Industrial equipment health state prediction method based on multi-source data fusion

Abstract

The invention relates to the field of health monitoring, and discloses an industrial equipment health state prediction method based on multi-source data fusion, and the method comprises the steps: firstly, synchronously collecting the electrical performance monitoring data and the insulation state monitoring data of a transformer; fusing the characteristics of the electrical performance monitoring data and the insulation state monitoring data, inputting the fused data into a pre-trained health assessment model, and calculating to obtain a comprehensive health index; judging whether the equipment enters an abnormal state or not according to a comparison result of the comprehensive health index and a preset threshold value; then, when the industrial equipment does not enter the abnormal state, continuously calculating the short-term change trend of the electrical performance monitoring data; dynamically adjusting the acquisition frequency or scanning intensity of the insulation state monitoring data according to the value of the short-term change trend; according to the invention, when the abnormal trend is detected, the resource is automatically focused for deep exploration; and when the equipment state is stable, high-density scanning is cancelled, so that the utilization rate of computing resources and communication bandwidth is improved.

Description

Technical Field

[0001] This invention relates to the field of health monitoring, and more specifically to a method for predicting the health status of industrial equipment based on multi-source data fusion. Background Technology

[0002] Industrial equipment, such as transformers and energy storage devices in power systems, and pumps, fans, and compressors in process industries, are core assets for ensuring social production and energy security. Accurate prediction of their health status and early diagnosis of faults are crucial for preventing unplanned downtime, reducing maintenance costs, and avoiding catastrophic accidents.

[0003] In recent years, with the popularization of sensor technology and the Internet of Things (IoT), data-driven predictive maintenance (PdM) strategies are gradually replacing traditional periodic maintenance and reactive repair, becoming the mainstream approach to industrial equipment health management. Existing predictive maintenance methods can be broadly categorized as follows: 1. Analysis methods based on a single data source; These methods rely on single-type sensor data. For example, vibration analysis is commonly used for fault diagnosis in rotating machinery; direct gas chromatography (DGA) is the "gold standard" for identifying internal insulation faults in transformers; and current characteristic analysis (MCSA) is used to diagnose electrical faults such as broken rotor bars and eccentricity in motors. However, these methods have significant limitations: First, the fault modes of industrial equipment are complex and diverse, and a single type of data can only reflect the state of one aspect of the equipment, lacking a global perspective and easily leading to misdiagnosis or missed diagnosis. For example, vibration data alone may not be sufficient to effectively distinguish between mechanical imbalance and electrical faults. Second, these methods typically rely on expert experience to set thresholds, resulting in limited intelligence and difficulty in automatically identifying early, subtle fault symptoms.

[0004] 2. A simple method for parallel monitoring of multi-source data; To obtain more comprehensive information, some systems have begun deploying multiple sensors to collect various signals such as vibration, temperature, current, and acoustics in parallel. However, most current systems only display data in parallel or trigger independent alarms, lacking a deep information fusion mechanism. Operators need to manually compare alarm information from different systems to make a comprehensive judgment on the fault location, which is not only inefficient but also heavily reliant on personal experience and lacks consistency.

[0005] 3. Preliminary data fusion attempt; Existing technologies have attempted to fuse multi-source data. For example, feature vectors from multiple sensors are simply concatenated and input into a classifier for state recognition. However, such methods remain static and rigid. Their detection strategies (such as sensor sampling frequency and scanning range) are usually pre-set and cannot be dynamically adjusted according to the real-time status of the device. When the system detects an abnormal trend, it cannot automatically focus resources for in-depth investigation; while when the device status is stable, it continues to perform high-density scanning, resulting in a significant waste of computing resources and communication bandwidth, limiting the economic viability of large-scale deployment of this technology on resource-constrained edge devices.

[0006] In summary, existing technologies lack a solution that can adaptively, efficiently, and accurately predict the health status of industrial equipment. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the health status of industrial equipment based on multi-source data fusion, thereby solving at least one of the above-mentioned technical problems.

[0008] The objective of this invention can be achieved through the following technical solutions: The industrial equipment health status prediction method based on multi-source data fusion includes the following steps: S1. Synchronously collect electrical performance monitoring data and insulation status monitoring data of the transformer; S2. The features of the electrical performance monitoring data and insulation status monitoring data are fused and input into a pre-trained health assessment model to calculate a comprehensive health index; based on the comparison result of the comprehensive health index and a preset threshold, it is determined whether the equipment has entered an abnormal state. S3. When the industrial equipment is not in an abnormal state, continuously calculate the short-term trend of the electrical performance monitoring data; and dynamically adjust the acquisition frequency or scanning intensity of the insulation status monitoring data according to the value of the short-term trend. The value of the short-term trend is negatively correlated with the acquisition frequency or scanning intensity.

