Transformer health state assessment method and system

Abstract

The invention provides a transformer health state assessment method and system, and the method comprises the steps: obtaining the multi-source operation data of a target transformer, generating a multi-criterion discrimination value sequence, carrying out the deep feature obtaining of the operation features of the transformer obtained by the sequence, and obtaining the operation deep features of the transformer, obtaining a criterion weight sequence under the weight transformation matrix and the bias term according to the transformer operation deep characteristics; generating a health state for maintenance according to the multi-criterion discriminant value sequence and the criterion weight sequence; the deep feature acquisition step comprises the following steps: performing multi-head self-attention processing according to transformer operation features to obtain an operation feature incidence matrix, and obtaining preliminary transformer deep features according to the incidence matrix and the transformer operation features; and repeatedly executing the deep feature acquisition step according to the preliminary transformer deep features until the execution times are not lower than the coding level, and taking the preliminary transformer deep features as transformer operation deep features. According to the technical scheme, the stability of transformer health state evaluation can be improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method and system for assessing the health status of transformers. Background Technology

[0002] The reliability of transformers directly determines the regional power supply quality and the grid's ability to withstand disturbances. However, with the increasing number of transformers in the power system, their expanding distribution range, and the increasing operating pressure, accurate assessment of transformer health status has become a key prerequisite for ensuring the safe and stable operation of the power grid.

[0003] To improve the comprehensiveness of transformer health status assessment, existing technologies utilize multiple metrics to jointly learn the correlations between variables and construct graph-structured data. A dual-input residual graph convolutional network based on Chebyshev graph convolution is then used to extract features from the graph-structured data, followed by feature fusion through a self-attention mechanism to obtain the transformer's status assessment results. Another existing technology employs a bidirectional gated neural network to extract important feature vectors for different modes from the textual information, frequency domain graph, and infrared image of the vibration signal. Subsequently, a cross-attention mechanism is used to establish connections between different modes, and after feature vector fusion, the fault status of the power transformer is output through convolutional layers and fully connected layers.

[0004] In the aforementioned existing technologies, the self-attention mechanism in the dual-input residual graph convolutional network is only used for the final fusion of the extracted graph features, and the bidirectional gated neural network and cross-attention mechanism are only used to fuse feature vectors of different modalities. Therefore, the existing technologies lack the process of adaptively adjusting the criterion weights based on the inherent connection of the multi-criteria operating data, and rely on a large amount of outdated historical fault data for supervised training. As a result, when facing the complex and ever-changing continuous environmental fluctuations in actual operating scenarios, the existing technologies have difficulty distinguishing between real transformer condition deterioration and instantaneous interference, which greatly reduces the stability of transformer health status assessment. Summary of the Invention

[0005] The present invention aims to provide a method and system for assessing the health status of transformers, in order to solve the above-mentioned technical problems, suppress external operating environment fluctuations and interference, and improve the stability of transformer health status assessment.

[0006] To address the aforementioned technical problems, this invention provides a method for assessing the health status of a transformer, comprising the following steps: Acquire multi-source operating data of the target transformer under continuous differential protection cycles, and generate a multi-criteria discriminant value sequence based on the multi-source operating data; Based on the multi-criteria discriminant value sequence, the transformer operation characteristics are obtained, and based on the transformer operation characteristics, a deep feature acquisition step is performed to obtain the deep features of transformer operation. The deep operating features of the transformer are pooled and compressed to obtain pooled operating features, and a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term. The state evaluation score is obtained by weighting the multi-criteria discriminant value sequence and the criterion weight sequence. A health status is generated based on the status assessment score, and maintenance is performed based on the health status. The deep feature acquisition steps include: Based on the transformer operating characteristics, an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix are obtained under a preset linear transformation matrix set. Multi-head self-attention processing is then performed based on the operating characteristic query matrix, the operating characteristic key matrix, and the operating characteristic value matrix to obtain an operating characteristic correlation matrix. Based on the operational feature correlation matrix and the transformer operational features, feedforward network processing is performed under a preset feature transformation matrix to obtain preliminary deep transformer features. If the number of times the deep feature acquisition step is executed is less than the preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped and the preliminary transformer deep features are used as transformer operating deep features.

[0007] In the above scheme, transformer operating characteristics are obtained based on a multi-criteria discriminant value sequence. By repeatedly executing deep feature acquisition steps involving multi-head self-attention and feedforward network processing at a preset encoding level, the scheme can mine and strengthen the stable and inherent intrinsic correlations between different criteria from the transformer operating characteristics, thereby obtaining deep transformer operating characteristics. When external environmental fluctuations cause a brief abnormal change in a certain criterion, this scheme can effectively determine the inconsistency between the abnormal change and the aforementioned intrinsic correlation, effectively distinguishing the true health state of the transformer from the instantaneous noise caused by the incoordination of operating data between criteria due to external environmental fluctuations. This suppresses interference from external operating environment fluctuations, thereby improving the stability of subsequent transformer health status assessments. Furthermore, the above scheme uses the deep transformer operating characteristics obtained after filtering external operating environment fluctuation interference to generate a criterion weight sequence. When the data of a certain criterion changes abruptly due to brief environmental interference rather than actual degradation, this scheme can dynamically adjust the criterion weight sequence according to the deep transformer operating characteristics, reducing the impact of the operating data corresponding to that criterion on the health assessment results and improving the stability of the transformer health status assessment.

[0008] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: performing a criterion weight acquisition step based on the deep operating features of the transformer to obtain a criterion weight sequence; the criterion weight acquisition step includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function; if it is confirmed that the operating feature similarity loss values ​​do not meet the preset convergence condition, then unsupervised training is performed based on the operating feature similarity loss values ​​to obtain an updated linear transformation matrix and an updated feature transformation matrix, and a deep feature acquisition step is performed based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features to update the deep operating features of the transformer, and the criterion weight acquisition step is re-executed based on the updated deep operating features of the transformer; otherwise, a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term.

[0009] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function, includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and performing cosine similarity calculation based on the pooled operating features to obtain feature similarity; and generating operating feature similarity loss values ​​based on the feature similarity under a preset loss function.

[0010] In the above scheme, the deep features of transformer operation are pooled and compressed into pooled operation features. Based on these, feature similarity and operation feature similarity loss values ​​are calculated. Unsupervised training is then conducted with the operation feature similarity loss value satisfying a preset convergence condition as the training objective. After continuously executing the deep feature acquisition steps, the final result is a deep feature of transformer operation with high similarity in operation features at adjacent times. This ensures that the final pooled operation features can stably reflect the inherent consistency of the multi-source transformer operation data, filtering out abnormal transformer operation features that are discontinuous in time due to random environmental interference. Ultimately, a criterion weight sequence is generated that can more accurately quantify the relative importance of each criterion under the current environment. Therefore, even if individual criteria in the multi-source operation data are affected by environmental interference, the misjudgment rate of the transformer health status generated subsequently based on the multi-criterion discriminant value sequence and the aforementioned criterion weight sequence is reduced, improving the accuracy of transformer health status assessment.

