain a segment attention feature vector that integrates intra-segment and inter-segment dual attention.

[0080] In this embodiment, the timing segment length The physical meaning is that each segment contains 2-8 power fundamental frequency cycles, which is specifically implemented through segmentation factors. Number of control cycles : , .

[0081] The liquid neuron dynamic modeling module constructs a neural ODE system: ,in, This indicates that the liquid neuron is at time 10:00. The internal state, This represents the weight matrix, used to analyze the neuron state. Perform a weighted operation. This represents the input weight matrix, used to weight external inputs. Perform a weighted operation. Represents the external input signal, dynamic time constant. , Seconds; the ODE hidden trajectory is obtained by integrating the segment attention feature vector through a neural ODE system. The neural ODE solution employs an adaptive Dormand-Prince algorithm with a relative error tolerance. .

[0082] The spatiotemporal interaction layer first injects spatiotemporal position encoding into the hidden trajectory of ODE: the time encoding adopts... ,in, It is usually related to the feature dimensions of the model, etc. These are parameters such as the index, which are used to control the frequency and other characteristics of the encoding to distinguish the encoding at different time positions; spatial encoding is a learnable sensor embedding matrix; then it goes through parallel temporal and spatial paths; temporal path: multi-scale TCN convolutional kernel size Dynamic weighting through channel attention; Spatial path: Graph attention mechanism based on vibration propagation topology, attention coefficient calculation formula: ,in, and This represents the node index in the graph. This represents the total number of nodes in the graph. This is used to transform the attention score into a non-negative exponential function. For activation function, For learnable parameters, , , It is a learnable weight matrix used to perform linear transformations on different node features; finally, through gated fusion, it outputs joint spatiotemporal features. ,in, The gating weight vector representing the time path. Features representing time path extraction The gating weight vector representing the spatial path. This represents the features extracted from the spatial path; the feature fusion layer will combine spatiotemporal features. The fused feature vector is then fused with the high-dimensional feature vector output by the feature extractor to obtain the fused enhanced feature.

[0083] The augmented classifier calculates the fault category confidence of the fused augmented features, obtaining the score components of each fault category label. The output layer calculates the probability based on the score components of each fault category label to obtain the predicted fault category label. The parameters of the fault diagnosis model are adjusted according to the loss between the fault detection results and the corresponding true fault category labels.

[0084] This embodiment comprehensively evaluates the classification performance of the proposed STDW-LNN model using confusion matrix, precision / recall / F1 score, and PR curve. A systematic evaluation is also conducted based on training dynamics, adaptability to noisy environments, and engineering deployment efficiency. The results show that it exhibits excellent accuracy and reliability in fault type identification tasks and demonstrates superior overall performance in complex industrial scenarios. Figure 3 As shown, based on the confusion matrix analysis of 1800 samples, the overall accuracy of the model reached 99.8%, with only 3 false negative samples. These were all misclassified as normal in the "radial loosening of the iron core" category. No false positive results were found, demonstrating a very strong ability to distinguish between positive and negative samples. The proportion of diagonal elements in the confusion matrix was significantly higher than that of off-diagonal elements, indicating that the model has extremely high consistency in classifying samples of each category, with almost no class confusion.

[0085] Further quantitative analysis of the classification indicators for each fault type, such as... Figure 4 The precision, recall, and F1 score for all categories shown are extremely high: the precision for the "normal" class is 0.990, the recall for the "radial loosening of the iron core" class is 0.990, and the metrics for the remaining classes are all 1.000. The F1 score for all classes exceeds 0.995. This indicates that the model is not only stable in recognizing majority class samples but also has excellent ability to capture minority class samples, solving the performance degradation problem of traditional models in imbalanced data.

[0086] like Figure 5 As shown, the Precision-Recall Curve further validates the model's robustness. The Average Precision (AP) for all fault categories exceeds 0.98, indicating that the model's precision consistently remains at the theoretically optimal level under different recall thresholds, achieving a perfect trade-off between precision and recall. This result far surpasses traditional feature engineering methods, confirming the model's ability to capture complex fault modes after enhancing feature discriminative power.

