Transformer winding mechanical fault severity identification method, device and storage medium

Abstract

The present invention discloses a method, device and storage medium for identifying the severity of mechanical faults of transformer windings, comprising first collecting vibration signals of transformers during operation through sensors; dividing the collected vibration signals into two groups, one group being training data and the other group being test data; using the training data to train a CLSTM model enhanced by a multi-head attention mechanism, and using the test data to test the generalization ability of the trained CLSTM model and locally fine-tuning the model based on the results, and finally using the locally fine-tuned CLSTM model to identify mechanical faults of transformers, and on this basis, using the SSDA model to identify the severity of the diagnosed mechanical faults. The present invention can reduce the workload of substation operation and maintenance personnel, realize active identification of mechanical faults of transformer windings and their severity, and has important significance for the digital and intelligent transformation of substation equipment operation and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and in particular to a method, a device and a storage medium for identifying the severity of a transformer winding mechanical fault. Background Art

[0002] In substations, the operating status of transformers is directly related to the safety and reliability of power generation and power supply systems. Existing methods for transformer fault diagnosis mainly include oil chromatography, low-voltage pulse method, short-circuit impedance method, and frequency response analysis method. Existing methods all focus on identifying mechanical faults of transformers, but cannot determine the severity of mechanical faults of transformers. This makes it impossible for operation and maintenance personnel to arrange the order of maintenance according to the diagnosis results, which easily leads to a waste of manpower and maintenance resources. Summary of the invention

[0003] The present invention provides a method, device and storage medium for identifying the severity of a transformer winding mechanical fault, which can at least solve one of the technical problems in the background technology.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A transformer mechanical fault diagnosis method, which performs the following steps by computer equipment:

[0006] The vibration signals generated by the transformer operation are collected by sensors, and the collected vibration signals are randomly divided into training data sets and test data sets. The training data sets and test data sets are processed separately using Ensemble empirical mode decomposition (EEMD) to obtain the corresponding intrinsic mode functions (IMFs).

[0007] The corresponding envelope spectrum function is calculated according to the obtained inherent mode function, and the correlation between each envelope spectrum signal and the original signal is calculated according to the correlation calculation formula. The relevant envelope spectrum function with a correlation coefficient above 0.5 is selected, and the relevant envelope spectrum function is divided into a training set and a test set at a ratio of 5:1. The CLSTM model and the SSDA model are trained based on the training set data, and the test set data is used to perform reverse local fine-tuning on the trained CLSTM model and the SSDA model. The final model output result is the severity of the transformer winding mechanical fault.

[0008] Furthermore, the integrated empirical mode decomposition step includes,

[0009] Assume the original vibration signal of the transformer is x(t) , adding noise to the original model to obtain the reconstructed signal:

[0010]

[0011] in represents the signal after adding noise, represents the mth noise component;

[0012] The reconstructed signal is processed using empirical mode decomposition to obtain the first IMF function and the corresponding residual term:

[0013]

[0014] Where r1 represents the residual term, d1 represents the first IMF function, T {.} indicates the empirical mode decomposition calculation;

[0015] Repeat the above steps until the decomposition is complete.

[0016] Furthermore, the method further includes a step of calculating a related envelope spectrum signal.

[0017] First calculate the envelope spectrum function of the signal:

[0018]

[0019]

[0020] Where d i (t) represents the ith IMF function, H[d i (t)] represents the Hilbert spectrum signal of the i-th IMF function, B j (t) represents the jth envelope spectrum signal;

[0021] Furthermore, on this basis, the correlation calculation formula is used to calculate the correlation between each envelope spectrum and the original signal, and the one with a correlation coefficient above 0.5 is selected as the correlation envelope spectrum function:

[0022]

[0023] In the formula, s e express B e (t) and x ( t ) is the correlation coefficient between them.

