Power equipment fault monitoring method and system based on big data

Abstract

This invention belongs to the field of power equipment fault monitoring technology, specifically relating to a power equipment fault monitoring method and system based on big data. The method includes: acquiring temperature, current, and partial discharge pulse sequences within a preset period; mapping the feature data of these sequences within a sliding time window to obtain a continuous time-step thermo-coupling index sequence and a partial discharge activity sequence; inputting the acquired sequences into a long short-term memory network to extract temporal hidden layer state vectors, and combining this with attention weighting based on a physical bias factor to obtain a degradation characterization vector; obtaining a comprehensive fault index based on the degradation characterization vector and a preset standard health vector, and performing final fault state detection and early warning. This invention improves the accuracy of power equipment condition monitoring and the timeliness of early warning.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault monitoring technology. More specifically, this invention relates to a power equipment fault monitoring method and system based on big data. Background Technology

[0002] Condition monitoring of power equipment refers to the process of real-time collection and analysis of the operating physical parameters of key equipment in power systems, such as substations. It is primarily used to assess the health of the equipment and prevent potential breakdown faults. Because the physical degradation process within power equipment is accompanied by abnormal thermal field distribution and high-frequency electromagnetic radiation, and because insulation degradation exhibits both long-term concealment and short-term suddenness, continuous monitoring and evaluation of the operating status of power equipment is necessary to avoid grid-related accidents such as large-scale power outages caused by sudden faults.

[0003] Existing power equipment condition monitoring methods typically rely on a single type of sensor to collect single-dimensional physical parameters such as temperature, load current, or partial discharge pulses for independent analysis, or use time-series data processing models such as long short-term memory networks to extract time-series features and determine the condition of the collected long-term operating data with equal weights.

[0004] However, sensor data collected from a single location are susceptible to interference from background electromagnetic white noise at the substation site, and the cumulative calculation of isolated physical quantities cannot characterize the degree of coupled damage to insulation materials caused by heat accumulation under real-time current impact. When processing long-term monitoring data, conventional long short-term memory networks (LSTMs) are prone to having their forget gate mechanism weaken the proportion of early transient discharge pulses retained in the final hidden layer state. This causes early, minute, sudden abnormal signals to be diluted in long-term stable data, resulting in the loss of key early fault precursor information during time-series fusion. Consequently, the accuracy of determining the overall degradation trend of the equipment and the timeliness of early warning are reduced. Summary of the Invention

[0005] To address the technical problems mentioned above, such as the inability of existing single-dimensional monitoring to characterize thermal coupling failure and the tendency of conventional long short-term memory networks to forget early, minute, sudden abnormal signals during long-term processing, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a power equipment fault monitoring method based on big data, comprising: acquiring a temperature sequence, a current sequence, and a partial discharge pulse sequence within a preset period; determining a relative temperature deviation based on the mean of the temperature sequence within a sliding time window; obtaining a relative current fluctuation rate based on the difference between the maximum and minimum values ​​of the current sequence within the sliding time window; obtaining a thermodynamic coupling index by combining the relative temperature deviation and the relative current fluctuation rate; linearly normalizing the cumulative pulse count of the partial discharge pulse sequence within the sliding time window to obtain the partial discharge activity; and acquiring a thermodynamic coupling index sequence and a partial discharge activity sequence for consecutive time steps; and combining the thermodynamic coupling index sequence and the partial discharge activity sequence for the same time step. The coupling index and partial discharge activity are concatenated into a merged feature matrix and input into a long short-term memory network. The temporal hidden state vectors corresponding to each time step are extracted. The physical bias factor is the absolute value of the difference between the partial discharge activity of each time step and the partial discharge activity of the previous adjacent time step. The attention weights corresponding to each time step are obtained by combining the physical bias factor and the temporal hidden state vectors. The degradation characterization vector is obtained by weighted summation of all temporal hidden state vectors according to the attention weights corresponding to each time step. The comprehensive fault index is obtained by Euclidean distance between the degradation characterization vector and the standard healthy vector. Fault state detection and early warning are performed based on the comprehensive fault index.