[0011] Further, the step of acquiring multi-source operating data of the target transformer under continuous differential protection cycles and generating a multi-criteria discriminant value sequence based on the multi-source operating data includes: acquiring multi-source operating data of the target transformer under continuous differential protection cycles; the multi-source operating data includes a sequence of oil gas concentration, a sequence of charged particle numbers, a sequence of vibration signals, a sequence of monitoring images, and a sequence of transformer secondary currents; obtaining a relative gas production rate sequence based on the oil gas concentration sequence, and obtaining a sequence of oil gas concentration discriminant values ​​based on the relative gas production rate sequence at a preset gas production rate threshold; obtaining a dielectric loss factor sequence based on the charged particle number sequence, and obtaining a sequence of dielectric loss factor discriminant values ​​based on the dielectric loss factor sequence at a preset loss threshold; acquiring the number of vibration breakthroughs in the vibration signal sequence where the amplitude of each vibration signal exceeds a preset maximum vibration threshold, and integrating the vibration breakthroughs to obtain... The vibration breakthrough number sequence is used to obtain a tap changer discrimination value sequence under a preset breakthrough number threshold. An initial abnormal region sequence is identified from the monitoring image sequence using image difference method. The initial abnormal region sequence is then denoised and segmented to obtain a denoised abnormal region sequence. Color feature determination is performed based on the initial abnormal region sequence and the denoised abnormal region sequence to obtain an oil leakage discrimination value sequence. A differential current sequence is obtained by normalizing the secondary current sequence of the current transformer, and a differential current distance sequence is generated based on the differential current sequence under a preset standard sine wave sequence. A differential protection discrimination value sequence is obtained under a preset distance threshold based on the differential current distance sequence. The oil gas concentration discrimination value sequence, the medium loss factor discrimination value sequence, the tap changer discrimination value sequence, the oil leakage discrimination value sequence, and the differential protection discrimination value sequence are used as a multi-criteria discrimination value sequence.

[0012] The multi-source operational data collected by the above scheme includes oil gas concentration sequences, charged particle quantity sequences, vibration signal sequences, monitoring image sequences, and transformer secondary current sequences, which are sensitive to long-term operational degradation but not to short-term environmental fluctuations. Therefore, when the operational characteristic corresponding to a certain criterion is abnormal due to a brief environmental disturbance such as a non-fault impact causing a momentary exceedance of vibration, other criteria unaffected by this disturbance will provide stable operational characteristics. This allows for the identification of deep-seated transformer operational characteristics that suppress external environmental fluctuations, thereby suppressing the interference of the operational data corresponding to the abnormal criteria on the final state assessment score and ensuring the stability of the transformer health status assessment.

[0013] Further, the step of obtaining transformer operating characteristics based on the multi-criteria discriminant value sequence and performing a deep feature acquisition step based on the transformer operating characteristics to obtain deep transformer operating characteristics includes: obtaining the oil leakage discriminant value sequence and the dielectric loss factor discriminant value sequence from the multi-criteria discriminant value sequence; aligning and integrating the oil leakage discriminant value sequence, the dielectric loss factor discriminant value sequence, the relative gas generation rate sequence, the vibration breakthrough number sequence, and the differential current distance sequence to obtain transformer operating characteristics; and performing a deep feature acquisition step based on the transformer operating characteristics to obtain deep transformer operating characteristics.

[0014] Further, the step of obtaining an operating feature query matrix, an operating feature key matrix, and an operating feature value matrix based on the transformer operating characteristics under a preset linear transformation matrix set, and performing multi-head self-attention processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain an operating feature association matrix includes: obtaining a query linear change weight matrix, a key linear change weight matrix, and a value linear change weight matrix from the preset linear transformation matrix set; obtaining an operating feature query matrix based on the query linear change weight matrix and the transformer operating characteristics; obtaining an operating feature key matrix based on the key linear change weight matrix and the transformer operating characteristics; obtaining an operating feature value matrix based on the value linear change weight matrix and the transformer operating characteristics; performing projection mapping processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain a sequence of projected operating feature query matrices, a sequence of projected operating feature key matrices, and a sequence of projected operating feature value matrices; and performing multi-head self-attention processing on the sequence of projected operating feature query matrices, the sequence of projected operating feature key matrices, and the sequence of projected operating feature value matrices to obtain an operating feature association matrix.

[0015] In the above scheme, the transformer operating characteristics are first transformed into an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix by linearly changing the weight matrix set. Then, after projecting and mapping these three matrices, multi-head self-attention processing is performed, enabling the scheme to explore and mine multiple possible intrinsic correlations between various criteria in parallel. When external environmental fluctuations cause temporary interference to certain criteria, this interference may only abnormally affect specific correlations of some self-attention heads, while other self-attention heads can still capture stable correlations from the operating data unaffected by the interference to dilute the abnormal interference. This ensures the stability of the final operating characteristic correlation matrix, thereby ensuring the generation of a stable and reliable criterion weight sequence and improving the stability of transformer health status assessment.

[0016] Further, the step of performing feedforward network processing based on the operational feature correlation matrix and the transformer operational features under a preset feature transformation matrix to obtain preliminary transformer deep features includes: performing residual connection and layer normalization based on the operational feature correlation matrix and the transformer operational features to obtain preliminary operational features; performing feedforward network processing based on the preliminary operational features under a preset feature transformation matrix to obtain preprocessed transformer operational deep features; and performing residual connection and layer normalization based on the preprocessed transformer operational deep features and the preliminary operational features to obtain preliminary transformer deep features.

[0017] Furthermore, in the process of generating a health status based on the status assessment score and performing maintenance based on the health status, generating a health status based on the status assessment score includes: confirming that the status assessment score is within a preset health score range, then generating a health status set as healthy; otherwise, determining whether the status assessment score is within a preset minor fault score range; confirming that the status assessment score is within the preset minor fault score range, then generating a health status set as minor fault; otherwise, determining whether the status assessment score is within a preset moderate fault score range; confirming that the status assessment score is within the preset moderate fault score range, then generating a health status set as moderate fault; otherwise, generating a health status set as severe fault.