[0024] Furthermore, the CLSTM model enhanced by the multi-head attention mechanism is trained and locally fine-tuned, including dividing the relevant envelope spectrum signal into a training data set and a test data set in a 5:1 manner, and then using the training data set to train the CLSTM model enhanced by the multi-head attention mechanism; the trained CLSTM model is further verified using the test data set, and locally fine-tuned according to the verification results until the best fault identification effect is achieved.

[0025] Furthermore, using the locally fine-tuned CLSTM model to realize the transformer mechanical fault vibration includes processing the vibration signal of the transformer to be diagnosed to obtain the relevant envelope spectrum signal, and using the locally fine-tuned multi-head attention mechanism enhanced CLSTM model to process the relevant envelope spectrum signal to obtain the identification result of the transformer winding mechanical fault to be diagnosed.

[0026] Furthermore, based on the fault identification, the SSDA model is used to judge the fault severity and obtain the transformer mechanical fault severity result.

[0027] Furthermore, the CLSTM model enhanced by the multi-head attention mechanism is used to realize transformer mechanical fault diagnosis based on the correlation envelope spectrum signal, including the following steps:

[0028] S21: Construct a CNN convolution layer to extract the correlation between high-dimensional time variables in the relevant envelope spectrum signal. The convolution kernel of the convolution layer is a*b, where a is the width and b is the height. The number of convolution kernels is consistent with the dimension of the relevant envelope spectrum signal. The i-th convolution kernel scans the input relevant envelope spectrum signal R b The resulting output is:

[0029]

[0030] Where RELU(V)=MAX(0,V), Vi represents the weighting coefficient of the input signal, and e represents the bias;

[0031] S22: Construct a basic LSTM unit, including a forget gate, an input gate, and a storage unit:

[0032]

[0033]

[0034]

[0035] In the formula W f is the weight of the forget gate, b f is the bias coefficient, xt express t The input value at the time, W i and b i Represent the weight and bias coefficient of the input gate respectively, Indicates the value of the storage unit currently involved in the calculation, s Represents the sigmoid function:

[0036]

[0037] S23: Establish a CLSTM model and connect the CNN convolutional layer with the LSTM model to ensure that the dependencies between time-related signals are fully mined. The connection method is as follows:

[0038]

[0039] Where P represents the output of the CNN convolutional layer connected to the LSTM model. Represents a concatenation operation.

[0040] S24: The obtained CLSTM model is enhanced by using a multi-head attention mechanism, and a multi-head attention layer is added to the LSTM to realize the extraction of global features of the input related envelope spectrum signal.

[0041] Furthermore, the multi-head attention mechanism, based on the CLSTM model mining the interdependent correlation of each time component, enhances the CLSTM model's learning of the global features of the input related envelope spectrum signal. The specific steps are as follows:

[0042] S31: For the input correlation envelope spectrum signal T, it is assumed to be n-dimensional data with m sampling points, that is:

[0043] T = [ t 1 , t 2 , …, t n ], n ∈R. t n = [ , , …, ] T , m ∈R

[0044] Where R represents the real number domain;

[0045] S32: Embed the concat function into each node of the CLSTM model to obtain the combined representation of each node:

[0046]

[0047] S33: T is calculated by n sets of linear transformation matrices m The attention parameter Q m , K m , V m And its single head feature expression head m :

[0048]

[0049] Where Q m , K m , V m Respectively represent T m The attention parameter, head m represents the single-head feature expression of the mth node, softmax is the activation function, Used to avoid dimensionality explosion during calculations.

[0050] S34: output data of each node after multi-head attention enhancement;

[0051]

[0052] In the formula x n represents the original input data of the CLSTM model, Represents the output data after the introduction of the multi-head attention mechanism.