[0007] This invention achieves synchronous fusion processing of multi-dimensional physical information, solving the problem that single-dimensional physical information cannot accurately reflect the true state of equipment insulation degradation. This invention combines bias factors of partial discharge physical changes to process hidden-layer state features, reducing the risk of early abrupt changes being gradually forgotten or lost in long-cycle time-series processing, and improving the accuracy of hidden-layer features in reflecting the actual physical degradation process of equipment. This invention utilizes spatial distance differences to transform time-series features into intuitive fault assessment indices, enabling feature processing results to directly correspond to specific equipment health deviations, facilitating equipment status judgment and early warning, thereby improving the accuracy of power equipment status monitoring and the timeliness of early warning, and helping to prevent sudden power equipment failures.

[0008] Preferably, the acquisition of the temperature sequence, current sequence, and partial discharge pulse sequence within a preset period includes: during the system power-on initialization phase, aligning the timestamps of the fiber optic temperature sensor, open-type current transformer, and transient ground voltage sensor deployed inside the power equipment using a network time protocol; retrieving the absolute temperature values ​​continuously collected by the fiber optic temperature sensor within the preset period to construct a temperature sequence; extracting the transient load current values ​​collected by the open-type current transformer within the corresponding time period to construct a current sequence; and counting the number of high-frequency discharge pulse events captured by the transient ground voltage sensor within a single time step to construct a partial discharge pulse sequence.

[0009] Preferably, the step of determining the relative temperature deviation based on the average of the temperature sequence within the sliding time window and obtaining the relative current fluctuation rate based on the difference between the maximum and minimum values ​​of the current sequence within the sliding time window includes: reading the reference temperature and rated current parameters preset by the device's factory specification nameplate; determining the relative temperature deviation based on the difference between the average of the temperature sequence within the sliding time window and the reference temperature; and determining the relative current fluctuation rate by mapping the difference between the maximum and minimum values ​​of the current sequence within the sliding time window and the rated current parameters.

[0010] This invention reads the preset reference temperature and rated current parameters on the equipment's factory specification nameplate, and maps them to the mean of the temperature series within a sliding time window and the difference between the maximum and minimum values ​​of the current series within the sliding time window, respectively, to determine the relative temperature deviation and relative current fluctuation rate. This enables the measurement of the deviation of the equipment's current operating parameters from its design standard, eliminates the differences in absolute physical values ​​between different models and capacities of power equipment, and allows the extracted features to uniformly reflect the relative degree of equipment deviation from normal operating conditions, thereby enhancing the generalization ability of multi-dimensional features in different equipment.

[0011] Preferably, the thermodynamic coupling index satisfies the following relationship: In the formula, For the first Thermocoupling index of a sliding time window For the temperature sequence at the 1st Mean within a sliding time window The preset reference temperature parameter, For the current sequence in the th The maximum value within a sliding time window For the current sequence in the th The minimum value within a sliding time window The preset rated current parameters, It is an exponential function with the natural constant as its base. Let be a logarithmic function with the natural constant as the base. For temperature sensitivity coefficient, Weighting for current surge amplification.

[0012] This invention obtains a nonlinearly increasing temperature influence term based on relative temperature deviation, a current fluctuation term subject to numerical compression based on relative current fluctuation rate, and uses the temperature influence term to perform multiplicative correction on the current fluctuation term to obtain a thermo-coupling index. This realizes the restoration of the amplified effect of temperature accumulation on insulation damage caused by current surges, retains the severity of heat accumulation by utilizing the positive growth trend, and suppresses the risk of data overflow caused by extreme electromagnetic interference, thereby improving the rationality of the thermo-coupling index in reflecting the actual damage to power equipment.

[0013] Preferably, the training method of the Long Short-Term Memory (LSTM) network includes: acquiring multimodal sensing data of power equipment in a healthy state and at different stages of degradation as training sample data; constructing a network structure consisting of an input layer, an LSM network layer containing a preset number of hidden nodes, and a fully connected layer; using the mean squared error function as the loss function, performing forward propagation on the network structure based on the training sample data to obtain predicted features; determining the loss value based on the predicted features and the true labels of the training sample data; performing backpropagation based on the loss value to update the network weights until the loss value converges or reaches a preset number of iterations, thereby obtaining a pre-trained LSM network.

[0014] Preferably, the attention weights satisfy the following relationship: In the formula, For the first Attention weights at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step The total number of time steps. It is an exponential function with the natural constant as its base. This is the bias gain constant.