[0018] This invention also provides a transformer health status assessment system for implementing any of the above transformer health status assessment methods, comprising: a discriminant value generation module for acquiring multi-source operating data of a target transformer under continuous differential protection cycles, and generating a multi-criteria discriminant value sequence based on the multi-source operating data; a deep feature acquisition module for obtaining transformer operating features based on the multi-criteria discriminant value sequence, and performing a deep feature acquisition step based on the transformer operating features to obtain deep transformer operating features; a criterion weight acquisition module for pooling and compressing the deep transformer operating features to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term; a status assessment module for weighting the multi-criteria discriminant value sequence and the criterion weight sequence to obtain a status assessment score; and a health status determination module for determining the health status based on the... The health status is generated by the state assessment score, and maintenance is carried out based on the health status. The deep feature acquisition step includes: obtaining an operating feature query matrix, an operating feature key matrix, and an operating feature value matrix based on the transformer operating features under a preset linear transformation matrix set; performing multi-head self-attention processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain an operating feature association matrix; performing feedforward network processing based on the operating feature association matrix and the transformer operating features under a preset feature transformation matrix to obtain preliminary transformer deep features; if it is confirmed that the number of times the deep feature acquisition step is executed is less than a preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped and the preliminary transformer deep features are used as transformer operating deep features.

[0019] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: performing a criterion weight acquisition step based on the deep operating features of the transformer to obtain a criterion weight sequence; the criterion weight acquisition step includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function; if it is confirmed that the operating feature similarity loss values ​​do not meet the preset convergence condition, then unsupervised training is performed based on the operating feature similarity loss values ​​to obtain an updated linear transformation matrix and an updated feature transformation matrix, and a deep feature acquisition step is performed based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features to update the deep operating features of the transformer, and the criterion weight acquisition step is re-executed based on the updated deep operating features of the transformer; otherwise, a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term.

[0020] In the above scheme, transformer operating characteristics are obtained based on a multi-criteria discriminant value sequence. By repeatedly executing deep feature acquisition steps involving multi-head self-attention and feedforward network processing at a preset encoding level, stable and inherent intrinsic correlations between different criteria can be extracted and strengthened from the transformer operating characteristics. This allows for the acquisition of deep transformer operating characteristics. When external environmental fluctuations cause a brief abnormal change in a certain criterion, the scheme can effectively determine the inconsistency between this abnormal change and the aforementioned intrinsic correlation, effectively distinguishing the true health state of the transformer from transient noise caused by external environmental fluctuations. This suppresses interference from external operating environment fluctuations and improves the stability of subsequent transformer health status assessments. Furthermore, when the data of a certain criterion undergoes a sudden change due to brief environmental interference rather than actual degradation, the scheme can dynamically adjust the criterion weight sequence based on the deep transformer operating characteristics, reducing the impact of the operating data corresponding to that criterion on the health assessment results and improving the stability of the transformer health status assessment. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the technical implementation of a transformer health status assessment method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a transformer health status assessment system architecture provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This embodiment provides a method for assessing the health status of a transformer, including the following steps: Step S1: Obtain multi-source operation data of the target transformer under continuous differential protection cycle, and generate a multi-criteria discriminant value sequence based on the multi-source operation data; Step S2: Obtain transformer operating characteristics based on the multi-criteria discriminant value sequence, and perform a deep feature acquisition step based on the transformer operating characteristics to obtain deep transformer operating characteristics; Step S3: Pool the deep operating features of the transformer to obtain pooled operating features, and obtain the criterion weight sequence based on the pooled operating features under the preset weight transformation matrix and preset bias term; Step S4: Weight the results based on the multi-criteria discriminant value sequence and the criterion weight sequence to obtain the state evaluation score; Step S5: Generate a health status based on the status assessment score, and perform maintenance according to the health status; The deep feature acquisition steps include: Based on the transformer operating characteristics, an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix are obtained under a preset linear transformation matrix set. Multi-head self-attention processing is then performed based on the operating characteristic query matrix, the operating characteristic key matrix, and the operating characteristic value matrix to obtain an operating characteristic correlation matrix. Based on the operational feature correlation matrix and the transformer operational features, feedforward network processing is performed under a preset feature transformation matrix to obtain preliminary deep transformer features. If the number of times the deep feature acquisition step is executed is less than the preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped and the preliminary transformer deep features are used as transformer operating deep features.

[0024] In the above embodiments, transformer operating characteristics are obtained based on a multi-criteria discriminant value sequence. By repeatedly executing deep feature acquisition steps involving multi-head self-attention and feedforward network processing at a preset encoding level, stable and inherent intrinsic correlations between different criteria can be extracted and strengthened from the transformer operating characteristics, thus obtaining deep transformer operating characteristics. In this embodiment, when external environmental fluctuations cause a brief abnormal change in a certain criterion, the inconsistency between this abnormal change and the aforementioned intrinsic correlation can be effectively determined. This effectively distinguishes the true health state of the transformer from the instantaneous noise caused by the incoordination of operating data between criteria due to external environmental fluctuations, thereby suppressing interference from external operating environment fluctuations and improving the stability of subsequent transformer health status assessments. Furthermore, the above embodiments utilize the deep transformer operating characteristics obtained after filtering external operating environment fluctuation interference to generate a criterion weight sequence. When the data of a certain criterion undergoes a sudden change due to brief environmental interference rather than actual degradation, this embodiment can dynamically adjust the corresponding criterion weights in the criterion weight sequence based on the deep transformer operating characteristics, reducing the impact of the operating data corresponding to that criterion on the health assessment results and improving the stability of the transformer health status assessment.

[0025] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: performing a criterion weight acquisition step based on the deep operating features of the transformer to obtain a criterion weight sequence; the criterion weight acquisition step includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function; if it is confirmed that the operating feature similarity loss values ​​do not meet the preset convergence condition, then unsupervised training is performed based on the operating feature similarity loss values ​​to obtain an updated linear transformation matrix and an updated feature transformation matrix, and a deep feature acquisition step is performed based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features to update the deep operating features of the transformer, and the criterion weight acquisition step is re-executed based on the updated deep operating features of the transformer; otherwise, a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term.

[0026] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function, includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and performing cosine similarity calculation based on the pooled operating features to obtain feature similarity; and generating operating feature similarity loss values ​​based on the feature similarity under a preset loss function.

[0027] In the above embodiments, the deep features of transformer operation are pooled and compressed into pooled operation features. Based on these pooled features, feature similarity and operation feature similarity loss values ​​are calculated. Unsupervised training is performed with the operation feature similarity loss value satisfying a preset convergence condition as the training objective. After continuously executing the deep feature acquisition steps, the resulting deep features of transformer operation exhibit high similarity in operation features at adjacent times. This ensures that the final pooled operation features stably reflect the inherent consistency of the multi-source transformer operation data, filtering out abnormal transformer operation features that are discontinuous in time due to random environmental interference. Ultimately, a criterion weight sequence is generated that more accurately quantifies the relative importance of each criterion under the current environment. Therefore, even if individual criteria in the multi-source operation data are affected by environmental interference, the misjudgment rate of the transformer health status generated subsequently based on the multi-criterion discriminant value sequence and the aforementioned criterion weight sequence is reduced, improving the accuracy of transformer health status assessment.