[0053] Furthermore, the mechanical fault severity identification uses the output data obtained by the CLSTM model enhanced by the multi-head attention mechanism as the input data of the SSDA model, and uses the sparse characteristics of the SSDA model and the global feature mining and comparison capabilities to realize the identification of the fault severity. The specific steps are as follows:

[0054] S41: Let the output data of the CLSTM model be T1, randomly insert Masking noise into the data set T1, and map it to the hidden layer of the DA model:

[0055]

[0056] in i h = { , } is a parameter set, The weight matrix in the mapping process, Bias term vector, s ( x ) = is the sigmoid function, which is used to implement the above mapping, and h represents the data in the hidden layer;

[0057] S42: Reverse mapping is performed on the data of the hidden layer to obtain the reconstructed input data T11, and the loss function value between T1 and T11 is calculated. When the loss function value reaches the minimum value, the DA model is constructed:

[0058]

[0059]

[0060] In the formula i z = { W z , b z} is a parameter set, The weight matrix in the mapping process, Bias term vector. In order to reduce the number of parameters that need to be solved, it is usually W h and W z Make the following restrictions: W h ={ W z} T , T represents transpose; p is the number of original input data, L (x i ,z i ) is the reconstruction error function.

[0061] S43: Train multiple DA models in sequence according to the above steps, extract the hidden layers of each DA model and connect them in sequence, which is the SDA model. On this basis, add a sparse penalty term to the last layer of the SDA model, which is the SSDA model, to avoid the weight explosion in multiple feature captures. The sparse penalty term formula is as follows:

[0062]

[0063] In the formula r Sparse target value, β is the weight value, r j is the average number of activated neurons in hidden unit j;

[0064] S44: The final output result of the SSDA model is the classification value of the severity of the mechanical fault of the transformer winding.

[0065] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0066] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0067] In summary, the transformer winding mechanical fault severity identification method of the present invention fully mines the deep fault features of the transformer vibration signal based on the CLSTM model, and at the same time introduces a multi-head attention mechanism to strengthen the CLSTM model, so that CLSTM takes into account the extraction of long-time scale features on the basis of short-time scale feature extraction, and enhances the CLSTM model's attention to the global features of the input fault signal, thereby improving the model's accuracy in identifying transformer winding mechanical faults. On this basis, the present invention takes the lead in using the SSDA model to sort and divide the severity of the identified mechanical faults. The SSDA model has good anti-interference ability corresponding to the noise component in the transformer winding fault signal due to its introduction of a denoising mechanism and a sparse penalty term, and is more accurate in fault severity classification, and its classification results are more valuable for operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is an overall block diagram of an embodiment of the present invention;

[0069] Figure 2 The CLSTM model structure of the embodiment of the present invention;

[0070] Figure 3 Schematic diagram of a multi-head attention mechanism according to an embodiment of the present invention;

[0071] Figure 4 This is a structural diagram of the SSDA model of an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0073] A method for identifying the severity of a transformer winding mechanical fault according to an embodiment of the present invention uses a sensor to collect vibration signals generated by transformer operation, uses integrated empirical mode decomposition to obtain corresponding intrinsic mode functions (IMFs) from the collected vibration signals and calculates corresponding envelope spectrum functions, selects relevant envelope spectrum signals with a correlation degree of more than 0.5 from all envelope spectrum signals based on correlation coefficients, and divides them into a training data set and a test data set at a ratio of 5:1; uses the training data set to train a CLSTM model enhanced by a multi-head attention mechanism, further uses the test data set to locally fine-tune the trained CLSTM model, and uses the locally fine-tuned CLSTM model to realize transformer mechanical fault identification; further, divides the fault signal into a training set and a data set at a ratio of 4:1, uses the training set to train an SSDA model, and uses the test data set to reversely locally fine-tune the trained SSDA model, and the output of the fine-tuned SSDA model is the severity of the transformer winding mechanical fault.

[0074] The specific steps are as follows:

[0075] S1: Integrated variational empirical mode decomposition. Assume that the original vibration signal of the transformer is x(t) , adding noise to the original model to obtain the reconstructed signal:

[0076]

[0077] in represents the signal after adding noise, represents the mth noise component;

[0078] The reconstructed signal is processed using empirical mode decomposition to obtain the first IMF function and the corresponding residual term:

[0079]

[0080] Where r1 represents the residual term, d1 represents the first IMF function, T {.} indicates the empirical mode decomposition calculation; repeat the above steps until the decomposition is completed.