[0015] This invention combines the physical bias factors of each time step with the spatial length characteristics of the temporal hidden layer state vector, uses nonlinear numerical transformation for enhancement or suppression, and accumulates and normalizes the overall feature response values ​​of all time steps to obtain the attention weights by adding single-step weight numerators. This achieves dynamic allocation of feature weights based on the severity of discharge mutations. When the local discharge deteriorates more severely, the corresponding hidden layer features are allocated a larger proportion in the final temporal fusion. This overcomes the problem of weakening key fault precursor information caused by equal weight output in the network, and improves the accuracy of characterizing the hidden degradation process of the equipment.

[0016] Preferably, the comprehensive failure index is calculated as follows: In the formula, The comprehensive failure index, To degrade the Euclidean distance between the characterization vector and the standard health vector, It is an exponential function with the natural constant as its base. This is the spatial mapping smoothing parameter.

[0017] This invention reflects the degree of physical deviation by squaring the Euclidean distance between the degradation characteristic vector and the standard health vector. It then transforms the physical deviation using a continuous and smooth numerical interval mapping rule, combined with a spatial mapping smoothing parameter, forcibly mapping the spatial distance to a finite numerical interval to obtain a comprehensive fault index. This invention achieves the transformation of abstract spatial distance into an intuitive fault assessment indicator, ensuring that operating states deviating further from the factory health benchmark correspond to higher fault indices. This facilitates direct reading and judgment by the operation and maintenance scheduling end to determine whether equipment exhibits significant degradation characteristics.

[0018] Preferably, the detection and early warning of fault status based on the comprehensive fault index includes: in response to the comprehensive fault index calculated within a continuous preset number of time steps of the evaluation period being greater than a preset severe alarm threshold, determining that the equipment is in a physical state on the verge of a breakdown fault, and automatically triggering a first-level severe early warning command to control the incoming circuit breaker to disconnect the load.

[0019] Preferably, the detection and early warning of fault status based on the comprehensive fault index includes: in response to the comprehensive fault index being within a preset maintenance range and the acquired partial discharge pulse sequence showing an increasing frequency trend over a long time span, determining that there are signs of gradual deterioration in the local insulation, generating a secondary maintenance early warning work order to schedule on-site manual testing instruments to retest the insulation withstand voltage.

[0020] Secondly, the present invention provides a power equipment fault monitoring system based on big data, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned power equipment fault monitoring method based on big data is implemented.

[0021] By adopting the above technical solution, the above-mentioned big data-based power equipment fault monitoring method is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0022] The beneficial effects of this invention are as follows: By collecting and synchronously aligning multi-dimensional physical signals within power equipment, this invention jointly processes the heating, load fluctuations, and high-frequency discharge pulses generated during equipment operation. It extracts features of the superimposed destructive effects of environmental heating and current surges on insulation materials, restoring the true stress state of the equipment under complex operating conditions. This allows the extracted features to more closely reflect the physical degradation phenomena within the equipment. Furthermore, considering the coexistence of long-term, hidden, and short-term sudden deterioration processes in power equipment, this invention, when processing long-term sequence features, utilizes the differences in changes of specific physical features to adjust the output at different time points. The results are assigned different weights; this invention amplifies the presence of transient deterioration features in the overall assessment, ensuring that even extremely brief early anomalies retain a corresponding response proportion in the final fused monitoring results, thus improving the feature processing's ability to capture early, minor fault precursors. This invention compares the spatial differences between the time-series fusion-processed feature results and the baseline features of the equipment in a fault-free state, transforming complex abstract features into intuitive fault assessment indices. These indices directly determine the degree to which the equipment's current operating state deviates from its normal safe state, triggering corresponding graded early warnings and on-site retesting and dispatching instructions. This improves the accuracy of determining power equipment degradation trends, provides reliable judgment support for preventative intervention in on-site operation and maintenance, and reduces the harm caused by sudden equipment breakdown faults to the overall power grid. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the big data-based power equipment fault monitoring method of the present invention. Detailed Implementation

[0024] 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, not all, of the embodiments of the present invention. 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.

[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] This invention discloses a power equipment fault monitoring method based on big data, referring to... Figure 1 This includes steps S1-S4: S1. Obtain the temperature sequence, current sequence, and partial discharge pulse sequence within a preset period through a multimodal sensor network.