[0028] Further, the step of acquiring multi-source operating data of the target transformer under continuous differential protection cycles and generating a multi-criteria discriminant value sequence based on the multi-source operating data includes: acquiring multi-source operating data of the target transformer under continuous differential protection cycles; the multi-source operating data includes a sequence of oil gas concentration, a sequence of charged particle numbers, a sequence of vibration signals, a sequence of monitoring images, and a sequence of transformer secondary currents; obtaining a relative gas production rate sequence based on the oil gas concentration sequence, and obtaining a sequence of oil gas concentration discriminant values ​​based on the relative gas production rate sequence at a preset gas production rate threshold; obtaining a dielectric loss factor sequence based on the charged particle number sequence, and obtaining a sequence of dielectric loss factor discriminant values ​​based on the dielectric loss factor sequence at a preset loss threshold; acquiring the number of vibration breakthroughs in the vibration signal sequence where the amplitude of each vibration signal exceeds a preset maximum vibration threshold, and integrating the vibration breakthroughs to obtain... The vibration breakthrough number sequence is used to obtain a tap changer discrimination value sequence under a preset breakthrough number threshold. An initial abnormal region sequence is identified from the monitoring image sequence using image difference method. The initial abnormal region sequence is then denoised and segmented to obtain a denoised abnormal region sequence. Color feature determination is performed based on the initial abnormal region sequence and the denoised abnormal region sequence to obtain an oil leakage discrimination value sequence. A differential current sequence is obtained by normalizing the secondary current sequence of the current transformer, and a differential current distance sequence is generated based on the differential current sequence under a preset standard sine wave sequence. A differential protection discrimination value sequence is obtained under a preset distance threshold based on the differential current distance sequence. The oil gas concentration discrimination value sequence, the medium loss factor discrimination value sequence, the tap changer discrimination value sequence, the oil leakage discrimination value sequence, and the differential protection discrimination value sequence are used as a multi-criteria discrimination value sequence.

[0029] The multi-source operational data collected in the above embodiments include oil gas concentration sequences, charged particle quantity sequences, vibration signal sequences, monitoring image sequences, and transformer secondary current sequences, which are sensitive to long-term operational degradation but not to short-term environmental fluctuations. Therefore, when an operational characteristic corresponding to a certain criterion is abnormal due to a brief environmental disturbance such as a non-fault impact causing a momentary exceedance of vibration, other criteria unaffected by this disturbance will provide stable operational characteristics. This allows for the identification of deep-seated transformer operational characteristics that suppress external environmental fluctuations, thereby suppressing the interference of operational data corresponding to abnormal criteria on the final state assessment score and ensuring the stability of the transformer health status assessment.

[0030] Further, the step of obtaining transformer operating characteristics based on the multi-criteria discriminant value sequence and performing a deep feature acquisition step based on the transformer operating characteristics to obtain deep transformer operating characteristics includes: obtaining the oil leakage discriminant value sequence and the dielectric loss factor discriminant value sequence from the multi-criteria discriminant value sequence; aligning and integrating the oil leakage discriminant value sequence, the dielectric loss factor discriminant value sequence, the relative gas generation rate sequence, the vibration breakthrough number sequence, and the differential current distance sequence to obtain transformer operating characteristics; and performing a deep feature acquisition step based on the transformer operating characteristics to obtain deep transformer operating characteristics.

[0031] Further, the step of obtaining an operating feature query matrix, an operating feature key matrix, and an operating feature value matrix based on the transformer operating characteristics under a preset linear transformation matrix set, and performing multi-head self-attention processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain an operating feature association matrix includes: obtaining a query linear change weight matrix, a key linear change weight matrix, and a value linear change weight matrix from the preset linear transformation matrix set; obtaining an operating feature query matrix based on the query linear change weight matrix and the transformer operating characteristics; obtaining an operating feature key matrix based on the key linear change weight matrix and the transformer operating characteristics; obtaining an operating feature value matrix based on the value linear change weight matrix and the transformer operating characteristics; performing projection mapping processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain a sequence of projected operating feature query matrices, a sequence of projected operating feature key matrices, and a sequence of projected operating feature value matrices; and performing multi-head self-attention processing on the sequence of projected operating feature query matrices, the sequence of projected operating feature key matrices, and the sequence of projected operating feature value matrices to obtain an operating feature association matrix.

[0032] In the above embodiments, the transformer operating characteristics are first transformed into an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix by linearly changing the weight matrix set. Then, after projection mapping of these three matrices, multi-head self-attention processing is performed, enabling this embodiment to explore and mine multiple possible intrinsic correlations between various criteria in parallel. When external environmental fluctuations cause temporary interference to certain criteria, this interference may only abnormally affect specific correlations of some self-attention heads, while other self-attention heads can still capture stable correlations from the operating data unaffected by the interference to dilute the abnormal interference. This ensures the stability of the final operating characteristic correlation matrix, thereby ensuring the generation of a stable and reliable criterion weight sequence and improving the stability of transformer health status assessment.

[0033] Further, the step of performing feedforward network processing based on the operational feature correlation matrix and the transformer operational features under a preset feature transformation matrix to obtain preliminary transformer deep features includes: performing residual connection and layer normalization based on the operational feature correlation matrix and the transformer operational features to obtain preliminary operational features; performing feedforward network processing based on the preliminary operational features under a preset feature transformation matrix to obtain preprocessed transformer operational deep features; and performing residual connection and layer normalization based on the preprocessed transformer operational deep features and the preliminary operational features to obtain preliminary transformer deep features.

[0034] Furthermore, in the process of generating a health status based on the status assessment score and performing maintenance based on the health status, generating a health status based on the status assessment score includes: confirming that the status assessment score is within a preset health score range, then generating a health status set as healthy; otherwise, determining whether the status assessment score is within a preset minor fault score range; confirming that the status assessment score is within the preset minor fault score range, then generating a health status set as minor fault; otherwise, determining whether the status assessment score is within a preset moderate fault score range; confirming that the status assessment score is within the preset moderate fault score range, then generating a health status set as moderate fault; otherwise, generating a health status set as severe fault.