[0081] S2: Calculation of relevant envelope spectrum function. First, calculate the envelope spectrum function of the signal:

[0082]

[0083]

[0084] Where d i (t) represents the ith IMF function, H[d i (t)] represents the Hilbert spectrum signal of the i-th IMF function, B j(t) represents the jth envelope spectrum signal;

[0085] On this basis, the correlation calculation formula is used to calculate the correlation between each envelope spectrum and the original signal, and the correlation coefficient above 0.5 is selected as the correlation envelope spectrum function:

[0086]

[0087] In the formula, s e express B e (t) and x ( t ) is the correlation coefficient between them.

[0088] S3: Training and local fine-tuning of the multi-head attention enhanced CLSTM model. The relevant envelope spectrum signal is divided into a training data set and a test data set in a 5:1 ratio, and then the multi-head attention enhanced CLSTM model is trained using the training data set; the trained CLSTM model is further verified using the test data set, and local fine-tuning is performed based on the verification results until the best fault diagnosis and identification effect is achieved.

[0089] S4: Identification of mechanical faults of transformer windings. The vibration signal of the transformer to be diagnosed is processed to obtain the relevant envelope spectrum function, and the relevant envelope spectrum function is processed using the multi-head attention enhanced CLSTM model after local fine-tuning in S3 to realize the identification of the mechanical faults of the transformer winding to be diagnosed.

[0090] The multi-head attention enhanced CLSTM model is used to realize transformer mechanical fault identification based on the relevant envelope spectrum function, and includes the following steps:

[0091] S21: Construct a CNN convolution layer to extract the correlation between high-dimensional time variables in the relevant envelope spectrum signal. The convolution kernel of the convolution layer is a*b, where a is the width and b is the height. The number of convolution kernels is consistent with the dimension of the relevant envelope spectrum signal. The i-th convolution kernel scans the input relevant envelope spectrum signal R b The resulting output is:

[0092]

[0093] In the formula, RELU(V)=MAX(0,V), Vi represents the weighting coefficient of the input signal, and e represents the bias.

[0094] S22: Construct a basic LSTM unit, including a forget gate, an input gate, and a storage unit:

[0095]

[0096]

[0097]

[0098] In the formula W f is the weight of the forget gate, b f is the bias coefficient, x t express t The input value at the time, W i and b i Represent the weight and bias coefficient of the input gate respectively, Indicates the value of the storage unit currently involved in the calculation, s Represents the sigmoid function:

[0099]

[0100] S23: Establish a CLSTM model and connect the CNN convolutional layer with the LSTM model to ensure that the dependencies between time-related signals are fully mined. The connection method is as follows:

[0101]

[0102] Where P represents the output of the CNN convolutional layer connected to the LSTM model. Represents a concatenation operation.

[0103] S24: The obtained CLSTM model is enhanced by using a multi-head attention mechanism, and a multi-head attention layer is added to the LSTM to realize the extraction of global features of the input related envelope spectrum signal.

[0104] The multi-head attention mechanism, based on the CLSTM model mining the interdependent correlation of each time component, enhances the CLSTM model's learning of the global features of the input related envelope spectrum signal. The specific steps are as follows:

[0105] S31: For the input correlation envelope spectrum signal T, it is assumed to be n-dimensional data with m sampling points, that is:

[0106] T = [ t 1 , t 2 , …, t n ], n ∈R. t n = [ , , …, ] T , m ∈R

[0107] Where R represents the real number domain;

[0108] S32: Embed the concat function into each node of the CLSTM model to obtain the combined representation of each node:

[0109]

[0110] S33: T is calculated by n sets of linear transformation matrices m The attention parameter Q m , K m , V m And its single head feature expression head m :

[0111]

[0112] Where Q m , K m , V m Respectively represent T m The attention parameter, head m represents the single-head feature expression of the mth node, softmax is the activation function, Used to avoid dimensionality explosion during calculations.