[0027] It should be noted that the physical degradation process inside power equipment is accompanied by abnormal thermal field distribution and high-frequency electromagnetic radiation. Data collected by sensors at a single location are easily interfered with by background electromagnetic white noise in the substation. To extract multi-dimensional physical parameters with cross-validation properties, it is necessary to synchronously capture multi-modal signals from the power equipment. Therefore, this invention constructs synchronous monitoring data through a multi-modal sensor network.

[0028] Specifically, during the system power-on initialization phase, the timestamps of the fiber optic temperature sensors, open-type current transformers, and transient ground voltage sensors deployed inside the power equipment are aligned using the Network Time Protocol (NTP). Absolute temperature values ​​continuously collected by the fiber optic temperature sensors within a preset period are retrieved to construct a temperature sequence. Transient load current values ​​collected by the open-type current transformers within the corresponding time period are extracted to construct a current sequence. The number of high-frequency discharge pulse events captured by the transient ground voltage sensor within a single time step is counted to construct a partial discharge pulse sequence.

[0029] For example, the preset cycle length is 24 hours.

[0030] S2. Based on the temperature sequence, current sequence, and partial discharge pulse sequence, combined with the mean and range data of the characteristics within the sliding time window, obtain the thermo-coupling index sequence and the partial discharge activity sequence.

[0031] It should be noted that the abnormal temperature rise effect caused by the aging of the conductive circuit exhibits a strong positive correlation with the real-time fluctuating current carried by the equipment. The cumulative calculation of isolated physical quantities cannot characterize the degree of damage to the insulating material caused by heat accumulation under current surges. Therefore, this invention combines feature extraction from multiple physical quantities to construct a thermo-coupling index sequence and a partial discharge activity sequence.

[0032] Specifically, the reference temperature and rated current parameters preset on the device's factory specification nameplate are read. A preset sliding time window is established, and the mean value of the temperature sequence within the sliding time window is extracted. This mean value is then normalized and mapped to the reference temperature to determine the relative temperature deviation. The difference between the maximum and minimum values ​​of the current sequence within the same sliding time window is extracted, and this difference is mapped to the rated current to determine the relative current fluctuation rate. Combining the relative temperature deviation and the relative current fluctuation rate, the thermodynamic coupling index for the corresponding sliding time window is obtained. The cumulative pulse count of the partial discharge pulse sequence within the corresponding window is extracted and linearly normalized to obtain the partial discharge activity. The thermodynamic coupling index sequence and the partial discharge activity sequence for continuous time steps are obtained through time-series deduction.

[0033] For example, the length of the sliding time window is 1 hour.

[0034] Specifically, the thermo-coupling index satisfies the following relationship: ; In the formula, For the first Thermocoupling index of a sliding time window For the temperature sequence at the 1st Mean within a sliding time window The preset reference temperature parameter, For the current sequence in the th The maximum value within a sliding time window For the current sequence in the th The minimum value within a sliding time window The preset rated current parameters, It is an exponential function with the natural constant as its base. Let be a logarithmic function with the natural constant as the base. For temperature sensitivity coefficient, Weighting for current surge amplification, for example, The empirical value range is [1.5, 3]. The empirical value range is [0.8, 2]. In this embodiment... It is 2. The value is 1.2, and the implementers can determine it based on the actual situation. and When electrical equipment is in a harsh environment with high temperature and high humidity, the operating temperature can be appropriately reduced. To reduce abnormal fluctuations in non-fault-related indicators caused by ambient background temperature; when the equipment is in an enclosed indoor substation with its own cooling circulation system, the pressure can be appropriately increased. This improves the sensitivity of monitoring subtle abnormal heating in equipment. When the power grid is located in an industrial area with drastic load fluctuations, the sensitivity can be appropriately reduced. To prevent false alarms caused by normal fluctuations; when in a residential area with stable load, the load can be appropriately increased. To detect subtle, unusual shocks.