[0035] In one embodiment, a sampling period is defined as one cycle of the target transformer during differential protection, and multi-source operating data of the target transformer under continuous differential protection cycles is acquired; the multi-source operating data includes a sequence of gas concentrations in the oil. , number sequence of charged particles Vibration signal sequence Monitoring image sequence and the secondary current sequence of the transformer ;in, For the first Gas concentration in oil for each sampling period For the first The number of charged particles per sampling period For the first Vibration signal for one sampling period, For the first Monitoring images for each sampling period, For the first The secondary current of the current transformer in one sampling period.

[0036] Based on the gas concentration sequence in oil The relative gas production rate is calculated by considering the gas concentration in the oil between adjacent sampling periods. The formula for calculating the relative gas production rate for one sampling period is as follows: ; in, This represents the relative gas production rate for the (i+1)th sampling period. After obtaining the relative gas production rate sequence based on the gas concentration sequence in the oil, the gas production rate threshold is... Below, the relative gas production rate sequence is obtained where the relative gas production rate is greater than the gas production rate threshold. The percentage of the number of [a certain number] relative to the total number of relative gas production rates in the relative gas production rate sequence. Subsequently, the maximum threshold value for the percentage of relative gas production rates out of the total number was determined. ,in accordance with and The discriminant value sequence of gas concentration in oil was calculated, and the formula for calculating the discriminant value sequence of gas concentration in oil is as follows: ; in, This is a sequence of discriminant values ​​for gas concentration in oil.

[0037] Based on the sequence of charged particle numbers The volumetric conductivity of the medium is calculated based on the number of charged particles in each sampling period. The formula for calculating the volumetric conductivity of the medium in one sampling period is: ; in, For the first The volumetric conductivity of the dielectric in each sampling period For charge quantity, Let be the mobility. After obtaining the dielectric volume conductivity sequence, the formula for calculating the dielectric loss factor is as follows: ; in, For the first Dielectric loss factor per sampling period, For the first The dielectric loss angle per sampling period, This is the loss term corresponding to the complex relative permittivity. For the complex relative permittivity, the corresponding capacitance term is... is the vacuum dielectric constant. Subsequently, based on the obtained dielectric loss factor sequence, within a preset loss threshold... The formula for obtaining the dielectric loss factor sequence is as follows: ; in, For the first The dielectric loss factor discriminant value is obtained from each sampling period, ultimately resulting in a dielectric factor discriminant value sequence. .

[0038] Acquire vibration signal sequence Among the various vibration signals, the amplitude exceeds the preset maximum vibration threshold. Vibration breakthrough number And based on the number of vibration breakthroughs The vibration breakthrough number sequence is obtained by integration; based on the vibration breakthrough number sequence, a preset breakthrough number threshold is reached. The following is a sequence of tap changer discrimination values, and the formula for calculating the tap changer discrimination value is: ; in, For the first Each sampling period tap changer discrimination value, and the final tap changer discrimination value sequence. .

[0039] For monitoring image sequences The initial abnormal regions in each monitoring image are identified using the image difference method. These initial abnormal regions are then denoised and segmented to obtain a denoised abnormal region sequence. The color characteristics of the initial abnormal region and the denoised abnormal region are compared. If the gray-brown portion shows a significant change, an oil leak is considered to exist. The formula for determining the oil leak discrimination value is thus derived as follows: ; in, For the first Each sampling period yields an oil leakage discrimination value, ultimately resulting in an oil leakage discrimination value sequence. .

[0040] Based on the secondary current sequence of the mutual inductor Normalization is performed to obtain the differential current sequence corresponding to each sampling period. , The total number of sampling points in each sampling period and based on the differential sequence A in a preset standard sine wave sequence. The Hausdorff distance is then used to calculate the differential distance. The formula for calculating the differential distance is as follows: ; in: ; ; in, For the first Differential current distance for each sampling period. The complement of the differential current distance is obtained based on this differential current distance. The formula for calculating the complement of the differential current distance is: ; in, For the first Differential distance complement for each sampling period. Based on the differential distance complement at the distance threshold. The differential protection discrimination value is obtained, and the corresponding calculation formula is as follows: ; in, For the first Differential protection discrimination values ​​for each sampling period are obtained, ultimately yielding a differential protection discrimination value sequence. Subsequently, this embodiment uses a sequence of gas concentration discrimination values ​​in the oil. Dielectric loss factor discriminant value sequence Tap changer discrimination value sequence Oil leakage discrimination value sequence differential protection discrimination value sequence As a sequence of discriminant values ​​based on multiple criteria.

[0041] In the above embodiments, the leakage oil discrimination value sequence is... The dielectric loss factor discrimination value sequence The transformer operating characteristics are obtained by aligning and integrating the relative gas production rate sequence, vibration breakthrough number sequence, and differential current distance sequence. ,and: ; In this formula, For the first Discrimination value of dielectric loss factor within each sampling period For the first The number of times the vibration signal exceeds the threshold within a sampling period For the first The oil leakage discrimination value within each sampling period For the first The complement of the differential distance within each sampling period.

[0042] Transformer operating characteristics In the Column vector of each sampling period Defined as the first One sample. This embodiment is based on the operating characteristics of transformers. Perform the deep feature acquisition step to obtain the deep operational features of the transformer. In-depth characteristics of transformer operation Pooling compression is performed to obtain pooled runtime features. Cosine similarity is then calculated based on these features to obtain feature similarity. Finally, runtime feature similarity loss values ​​are generated under a preset loss function. The relevant formula for generating the runtime feature similarity loss values ​​is as follows: ; ; in, To calculate the loss value based on the similarity of the running features, This is the cosine similarity calculation function. Deep characteristics of transformer operation After pooling compression The feature vector corresponding to each sample, and for After pooling compression The set of positive samples corresponding to each sample. for After pooling compression The set of negative samples corresponding to each sample. For preset temperature parameters, Deep characteristics of transformer operation The total number of samples in the sample, This embodiment constructs a contrastive loss function based on temporal consistency. In the above embodiment, since the health state of the transformer is difficult to quantify and is not suitable as the target of the loss function, a contrastive loss function based on temporal consistency is constructed to build meaningful feature representations by learning the similarity of features of samples at adjacent time points.

[0043] This embodiment uses The minimum value is reduced to the convergence condition, confirming the similarity loss value of the running characteristics. If the convergence condition is not met, unsupervised training is performed based on the similarity loss value of the operating features to obtain the updated linear transformation matrix and the updated feature transformation matrix. Then, based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features... Perform the deep feature acquisition step to update the transformer's deep operating features. Based on the updated deep operating characteristics of the transformer Repeat the criterion weight acquisition step, i.e., the above-mentioned deep characteristics of transformer operation. Pooling compression is performed to obtain pooled runtime features. Cosine similarity is then calculated based on these features to obtain feature similarity. Finally, runtime feature similarity loss values ​​are generated under a preset loss function. The process of determining whether the runtime feature similarity loss values ​​meet the convergence condition is then repeated until the runtime feature similarity loss values ​​are confirmed. The convergence condition is met, that is Once the value drops to the minimum, the criterion weight acquisition step is stopped.