[0113] S34: output data of each node after multi-head attention enhancement;

[0114]

[0115] In the formula x n represents the original input data of the CLSTM model, Represents the output data after the introduction of the multi-head attention mechanism.

[0116] S5: Identification of the severity of mechanical faults in transformer windings. The output data obtained by the CLSTM model enhanced by the multi-head attention mechanism is used as the input data of the SSDA model. The sparse characteristics of the SSDA model and the global feature mining and comparison capabilities are used to identify the severity of the fault.

[0117] The SSDA model is used to identify the severity of the fault category obtained by the CLSTM model. The specific steps are as follows:

[0118] S41: Let the output data of the CLSTM model be T1, randomly insert Masking noise into the data set T1, and map it to the hidden layer of the DA model:

[0119]

[0120] in i h = { , } is a parameter set, The weight matrix in the mapping process, Bias term vector, s ( x ) = is the sigmoid function, which is used to implement the above mapping, and h represents the data in the hidden layer;

[0121] S42: Reverse mapping is performed on the data of the hidden layer to obtain the reconstructed input data T11, and the loss function value between T1 and T11 is calculated. When the loss function value reaches the minimum value, the DA model is constructed:

[0122]

[0123]

[0124] In the formula i z = { W z , b z} is a parameter set, The weight matrix in the mapping process, Bias term vector. In order to reduce the number of parameters that need to be solved, it is usually W h and W z Make the following restrictions: W h ={ W z} T , T represents transpose; p is the number of original input data, L (x i ,z i ) is the reconstruction error function.

[0125] S43: According to the above steps, multiple DA models are trained in sequence, and the hidden layers of each DA model are extracted and connected in sequence, which is the SDA model. On this basis, a sparse penalty term is added to the last layer of the SDA model to avoid the weight explosion in multiple feature captures. The sparse penalty term formula is as follows:

[0126]

[0127] In the formula r Sparse target value, β is the weight value, r j is the average number of activated neurons in hidden unit j.

[0128] S44: The final output result of the SSDA model is the classification value of the severity of the mechanical fault of the transformer winding.

[0129] The following are examples:

[0130] Reference Figure 1 A transformer winding mechanical fault severity identification method is proposed, including vibration signal acquisition, integrated empirical mode decomposition and correlation calculation, winding mechanical fault identification based on multi-head attention mechanism CLSTM model, and winding mechanical fault severity identification based on SSDA model.

[0131] The vibration signal acquisition is used to acquire the vibration signal generated when the transformer is running.

[0132] The integrated empirical mode decomposition and correlation calculation is used to decompose the transformer vibration signal obtained above and calculate the decomposed IMF and envelope spectrum function, and calculate the envelope spectrum function and the correlation coefficient between it and the original signal.

[0133] The fault identification based on the multi-head attention mechanism CLSTM model is used to identify the obtained relevant envelope spectrum signal to determine whether there is a mechanical fault in the transformer winding and the fault type.

[0134] The SSDA model is used to identify the severity of the fault categories obtained by the CLSTM model.

[0135] Figure 2 This is the structure diagram of the CLSTM model of the present invention, which consists of a CNN convolutional layer and an LSTM structure. The specific form is as follows:

[0136] S21: Construct a CNN convolution layer to extract the correlation between high-dimensional time variables in the relevant envelope spectrum signal. The convolution kernel of the convolution layer is a*b, where a is the width and b is the height. The number of convolution kernels is consistent with the dimension of the relevant envelope spectrum signal. The i-th convolution kernel scans the input relevant envelope spectrum signal R b The resulting output is:

[0137]

[0138] In the formula, RELU(V)=MAX(0,V), Vi represents the weighting coefficient of the input signal, and e represents the bias.