[0035] in, This represents the current relative temperature deviation of the equipment. A larger value indicates more severe heat accumulation, leading to… An increase in this value leads to an upward gain in the thermocoupling index; a smaller value indicates a more stable operating temperature of the equipment, resulting in... Approaching 1, thus making Maintain at the baseline response level. This represents the relative current fluctuation rate of the load current over a very short period of time. A larger value indicates a stronger load impact on the power grid, leading to... The larger the value, the better it is at suppressing computational overflow caused by extreme electromagnetic interference while maintaining a positive growth trend through the logarithmic function structure; the smaller the value, the more stable the current becomes. The smaller the value, the better. This is to avoid the device operating in a completely stable steady state within a certain sliding time window, which could lead to... To bring the thermal coupling index to zero, by right By performing multiplicative correction, the more severe the heat accumulation, the greater the amplification and correction of the insulation damage effect caused by the current surge, thus accurately restoring the accelerated material damage phenomenon under the superposition of electrothermal coupling.

[0036] S3. The thermal coupling index sequence and partial discharge activity sequence are processed by a long short-term memory network, and attention weighting with physical bias factor as the core is combined to obtain the degradation characterization vector of power equipment.

[0037] It should be noted that insulation degradation in power equipment exhibits both long-term concealment and short-term suddenness. When processing long-period data, the forgetting gate mechanism in Long Short-Term Memory (LSTM) networks can easily weaken the proportion of features retained in the final hidden state from early transient discharge pulses. Therefore, this invention corrects the network's equally weighted output by dynamically allocating attention based on physical features.

[0038] Specifically, the thermal coupling index and partial discharge activity at the same time step are concatenated into a merged feature matrix. This merged feature matrix is ​​then input into a pre-trained Long Short-Term Memory (LSTM) network to extract the temporal hidden state vectors output by the LTM network at each time step. The absolute value of the difference between the partial discharge activity at each time step and the partial discharge activity at the previous adjacent time step is calculated, and this absolute value is used as the physical bias factor. The attention weights corresponding to each time step are obtained by combining the physical bias factor and the temporal hidden state vectors. Finally, all temporal hidden state vectors are weighted and summed item by item according to the attention weights at each time step to obtain the overall degradation representation vector.

[0039] For example, the thermal coupling index is 1.25, the partial discharge activity is 0.45, and the merged feature matrix is... .

[0040] For example, the total number of time steps is 3, and the attention weights of each time step are 0.2, 0.3, and 0.5, respectively. The corresponding temporal hidden layer state vectors are [0.1, 0.2], [0.3, 0.4], and [0.5, 0.6], respectively. By weighting and aggregating the attention weights of each time step with their corresponding temporal hidden layer state vectors, the degradation representation vector obtained is [0.36, 0.46].

[0041] It should be added that the training method of the Long Short-Term Memory (LSTM) network includes: acquiring multimodal sensing data of power equipment in a healthy state and at different stages of degradation as training sample data; constructing a network structure consisting of an input layer, an LTM network layer containing a preset number of hidden nodes, and a fully connected layer; using the mean squared error function as the loss function, performing forward propagation on the network structure based on the training sample data to obtain predicted features; determining the loss value based on the predicted features and the true labels of the training sample data; and performing backpropagation based on the loss value to update the network weights, thereby obtaining a pre-trained LTM network.

[0042] Specifically, the attention weights satisfy the following relation: ; In the formula, For the first Attention weights at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step The total number of time steps. It is an exponential function with the natural constant as its base. The bias gain constant, for example The empirical value range is [1, 5]. In this embodiment... The value is 1.5. When the insulating material is sensitive to partial discharge and prone to breakdown, this parameter can be appropriately increased to amplify the weight of transient discharge change characteristics. When there is a lot of white noise in the field sensor, this parameter can be appropriately decreased to prevent noise disturbance from being erroneously amplified.

[0043] in, This represents the severity of transient abrupt changes in partial discharge activity. A higher value indicates a greater likelihood of a rapid deterioration in the partial discharge at the current time step, leading to… Increase in size, thus affecting This produces an enhancing effect, thereby assigning a greater attention weight to the current time step; a smaller value indicates more stable or non-discharging activity, leading to... The smaller the value, the more it suppresses the attention weight at that time step. Through The overall feature response values ​​of all time steps are accumulated, and the weight numerators of each single step are normalized. Attention is dynamically allocated through physical features. When the local discharge deteriorates more severely, the corresponding hidden layer features have a greater weight in the final time-series fusion, ensuring the retention of key fault precursor information.