[0044] After stopping the criterion weight acquisition step, the above embodiment obtains the criterion weight sequence based on the pooling operation characteristics under a preset weight transformation matrix and a preset bias term. The specific formula for obtaining the criterion weight sequence is as follows: ; in, For the criterion weight sequence and , For the first The weight vector of each criterion; The pre-defined weight transformation matrix is ​​trainable. For trainable preset bias terms, the preset weight change matrix and preset bias terms can be adjusted according to the above. The minimum value is determined during repeated unsupervised training to meet the convergence condition; GAP(·) is the pooling compression function, which represents global average pooling, and its function is to extract deep features of transformer operation. An averaging operation is performed over the time dimension to aggregate the feature representations of each criterion at different time points into a fixed-length feature vector.

[0045] After obtaining the criterion weight sequence Then, based on the multi-criteria discriminant value sequence The weighted average is used to obtain the state assessment score, and the specific formula is as follows: ; in, The condition assessment score reflects the overall health level of the target transformer in its current state. This embodiment sets a health score range for the condition assessment score. The score range for minor faults is: The score range for moderate faults is: The score range for severe faults is: If the status assessment score If the target transformer falls within the health score range, its health status is healthy, and no further action is required. If the status assessment score is... If the condition falls within the range of minor faults, the target transformer's health status is considered to be a minor fault, and it needs to be inspected during the next inspection. If the condition assessment score is... If the score falls within the moderate fault range, the target transformer's health status is considered moderately faulty, requiring maintenance within a short period. If the condition assessment score... If the score falls within the range of severe faults, the health status of the target transformer is considered to be severely faulty. In this case, power should be cut off immediately for inspection and emergency repairs should be arranged.

[0046] In one embodiment, the deep feature acquisition step is specifically as follows: Transformer operating features In a trainable preset query linearly changing weight matrix The following is the runtime feature query matrix. In a trainable preset key linearly changing weight matrix The following is the running feature key matrix. The weight matrix changes linearly with trainable preset values. The eigenvalue matrix obtained below And run the feature query matrix Running the feature key matrix and running eigenvalue matrix The corresponding calculation formula is: ; Based on runtime feature query matrix Running the feature key matrix and running eigenvalue matrix Projection mapping is performed to obtain a sequence of projected running feature query matrices, a sequence of projected running feature key matrices, and a sequence of projected running feature value matrices. Multi-head self-attention processing is then applied to these sequences to obtain the running feature correlation matrix. The corresponding formula for the running feature correlation matrix is ​​as follows: ; ; ; in, This is a pre-defined normalized exponential function; To run the feature key matrix The dimension of each row of key vectors determines the representational power of the key vectors in attention computation, and its function is to scale the dot product to prevent gradient vanishing. The function is computed for the self-attention mechanism; This is a concatenation function; The function is computed for the multi-head attention mechanism, and To run the feature correlation matrix; For the first A separate focus; For the first The projection run feature query matrix corresponds to each attention head. For the first The projection operation of the feature key matrix corresponds to each attention head. For the first The projection eigenvalue matrix corresponding to each attention head. This is the preset output projection matrix.

[0047] Based on the operational characteristic correlation matrix and transformer operational characteristics Residual connectivity and layer normalization are performed to obtain preliminary operational characteristics. The relevant formulas are as follows: ; in, Preliminary operational characteristics, This is a preset normalization function. The initial running characteristics will be... The input is processed by a feedforward network, and the corresponding formula is: ; in, To preprocess the deep characteristics of transformer operation, It is a non-linear activation function; and It is a trainable preset feature transformation matrix, and and , The hidden layer dimension of the feedforward network, For the dimensions of the intermediate layers of the feedforward network; and This is a preset feedforward network bias term. Based on the deep operational characteristics of the preprocessed transformer. and initial operational characteristics Residual connection and layer normalization are performed to obtain preliminary deep characteristics of the transformer, and the corresponding formulas are as follows: .

[0048] in, This is a preliminary description of the deep features of the transformer.

[0049] Confirm the current level, i.e., the number of times the deep feature acquisition step has been executed. Below the preset encoding level This will reveal the in-depth characteristics of the initial transformer operation. If the transformer operating characteristics are used as the basis for the next deep feature acquisition step, the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped, and the preliminary transformer operating deep features are used as the transformer operating deep features. The corresponding formula in the above process is: ; ; in, For the first The preliminary deep features of the transformer obtained from the hierarchical analysis are then... For the first The first level of the next level Preliminary deep-level characteristics of transformer operation obtained from the hierarchical analysis; This refers to the complete computation process of a single-level encoder, i.e., the deep feature acquisition steps. This describes the deep characteristics of transformer operation.

[0050] Under the same external operating environment, the target transformer's health status is assessed using the existing dual-input residual graph convolutional network method, with a highest coefficient of variation of 0.099 and a lowest coefficient of variation of 0.078. The target transformer's health status is assessed using the existing random forest method, with a highest coefficient of variation of 0.266 and a lowest coefficient of variation of 0.243. In contrast, under the same external operating environment, the target transformer's health status is assessed using the transformer health status assessment method described in the above embodiment, with a highest coefficient of variation of 0.085 and a lowest coefficient of variation of 0.063. It is evident that the highest and lowest coefficients of variation of the transformer health status assessment method in the above embodiment are significantly lower than those of the compared existing methods, demonstrating that the assessment results of the transformer health status assessment method in the above embodiment have superior stability.

[0051] Furthermore, under the same external operating environment, the target transformer's health status assessment using the existing dual-input residual graph convolutional network method yielded a maximum false alarm rate of 3.70% and a minimum false alarm rate of 2.60%. The target transformer's health status assessment using the existing random forest method yielded a maximum false alarm rate of 4.80% and a minimum false alarm rate of 3.50%. In contrast, under the same external operating environment, the target transformer's health status assessment using the transformer health status assessment method described in the above embodiments yielded a maximum false alarm rate of 2.60% and a minimum false alarm rate of 1.80%. It is evident that the maximum and minimum false alarm rates of the transformer health status assessment method in the above embodiments are significantly lower than those of the compared existing methods, demonstrating that the transformer health status assessment method in the above embodiments has higher assessment accuracy.