[0139] S22: Construct a basic LSTM unit, including a forget gate, an input gate, and a storage unit:

[0140]

[0141]

[0142]

[0143] In the formula W f is the weight of the forget gate, b f is the bias coefficient, x t express t The input value at the time, W i and b i Represent the weight and bias coefficient of the input gate respectively, Indicates the value of the storage unit currently involved in the calculation, s Represents the sigmoid function:

[0144]

[0145] S23: Establish a CLSTM model and connect the CNN convolutional layer with the LSTM model to ensure that the dependencies between time-related signals are fully mined. The connection method is as follows:

[0146]

[0147] Where P represents the output of the CNN convolutional layer connected to the LSTM model. Represents a concatenation operation.

[0148] Figure 3 This is a schematic diagram of the multi-head attention mechanism structure, its specific form is as follows:

[0149] S31: For the input correlation envelope spectrum signal T, it is assumed to be n-dimensional data with m sampling points, that is:

[0150] T = [ t 1 , t 2 , …, t n ], n ∈R. t n = [ , , …, ] T , m∈R

[0151] Where R represents the real number domain;

[0152] S32: Embed the concat function into each node of the CLSTM model to obtain the combined representation of each node:

[0153]

[0154] S33: T is calculated by n sets of linear transformation matrices m The attention parameter Q m , K m , V m And its single head feature expression head m :

[0155]

[0156] Where Q m , K m , V m Respectively represent T m The attention parameter, head m represents the single-head feature expression of the mth node, softmax is the activation function, Used to avoid dimensionality explosion during calculations.

[0157] S34: output data of each node after multi-head attention enhancement;

[0158]

[0159] In the formula x n represents the original input data of the CLSTM model, Represents the output data after the introduction of the multi-head attention mechanism.

[0160] Figure 4 This is a schematic diagram of the SSDA model structure of the present invention, and its specific form is as follows:

[0161] S41: Let the output data of the CLSTM model be T1, randomly insert Masking noise into the data set T1, and map it to the hidden layer of the DA model:

[0162]

[0163] in i h = { , } is a parameter set, The weight matrix in the mapping process, Bias term vector,s ( x ) = is the sigmoid function, which is used to implement the above mapping, and h represents the data in the hidden layer;

[0164] S42: Reverse mapping is performed on the data of the hidden layer to obtain the reconstructed input data T11, and the loss function value between T1 and T11 is calculated. When the loss function value reaches the minimum value, the DA model is constructed:

[0165]

[0166]

[0167] In the formula i z = { W z , b z} is a parameter set, The weight matrix in the mapping process, Bias term vector. In order to reduce the number of parameters that need to be solved, it is usually W h and W z Make the following restrictions: W h ={ W z} T , T represents transpose; p is the number of original input data, L (x i ,z i ) is the reconstruction error function.

[0168] S43: According to the above steps, multiple DA models are trained in sequence, and the hidden layers of each DA model are extracted and connected in sequence, which is the SDA model. On this basis, a sparse penalty term is added to the last layer of the SDA model to avoid the weight explosion in multiple feature captures. The sparse penalty term formula is as follows:

[0169]

[0170] In the formula r Sparse target value, β is the weight value, r j is the average number of activated neurons in hidden unit j.

[0171] S44: The final output result of the SSDA model is the classification value of the severity of the mechanical fault of the transformer winding.

[0172] The method for identifying the severity of transformer winding mechanical faults of the present invention realizes diagnosis and identification of transformer mechanical faults by analyzing vibration data generated during transformer operation. It has high diagnostic accuracy and short calculation time, and is suitable for large-scale promotion and use in substation operation and maintenance.

[0173] This embodiment uses LSTM, DBN and CLSTM for comparison to verify the effectiveness of the transformer winding mechanical fault severity identification method proposed in the present invention. All data are evenly divided into two groups, one group is training data and the other group is test data, and the fault severity is identified using the SSDA model. The diagnosis results under three different decomposition methods are shown in Table 1. It can be seen from Table 1 that compared with CLSTM, the data obtained by LSTM and DBN decomposition are used for transformer winding mechanical fault severity identification. The result is poor, mainly because LSTM and DBN have poor ability in comprehensive evaluation and trade-off between long time scale features and short time scale features when processing complex signals.