[0087] like Figure 6The results show that the model exhibits the dual advantages of rapid convergence and stable generalization during training: the accuracy on both the training and validation sets reaches 0.998, and the final training loss converges to 0.2711. The training process demonstrates that the model achieves a significant increase in accuracy within 10 iterations and maintains stability, without overfitting. The validation set performance is highly consistent with the training set performance, indicating that the enhanced feature extraction module effectively improves the model's ability to fit the data distribution and its generalization boundary.

[0088] like Figure 7 As shown, in to Under Gaussian noise interference, the model accuracy consistently remained above 0.995, especially under strong noise levels. The accuracy remained at 0.998, consistent with performance in noise-free scenarios. Robustness analysis of typical industrial equipment fault categories showed that the robustness score for the "winding nut detachment" fault reached 0.598, significantly higher than traditional CNN models, indicating that the model's ability to capture weak fault features remained stable even under noise interference. This anti-interference characteristic stems from the structured encoding of key fault modes by the feature enhancement module, effectively suppressing feature drift caused by noise.

[0089] like Figure 8 As shown, the model exhibits significant advantages in computational efficiency and resource consumption. During the training phase, each iteration takes only 1.50 seconds, with peak memory consumption controlled at 350 MB, enabling end-to-end training to be completed on a typical industrial server. During the inference phase, with a batch size of 60, the single-sample inference speed reaches 0.175 ms, meeting the sub-millisecond response requirements of industrial real-time detection systems.

[0090] Step S5: Application of Fault Diagnosis Model

[0091] Based on the historical data collection steps, vibration signals are collected from the surface of the transformer in operation. The collected vibration signal matrix to be diagnosed is input into the trained fault diagnosis model to obtain the predicted fault classification results, thereby achieving accurate diagnosis of mechanical faults in the transformer windings.

[0092] Example 2

[0093] This embodiment provides a transformer mechanical fault diagnosis device based on vibration signal detection, including:

[0094] Diagnostic signal acquisition module: used to acquire the matrix of diagnostic vibration signals on the surface of the transformer enclosure during operation;

[0095] Fault diagnosis result acquisition module: used to input the vibration signal matrix to be diagnosed into a pre-trained fault diagnosis model to obtain fault diagnosis results;

[0096] The training method for the fault diagnosis model includes:

[0097] Obtain a training sample set; wherein, the training sample set includes a historical vibration signal matrix of the transformer tank surface and its corresponding fault category labels;

[0098] The training sample set is input into a pre-built fault diagnosis model to obtain a trained fault diagnosis model. The fault diagnosis model enhances feature expression by fusing a multi-channel adaptive weight mechanism, a temporal segmented attention mechanism, and a spatiotemporal interaction mechanism with a liquid neural network.

[0099] The device provided in this embodiment can execute the transformer mechanical fault diagnosis method based on vibration signal detection provided in any step of Embodiment 1, and has the corresponding functional modules and beneficial effects of the method.

[0100] Example 3

[0101] This embodiment provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the transformer mechanical fault diagnosis method based on vibration signal detection as provided in any step of Embodiment 1.

[0102] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0104] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0106] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A transformer mechanical fault diagnosis method based on vibration signal detection, characterized in that, The method comprises the following steps: obtaining a vibration signal matrix to be diagnosed on the surface of a transformer tank during operation of the transformer; inputting the vibration signal matrix to be diagnosed into a pre-trained fault diagnosis model to obtain a fault diagnosis result; wherein the training method of the fault diagnosis model comprises: obtaining a training sample set, wherein the training sample set comprises historical vibration signal matrices on the surface of a transformer tank and corresponding fault category labels; inputting the training sample set into a pre-constructed fault diagnosis model to obtain a trained fault diagnosis model, wherein the fault diagnosis model is enhanced by fusing a multi-channel adaptive weight mechanism, a time sequence segmentation attention mechanism, a space-time interaction mechanism and a liquid state neural network to enhance feature expression.

2. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 1, characterized in that, The method for obtaining the historical vibration signal matrix on the surface of the transformer tank comprises: performing a short-circuit test on the transformer; obtaining multi-channel vibration signals through a plurality of vibration sensor monitoring points pre-set on the surface of the transformer tank; preprocessing the multi-channel vibration signals to obtain an effective historical vibration signal matrix.

3. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 1, characterized in that, The preprocessing of the multi-channel vibration signals to obtain the effective historical vibration signal matrix comprises: performing IQR abnormal value filtering, linear detrending, 1-200 Hz band-pass filtering and adaptive window smoothing on the multi-channel vibration signals respectively to obtain effective multi-channel vibration signals from which instantaneous high amplitude, baseline drift and noise frequency bands are removed.

4. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 1, characterized in that, The fault diagnosis model comprises an input layer, a feature extraction module, a feature enhancement module and an output module connected in series; The feature extraction module comprises a multi-channel adaptive weight module and a feature extractor connected in series; The feature enhancement module comprises a time sequence segmentation attention module, a liquid state neuron dynamic modeling module, a space-time interaction layer and a feature fusion layer connected in series; The output module comprises an enhanced classifier and an output layer connected in series.

5. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 4, characterized in that, The inputting of the training sample set into the pre-constructed fault diagnosis model to obtain the trained fault diagnosis model comprises: receiving and shaping the historical vibration signal matrix input into the fault diagnosis model through the input layer to obtain standard input features; extracting features from the standard input features through the feature extraction module to obtain a high-dimensional feature vector; extracting deep features from the high-dimensional feature vector through the feature enhancement module, and fusing and enhancing the features extracted in the deep layer to obtain fused and enhanced features; mapping the fused and enhanced features through the output module to obtain a fault diagnosis result; adjusting parameters of the fault diagnosis model according to the loss between the fault detection result and the corresponding true fault category label to obtain the trained fault diagnosis model.

6. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 5, characterized in that, The feature extraction from the standard input features through the feature extraction module to obtain the high-dimensional feature vector comprises: extracting physical features of each channel of the standard input features through the multi-channel adaptive module, calculating a value score of each channel according to the physical features of each channel, and obtaining input features weighted by multi-channel fusion weights according to the value score of each channel; performing high-dimensional feature mapping on the input features weighted by the multi-channel fusion weights through the feature extractor.

7. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 5, characterized in that, The high-dimensional feature vector is subjected to deep feature extraction by the feature enhancement module, and the features after deep extraction are fused and enhanced to obtain an enhanced feature vector, including: The high-dimensional feature vector is subjected to time sequence segmentation by the time sequence segmentation attention module to obtain a plurality of segment vectors, and the segment vectors are subjected to intra-segment self-attention calculation and inter-segment similarity calculation to obtain segment attention feature vectors fused with intra-segment and inter-segment double attention; The neural ODE system is constructed by the liquid neuron dynamic modeling module, and the segment attention feature vectors are integrated by the neural ODE system to obtain ODE hidden trajectories; The ODE hidden trajectories are subjected to parallel space-time attention weighting and gate fusion by the space-time interaction layer to obtain joint space-time features; The joint space-time features are fused with the high-dimensional feature vectors output by the feature extractor by the feature fusion layer to obtain fused and enhanced features.

8. The transformer mechanical fault diagnosis method based on vibration signal detection according to claim 5, characterized in that, The output module is used to map the fused and enhanced features to obtain a fault diagnosis result, including: The fused and enhanced features are subjected to fault category confidence calculation by the enhanced classifier to obtain score components of each fault category label; The probability calculation is performed according to the score components of each fault category label by the output layer to obtain a predicted fault category label.

9. A transformer mechanical fault diagnosis device based on vibration signal detection, characterized in that, It includes: A to-be-diagnosed signal acquisition module is configured to acquire a to-be-diagnosed vibration signal matrix on the surface of a transformer tank during operation of the transformer; A fault diagnosis result acquisition module is configured to input the to-be-diagnosed vibration signal matrix into a pre-trained fault diagnosis model to obtain a fault diagnosis result; The training method of the fault diagnosis model includes: A training sample set is acquired, wherein the training sample set includes historical vibration signal matrices on the surface of a transformer tank and corresponding fault category labels; The training sample set is input into a pre-constructed fault diagnosis model to obtain a trained fault diagnosis model, wherein the fault diagnosis model is fused and enhanced by a multi-channel adaptive weight mechanism, a time sequence segmentation attention mechanism, and a space-time interaction mechanism and a liquid neural network.

10. A computer storage medium having stored thereon a computer program, characterized in that When the computer program is executed by the processor, the transformer mechanical fault diagnosis method based on vibration signal detection is realized.

Images