[0052] Please see Figure 2This embodiment also provides a transformer health status assessment system, including: a discriminant value generation module, used to acquire multi-source operating data of a target transformer under continuous differential protection cycles, and generate a multi-criteria discriminant value sequence based on the multi-source operating data; a deep feature acquisition module, used to obtain transformer operating features based on the multi-criteria discriminant value sequence, and perform a deep feature acquisition step based on the transformer operating features to obtain deep transformer operating features; a criterion weight acquisition module, used to pool and compress the deep transformer operating features to obtain pooled operating features, and obtain a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term; a status assessment module, used to weight the multi-criteria discriminant value sequence and the criterion weight sequence to obtain a status assessment score; and a health status determination module, used to generate a health status determination based on the status assessment score. The state is determined, and maintenance is performed based on the health status. The deep feature acquisition step includes: obtaining an operating feature query matrix, an operating feature key matrix, and an operating feature value matrix based on the transformer operating features under a preset linear transformation matrix set; performing multi-head self-attention processing based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain an operating feature association matrix; performing feedforward network processing based on the operating feature association matrix and the transformer operating features under a preset feature transformation matrix to obtain preliminary transformer deep features; if the number of executions of the deep feature acquisition step is confirmed to be less than a preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped, and the preliminary transformer deep features are used as transformer operating deep features.

[0053] Further, the step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: performing a criterion weight acquisition step based on the deep operating features of the transformer to obtain a criterion weight sequence; the criterion weight acquisition step includes: pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function; if it is confirmed that the operating feature similarity loss values ​​do not meet the preset convergence condition, then unsupervised training is performed based on the operating feature similarity loss values ​​to obtain an updated linear transformation matrix and an updated feature transformation matrix, and a deep feature acquisition step is performed based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features to update the deep operating features of the transformer, and the criterion weight acquisition step is re-executed based on the updated deep operating features of the transformer; otherwise, a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term.

[0054] In the above embodiments, transformer operating characteristics are obtained based on a multi-criteria discriminant value sequence. By repeatedly executing deep feature acquisition steps involving multi-head self-attention and feedforward network processing at a preset encoding level, stable and inherent intrinsic correlations between different criteria can be mined and strengthened from the transformer operating characteristics. This allows for the acquisition of deep transformer operating characteristics. When external environmental fluctuations cause a temporary abnormal change in a certain criterion, the inconsistency between this abnormal change and the aforementioned intrinsic correlation can be effectively identified. This effectively distinguishes the true health state of the transformer from transient noise caused by external environmental fluctuations, suppressing interference from external operating environment fluctuations and improving the stability of subsequent transformer health status assessments. Furthermore, when the data of a certain criterion undergoes a sudden change due to temporary environmental interference rather than actual degradation, the above embodiments can dynamically adjust the criterion weight sequence based on the deep transformer operating characteristics, reducing the impact of the operating data corresponding to that criterion on the health assessment results and improving the stability of the transformer health status assessment.

[0055] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of a transformer, characterized in that, Includes the following steps: Acquire multi-source operating data of the target transformer under continuous differential protection cycles, and generate a multi-criteria discriminant value sequence based on the multi-source operating data; Based on the multi-criteria discriminant value sequence, the transformer operation characteristics are obtained, and based on the transformer operation characteristics, a deep feature acquisition step is performed to obtain the deep features of transformer operation. The deep operating features of the transformer are pooled and compressed to obtain pooled operating features, and a criterion weight sequence is obtained based on the pooled operating features under a preset weight transformation matrix and a preset bias term. The state evaluation score is obtained by weighting the multi-criteria discriminant value sequence and the criterion weight sequence. A health status is generated based on the status assessment score, and maintenance is performed based on the health status. The deep feature acquisition steps include: Based on the transformer operating characteristics, an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix are obtained under a preset linear transformation matrix set. Multi-head self-attention processing is then performed based on the operating characteristic query matrix, the operating characteristic key matrix, and the operating characteristic value matrix to obtain an operating characteristic correlation matrix. Based on the operational feature correlation matrix and the transformer operational features, feedforward network processing is performed under a preset feature transformation matrix to obtain preliminary deep transformer features. If the number of times the deep feature acquisition step is executed is less than the preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped and the preliminary transformer deep features are used as transformer operating deep features.

2. The method for assessing the health status of a transformer according to claim 1, characterized in that, The step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: Based on the deep operating characteristics of the transformer, a criterion weight acquisition step is performed to obtain a criterion weight sequence; The steps for obtaining the criterion weights include: The deep operating features of the transformer are pooled and compressed to obtain pooled operating features, and similarity loss values ​​of operating features are generated based on the pooled operating features under a preset loss function. If the similarity loss value of the operating features does not meet the preset convergence condition, then unsupervised training is performed based on the similarity loss value of the operating features to obtain the updated linear transformation matrix and the updated feature transformation matrix. Based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features, the deep feature acquisition step is performed to update the deep features of transformer operation. The criterion weight acquisition step is then re-executed based on the updated deep features of transformer operation. Otherwise, the criterion weight sequence is obtained based on the pooled operating features under the preset weight transformation matrix and the preset bias term.

3. The method for assessing the health status of a transformer according to claim 2, characterized in that, The step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and generating operating feature similarity loss values ​​based on the pooled operating features under a preset loss function, includes: The deep operating features of the transformer are pooled and compressed to obtain pooled operating features, and cosine similarity is calculated based on the pooled operating features to obtain feature similarity. Based on the feature similarity, a running feature similarity loss value is generated under a preset loss function.

4. The method for assessing the health status of a transformer according to claim 1, characterized in that, The process of acquiring multi-source operating data of the target transformer under continuous differential protection cycles and generating a multi-criteria discriminant value sequence based on the multi-source operating data includes: Acquire multi-source operating data of the target transformer under continuous differential protection cycles; the multi-source operating data includes oil gas concentration sequence, charged particle number sequence, vibration signal sequence, monitoring image sequence, and transformer secondary current sequence; Based on the gas concentration sequence in the oil, a relative gas production rate sequence is obtained, and based on the relative gas production rate sequence, a gas concentration discrimination value sequence in the oil is obtained under a preset gas production rate threshold. Based on the charged particle number sequence, a dielectric loss factor sequence is obtained, and based on the dielectric loss factor sequence, a dielectric loss factor discrimination value sequence is obtained under a preset loss threshold. The vibration breakthrough number is obtained for each vibration signal in the vibration signal sequence whose amplitude exceeds a preset maximum vibration threshold, and the vibration breakthrough number sequence is obtained by integrating the vibration breakthrough number. Based on the vibration breakthrough number sequence, a tap changer discrimination value sequence is obtained under a preset breakthrough number threshold; The monitoring image sequence is used to identify an initial abnormal region sequence using the image difference method. The initial abnormal region sequence is then denoised and segmented to obtain a denoised abnormal region sequence. Color feature determination is performed based on the initial abnormal region sequence and the denoised abnormal region sequence to obtain an oil leakage discrimination value sequence. The differential current sequence is obtained by normalizing the secondary current sequence of the current transformer, and a differential current distance sequence is generated based on the differential current sequence under a preset standard sine wave sequence; a differential protection discrimination value sequence is obtained based on the differential current distance sequence under a preset distance threshold. The oil gas concentration discrimination value sequence, the medium loss factor discrimination value sequence, the tap changer discrimination value sequence, the oil leakage discrimination value sequence, and the differential protection discrimination value sequence are used as multi-criteria discrimination value sequences.