[0044] S4. Based on the degradation characterization vector and the preset standard health vector, obtain the comprehensive fault index of the power equipment, and perform final fault state detection and early warning.

[0045] It should be noted that the hidden layer vector features extracted after multimodal feature fusion reside in an abstract space lacking clear physical units of measurement, making it difficult for the operation and maintenance scheduling system to directly read and formulate intervention strategies. To measure the difference between the current degree of equipment micro-deterioration and the deviation from the ideal safe state, this invention obtains a comprehensive fault index for power equipment based on the degradation characterization vector and a preset standard health vector, and performs final fault state detection and early warning.

[0046] Specifically, the pre-calibrated standard health vector stored in the system monitoring module is retrieved. The Euclidean distance between the degradation characterization vector and the standard health vector is calculated. The comprehensive failure index is obtained based on this Euclidean distance.

[0047] For example, the standard health vector is generated by calibrating the average hidden layer output of data from when the device is operating in a factory fault-free state.

[0048] Specifically, the comprehensive failure index is calculated as follows: ; In the formula, The comprehensive failure index, To degrade the Euclidean distance between the characterization vector and the standard health vector, It is an exponential function with the natural constant as its base. For spatial mapping smoothing parameters, The empirical value range is [5, 20]. In this embodiment... The value is 10, which can be determined by the implementers based on the actual situation. When it is necessary to intervene in minor defects at an early stage to prevent a complete power outage, this parameter can be appropriately reduced to improve the sensitivity of spatial mapping; when connecting old power equipment for condition monitoring, this parameter can be appropriately increased to increase the system's tolerance to normal fluctuations.

[0049] in, This reflects the absolute deviation of the current operating characteristics of power equipment from the factory health benchmark. A larger value indicates a greater deviation from normal physical health patterns in the equipment's operating state, and a higher likelihood of significant structural deterioration or discharge characteristics, leading to… The closer the value is to 1, the more likely it is to trigger a high-risk trip alarm; the smaller the value, the more similar the equipment's operating state is to the factory default state, leading to... The closer the deviation is to zero, the more likely the equipment is to be in a highly safe operating state. By converting physical deviations through an exponential function, spatial distances are forcibly mapped to the range of 0 to 1, facilitating the detection and early warning of eventual failure states.

[0050] Furthermore, fault warning commands are output based on the continuous monitoring results of the comprehensive fault index.

[0051] In one embodiment, in response to the comprehensive fault index calculated within three consecutive time steps of the evaluation period being greater than a preset severe alarm threshold, the equipment is determined to be in a physical state on the verge of a breakdown fault, and a first-level severe warning command is automatically triggered to control the incoming circuit breaker to disconnect the load. For example, the severe alarm threshold is 0.85.

[0052] In another embodiment, in response to the comprehensive fault index being within a preset maintenance interval and the acquired partial discharge pulse sequence showing an increasing frequency trend over a long time span, it is determined that there are signs of gradual deterioration in the local insulation, and a secondary maintenance early warning work order is generated to schedule on-site manual testing instruments to retest the insulation withstand voltage. For example, the preset maintenance interval is [0.60, 0.85].

[0053] The present invention also discloses a power equipment fault monitoring system based on big data, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the power equipment fault monitoring method based on big data according to the present invention is implemented.

[0054] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A method for monitoring power equipment faults based on big data, characterized in that, include: Acquire the temperature sequence, current sequence, and partial discharge pulse sequence within a preset period; The relative temperature deviation is determined based on the mean of the temperature sequence within the sliding time window. The relative current fluctuation rate is obtained based on the difference between the maximum and minimum values ​​of the current sequence within the sliding time window. The thermal coupling index is obtained by combining the relative temperature deviation and the relative current fluctuation rate. The partial discharge activity is obtained by linearly normalizing the cumulative pulse count of the partial discharge pulse sequence within the sliding time window. The thermal coupling index sequence and the partial discharge activity sequence of the continuous time step are obtained. The thermal coupling index and partial discharge activity at the same time step are concatenated into a merged feature matrix and input into the Long Short-Term Memory network. The temporal hidden state vectors corresponding to each time step are extracted. The absolute value of the difference between the partial discharge activity at each time step and the partial discharge activity at the previous adjacent time step is used as the physical bias factor. The attention weights corresponding to each time step are obtained by combining the physical bias factor and the temporal hidden state vectors. All temporal hidden state vectors are weighted and summed item by item according to the attention weights corresponding to each time step to obtain the degradation characterization vector. The comprehensive failure index is obtained according to the Euclidean distance between the degradation characterization vector and the standard health vector. Fault status is detected and early warning is given based on the comprehensive fault index.

2. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The acquisition of the temperature sequence, current sequence, and partial discharge pulse sequence within a preset period includes: during the system power-on initialization phase, aligning the timestamps of the fiber optic temperature sensor, open-type current transformer, and transient ground voltage sensor deployed inside the power equipment using the Network Time Protocol; retrieving the absolute temperature values ​​continuously collected by the fiber optic temperature sensor within the preset period to construct the temperature sequence; extracting the transient load current values ​​collected by the open-type current transformer within the corresponding time period to construct the current sequence; and counting the number of high-frequency discharge pulse events captured by the transient ground voltage sensor within a single time step to construct the partial discharge pulse sequence.

3. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The process of determining the relative temperature deviation based on the average of the temperature sequence within the sliding time window and obtaining the relative current fluctuation rate based on the difference between the maximum and minimum values ​​of the current sequence within the sliding time window includes: reading the reference temperature and rated current parameters preset by the equipment's factory specification nameplate; determining the relative temperature deviation based on the difference between the average of the temperature sequence within the sliding time window and the reference temperature; and determining the relative current fluctuation rate by mapping the difference between the maximum and minimum values ​​of the current sequence within the sliding time window and the rated current parameters.

4. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The thermo-coupling index satisfies the following relationship: ； In the formula, For the first Thermocoupling index of a sliding time window For the temperature sequence at the 1st Mean within a sliding time window The preset reference temperature parameter, For the current sequence in the th The maximum value within a sliding time window For the current sequence in the th The minimum value within a sliding time window The preset rated current parameters, It is an exponential function with the natural constant as its base. Let be a logarithmic function with the natural constant as the base. For temperature sensitivity coefficient, Weighting for current surge amplification.

5. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The training method of the Long Short-Term Memory (LSTM) network includes: acquiring multimodal sensing data of power equipment in a healthy state and at different stages of degradation as training sample data; constructing a network structure consisting of an input layer, an LSM network layer containing a preset number of hidden nodes, and a fully connected layer; using the mean squared error function as the loss function, performing forward propagation on the network structure based on the training sample data to obtain predicted features; determining the loss value based on the predicted features and the true labels of the training sample data; performing backpropagation based on the loss value to update the network weights until the loss value converges or reaches a preset number of iterations, thereby obtaining a pre-trained LSM network.

6. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The attention weights satisfy the following relationship: ； In the formula, For the first Attention weights at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step For the first The physical bias factor for each time step. For the first The L2 norm of the temporal hidden state vector at each time step The total number of time steps. It is an exponential function with the natural constant as its base. This is the bias gain constant.

7. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The comprehensive failure index is calculated as follows: ； In the formula, The comprehensive failure index, To degrade the Euclidean distance between the characterization vector and the standard health vector, It is an exponential function with the natural constant as its base. This is the spatial mapping smoothing parameter.

8. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The detection and early warning of fault status based on the comprehensive fault index includes: in response to the comprehensive fault index calculated within a continuous preset number of time steps of the evaluation period being greater than the preset severe alarm threshold, determining that the equipment is in a physical state on the verge of a breakdown fault, and automatically triggering a first-level severe early warning command to control the incoming circuit breaker to disconnect the load.

9. The power equipment fault monitoring method based on big data according to claim 1, characterized in that, The method of detecting and warning of fault status based on the comprehensive fault index includes: in response to the comprehensive fault index being within the preset maintenance range and the acquired partial discharge pulse sequence showing an increasing frequency trend over a long time span, determining that there are signs of gradual deterioration in the local insulation, generating a secondary maintenance warning work order to schedule on-site manual testing instruments to retest the insulation withstand voltage.

10. A power equipment fault monitoring system based on big data, characterized in that, include: A processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the big data-based power equipment fault monitoring method according to any one of claims 1-9.

Images