[0174] Table 1 Fault diagnosis results

[0175]

[0176] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0177] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0178] In another embodiment provided in the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute any transformer mechanical fault diagnosis method in the above embodiments.

[0179] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above method.

[0180] The embodiment of the present application also provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus.

[0181] Memory, used to store computer programs;

[0182] The processor is used to implement the above transformer mechanical fault diagnosis method when executing the program stored in the memory.

[0183] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus may be divided into an address bus, a data bus, a control bus, etc.

[0184] The communication interface is used for communication between the above electronic device and other devices.

[0185] The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0186] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0187] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0188] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0189] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0190] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the severity of a transformer winding mechanical fault, characterized in that: The following steps are involved: The vibration signals generated by the transformer operation are collected by sensors, and the collected vibration signals are randomly divided into training data sets and test data sets. The training data sets and test data sets are processed separately using integrated empirical mode decomposition to obtain the corresponding intrinsic mode function IMF and calculate the envelope spectrum signal. Selecting the relevant envelope signal with a correlation degree greater than 0.5 from all envelope spectrum signals based on the correlation coefficient; The CLSTM model enhanced by the multi-head attention mechanism is trained using the relevant envelope signals obtained from the training data set, and the trained CLSTM model is locally fine-tuned using the relevant envelope signals from the test data set. The locally fine-tuned CLSTM model is used to diagnose transformer mechanical faults. Finally, the sparse stacked denoising autoencoder SSDA model is used to identify the severity of the diagnosed faults. Among them, the corresponding intrinsic mode function IMF is obtained and the envelope spectrum signal is calculated, including, First calculate the envelope spectrum function of the signal: Where d i (t) represents the ith IMF function, H[d i (t)] represents the Hilbert spectrum signal of the i-th IMF function, B j (t) represents the jth envelope spectrum signal; The correlation calculation formula is used to calculate the correlation between each envelope spectrum and the original signal, and the correlation coefficient above 0.5 is selected as the correlation envelope spectrum function: In the formula, s e express B e (t) and x ( t ) between them; Finally, the sparse stacked denoising autoencoder SSDA model is used to identify the severity of the diagnosed fault. The output data obtained by the CLSTM model enhanced by the multi-head attention mechanism is used as the input data of the sparse stacked denoising autoencoder SSDA model. The sparse characteristics of the sparse stacked denoising autoencoder SSDA model and the global feature mining and comparison capabilities are used to identify the severity of the fault. The specific steps are as follows: S41: Let the output data of the CLSTM model enhanced by the multi-head attention mechanism be T1, randomly insert Masking noise into the data set T1, and map it to the hidden layer of the DA model: in θ h = { , } is a parameter set, The weight matrix in the mapping process, Bias term vector, s ( x ) = is the sigmoid function, which is used to implement the above mapping, and h represents the data in the hidden layer; S42: Reverse mapping is performed on the data of the hidden layer to obtain the reconstructed input data T11, and the loss function value between T1 and T11 is calculated. When the loss function value reaches the minimum value, the DA model is constructed: In the formula θ z = { W z , b z } is a parameter set, The weight matrix in the mapping process, Bias term vector; in order to reduce the number of parameters that need to be solved, W h and W z Make the following restrictions: W h ={ W z } T , T represents transpose; p is the number of original input data, L (x i , z i ) is the reconstruction error function; S43: Train multiple DA models in sequence according to the above steps, extract the hidden layers of each DA model and connect them in sequence, which is the SDA model. On this basis, add a sparse penalty term to the last layer of the SDA model, which is the SSDA model, to avoid the weight explosion in multiple feature captures. The sparse penalty term formula is as follows: In the formula ρ Sparse target value, β is the weight value, ρ j is the average number of activated neurons in hidden unit j; S44: The final output result of the SSDA model is the classification value of the severity of the mechanical fault of the transformer winding.