5. The method for assessing the health status of a transformer according to claim 4, characterized in that, The transformer operating characteristics are obtained based on the multi-criteria discriminant value sequence, and a deep feature acquisition step is performed based on the transformer operating characteristics to obtain deep transformer operating characteristics, including: Obtain the oil leakage discrimination value sequence and the medium loss factor discrimination value sequence from the multi-criteria discrimination value sequence; The transformer operating characteristics are obtained by aligning and integrating the leakage oil discrimination value sequence, the dielectric loss factor discrimination value sequence, the relative gas generation rate sequence, the vibration breakthrough number sequence, and the differential current distance sequence. Based on the transformer operating characteristics, a deep feature acquisition step is performed to obtain the deep operating characteristics of the transformer.

6. The method for assessing the health status of a transformer according to claim 1, characterized in that, The process involves obtaining an operating feature query matrix, an operating feature key matrix, and an operating feature value matrix based on the transformer operating characteristics under a preset linear transformation matrix set. Then, multi-head self-attention processing is performed based on the operating feature query matrix, the operating feature key matrix, and the operating feature value matrix to obtain an operating feature correlation matrix, including: Retrieve the query linear transformation weight matrix, key linear transformation weight matrix, and value linear transformation weight matrix from the preset linear transformation matrix set; Based on the linearly changing weight matrix and the transformer operating characteristics, an operating characteristic query matrix is ​​obtained; The operating feature key matrix is ​​obtained based on the linearly changing weight matrix of the keys and the operating characteristics of the transformer; The operating characteristic value matrix is ​​obtained based on the linearly changing weight matrix of the values ​​and the operating characteristics of the transformer; Based on the running feature query matrix, the running feature key matrix, and the running feature value matrix, a projection mapping process is performed to obtain a sequence of projected running feature query matrices, a sequence of projected running feature key matrices, and a sequence of projected running feature value matrices. The projection running feature query matrix sequence, the projection running feature key matrix sequence, and the projection running feature value matrix sequence are subjected to multi-head self-attention processing to obtain the running feature association matrix.

7. The method for assessing the health status of a transformer according to claim 1, characterized in that, The preliminary deep features of the transformer are obtained by performing feedforward network processing based on the operational feature correlation matrix and the transformer operational features under a preset feature transformation matrix, including: Based on the operational feature correlation matrix and the transformer operational features, residual connection and layer normalization are performed to obtain preliminary operational features; Based on the preliminary operating characteristics, a feedforward network is performed under a preset feature transformation matrix to obtain preprocessed deep operating characteristics of the transformer. Based on the pre-processed deep operating characteristics of the transformer and the preliminary operating characteristics, residual connection and layer normalization are performed to obtain the preliminary deep operating characteristics of the transformer.

8. The method for assessing the health status of a transformer according to claim 1, characterized in that, In the process of generating a health status based on the status assessment score and performing maintenance based on the health status, the step of generating a health status based on the status assessment score includes: If the status assessment score is confirmed to be within the preset health score range, a healthy status is generated and set as healthy; otherwise, it is determined whether the status assessment score is within the preset minor fault score range. If the status assessment score is confirmed to be within the preset range of minor fault scores, a healthy state set as minor fault is generated; otherwise, it is determined whether the status assessment score is within the preset range of moderate fault scores. If the status assessment score is confirmed to be within the preset range of moderate fault scores, a healthy state set as moderate fault is generated; otherwise, a healthy state set as severe fault is generated.

9. A transformer health status assessment system, characterized in that, A method for assessing the health status of a transformer as described in any one of claims 1 to 8, comprising: The discrimination value generation module is used to acquire multi-source operating data of the target transformer under continuous differential protection cycles, and generate a multi-criteria discrimination value sequence based on the multi-source operating data; The deep feature acquisition module is used to obtain transformer operation features based on the multi-criteria discriminant value sequence, and to perform a deep feature acquisition step based on the transformer operation features to obtain deep transformer operation features; The criterion weight acquisition module is used to pool and compress the deep operating features of the transformer to obtain pooled operating features, and to obtain a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term. The state assessment module is used to obtain a state assessment score by weighting the multi-criteria discriminant value sequence and the criterion weight sequence. The health status determination module is used to generate a health status based on the status assessment score and to perform maintenance based on the health status. The deep feature acquisition steps include: Based on the transformer operating characteristics, an operating characteristic query matrix, an operating characteristic key matrix, and an operating characteristic value matrix are obtained under a preset linear transformation matrix set. Multi-head self-attention processing is then performed based on the operating characteristic query matrix, the operating characteristic key matrix, and the operating characteristic value matrix to obtain an operating characteristic correlation matrix. Based on the operational feature correlation matrix and the transformer operational features, feedforward network processing is performed under a preset feature transformation matrix to obtain preliminary deep transformer features. If the number of times the deep feature acquisition step is executed is less than the preset coding level, then the preliminary transformer deep features are used as transformer operating features and the deep feature acquisition step is re-executed; otherwise, the deep feature acquisition step is stopped and the preliminary transformer deep features are used as transformer operating deep features.

10. A transformer health status assessment system according to claim 9, characterized in that, The step of pooling and compressing the deep operating features of the transformer to obtain pooled operating features, and obtaining a criterion weight sequence based on the pooled operating features under a preset weight transformation matrix and a preset bias term, further includes: Based on the deep operating characteristics of the transformer, a criterion weight acquisition step is performed to obtain a criterion weight sequence; The steps for obtaining the criterion weights include: The deep operating features of the transformer are pooled and compressed to obtain pooled operating features, and similarity loss values ​​of operating features are generated based on the pooled operating features under a preset loss function. If the similarity loss value of the operating features does not meet the preset convergence condition, then unsupervised training is performed based on the similarity loss value of the operating features to obtain the updated linear transformation matrix and the updated feature transformation matrix. Based on the updated linear transformation matrix, the updated feature transformation matrix, and the transformer operating features, the deep feature acquisition step is performed to update the deep features of transformer operation. The criterion weight acquisition step is then re-executed based on the updated deep features of transformer operation. Otherwise, the criterion weight sequence is obtained based on the pooled operating features under the preset weight transformation matrix and the preset bias term.

Images