2. The method for identifying the severity of a transformer winding mechanical fault according to claim 1, characterized in that: The steps of integrated empirical mode decomposition include: Assume the original vibration signal of the transformer is x(t) , adding noise to the original model to obtain the reconstructed signal: in represents the signal after adding noise, represents the mth noise component; The reconstructed signal is processed using empirical mode decomposition to obtain the first IMF function and the corresponding residual term: Where r1 represents the residual term, d1 represents the first IMF function, T {.} indicates the empirical mode decomposition calculation; Repeat the above steps until the decomposition is complete.

3. The method for identifying the severity of transformer winding mechanical fault according to claim 1, characterized in that: The CLSTM model enhanced by the multi-head attention mechanism is trained and locally fine-tuned, including dividing the original vibration signal into a training data set and a test data set in a 5:1 ratio, and then using the relevant envelope spectrum signal of the training data set to train the CLSTM model enhanced by the multi-head attention mechanism; After training, the CLSTM model is further verified using the relevant envelope spectrum signals of the test data set, and local fine-tuning is performed based on the verification results until the best fault diagnosis and identification effect is achieved.

4. The method for identifying the severity of transformer winding mechanical fault according to claim 1, characterized in that: The CLSTM model enhanced by the multi-head attention mechanism is used to realize transformer mechanical fault identification based on the relevant envelope spectrum signal, and includes the following steps: S21: Construct a CNN convolution layer to extract the correlation between high-dimensional time variables in the relevant envelope spectrum signal. The convolution kernel of the convolution layer is a*b, where a is the width and b is the height. The number of convolution kernels is consistent with the dimension of the relevant envelope spectrum signal. The i-th convolution kernel scans the input relevant envelope spectrum signal R b The resulting output is: Where RELU(V)=MAX(0,V), V i represents the weighting coefficient of the input signal, and e represents the bias; S22: Construct a basic LSTM unit, including a forget gate, an input gate, and a storage unit: In the formula W f is the weight of the forget gate, b f is the bias coefficient, x t express t The input value at the time, W i and b i Represent the weight and bias coefficient of the input gate respectively, Indicates the value of the storage unit currently involved in the calculation, σ Represents the sigmoid function: S23: Establish a CLSTM model and connect the CNN convolutional layer with the LSTM model to ensure that the dependencies between time-related signals are fully mined. The connection method is as follows: Where P represents the output of the CNN convolutional layer connected to the LSTM model. Represents a splicing operation; S24: The obtained CLSTM model is enhanced by using a multi-head attention mechanism, and a multi-head attention layer is added to the LSTM model to extract the global features of the input related envelope spectrum signal.

5. The method for identifying the severity of transformer winding mechanical fault according to claim 4, characterized in that: The multi-head attention mechanism, based on the CLSTM model mining the interdependent correlation of each time component, enhances the CLSTM model's learning of the global features of the input related envelope spectrum signal. The specific steps are as follows: S31: For the input correlation envelope spectrum signal T, it is assumed to be n-dimensional data with m sampling points, that is: T = [ t 1 , t 2 , …, t n ], n ∈R. t n = [ , , …, ] T , m ∈R Where R represents the real number domain; S32: Embed the concat function into each node of the CLSTM model to obtain the combined representation of each node: S33: T is calculated by n sets of linear transformation matrices m The attention parameter Q m , K m , V m And its single head feature expression head m : Where Q m , K m , V m Respectively represent T m The attention parameter, head m represents the single-head feature expression of the mth node, softmax is the activation function, Used to avoid dimensionality explosion during calculation; S34: output data of each node after multi-head attention enhancement; In the formula x n represents the original input data of the CLSTM model, Represents the output data after the introduction of the multi-head attention mechanism.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 5.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 5.

Images