[0009] As a further technical solution, the specific steps for dynamically adjusting the acquisition frequency or scanning intensity of the insulation condition monitoring data in step S3 are as follows: When the value of the short-term trend is less than the first trend threshold, increase the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor. When the value of the short-term trend exceeds the second trend threshold, reduce the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor.

[0010] As a further technical solution, the method also includes: S4. When the comprehensive health index exceeds the preset alarm threshold, the internal structure of the transformer tank is located; the steps for this location are as follows: S41. Perform the k-th scan on the target area, divide it into grids and calculate the local anomaly index of each grid to identify the k-th hotspot cells. S42. Using the kth hotspot unit as the center, a smaller scanning area is redefined; S43. Within the reduced scanning area, perform the (k+1)th scan, wherein the grid resolution of the (k+1)th scan is higher than that of the kth scan, calculate the local anomaly index of each grid and identify the (k+1)th hot spot cell. S44. Let k = k + 1, iteratively execute steps S42 and S43 until the physical size of the scanned area is less than or equal to the preset accuracy threshold, and output the center coordinates of the hot spot unit identified in the last iteration as the final coordinates of the abnormal part.

[0011] As a further technical solution, step S43, when performing the (k+1)th scan, also includes the following steps: The current scanning area is simultaneously divided into three grid levels: low, medium, and high resolution. Calculate and identify hotspot units at each level; If a hot spot cell at a high resolution level is contained within a hot spot cell at a low resolution level, it is determined that the focus is consistent, and step S44 is continued to shrink the region. Otherwise, it is determined to be a focus inconsistency, and multi-focus tracking is initiated.

[0012] As a further technical solution, the multi-focus tracking method is as follows: Sub-scan regions are created for each of the currently identified high-abnormality index units, and step S4 is executed in parallel for each sub-scan region.

[0013] As a further technical solution, in step S1, the electrical performance monitoring data includes high-frequency current signals, voltage harmonics, and reactive power curves. The insulation condition monitoring data includes dissolved gas content in the oil, partial discharge signal, and top oil temperature. The dissolved gas content in the oil is monitored by an online oil chromatograph, and the partial discharge signal is monitored by a partial discharge sensor.

[0014] As a further technical solution, the specific process of calculating the local anomaly index of each grid and identifying the k-th hotspot cell in step S41 is as follows: S441. Based on the spatial positioning monitoring data obtained from the k-th scan, the spatial positioning monitoring data includes one or more of ultrasonic signal amplitude, ultra-high frequency signal energy, and infrared thermometry data; the spatial positioning monitoring data is acquired based on a partial discharge sensor array, an ultrasonic sensor array, an ultra-high frequency sensor array, and an infrared thermal imager. S442. For each grid cell within the current scanning area, a comprehensive local anomaly index is calculated through weighted summation. The calculation formula is as follows: ; in, This is a local anomaly index. This is the normalized value of the ultrasonic signal amplitude. This is the normalized value of the UHF signal energy. This is the normalized value of the temperature. , , These are the weighting coefficients, and + + =1; S443. Compare the local anomaly indices of all grids, and identify the grid cell with the highest local anomaly index that exceeds the preset activity threshold as the k-th hotspot cell.

[0015] As a further technical solution, the steps for obtaining the short-term trend value are as follows: S31. Obtain the historical sequence of each parameter in the electrical performance monitoring data within the most recent time window T, and calculate the standard deviation of the historical sequence respectively; S32. Select the parameter with the largest standard deviation as the key electrical parameter; S33. Perform linear regression analysis on the historical sequences corresponding to key electrical parameters and calculate the slope values ​​of the key electrical parameters; S34. The absolute value of the slope value is taken as the numerical value of the short-term trend.

[0016] As a further technical solution, the pre-trained health assessment model is a machine learning model based on gradient boosting decision trees; The health assessment model is trained using historical data, which includes electrical performance monitoring data and insulation status monitoring data corresponding to the transformer under normal operating conditions and various known fault conditions. The comprehensive health index is a scalar between 0 and 1, used to quantitatively characterize the continuous health status of a transformer from complete health to complete failure.

[0017] The beneficial effects of this invention are: This invention establishes a dynamic adjustment mechanism that is negatively correlated with insulation monitoring intensity by quantifying the short-term changing trend of electrical data. When the equipment is in a stable state, the oil chromatography sampling frequency and partial discharge scanning density are automatically reduced, directly reducing sensor energy consumption, data storage and computing overhead. The monitoring intensity is only increased when the state changes rapidly, thereby realizing the transformation of monitoring resources from fixed consumption to on-demand allocation while ensuring safety, and reducing operation and maintenance costs. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a logical schematic diagram of the present invention. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 As shown, this invention is a method for predicting the health status of industrial equipment based on multi-source data fusion, comprising the following steps: S1. Synchronously collect electrical performance monitoring data and insulation status monitoring data of the transformer; S2. The features of the electrical performance monitoring data and insulation status monitoring data are fused and input into a pre-trained health assessment model to calculate a comprehensive health index; based on the comparison result of the comprehensive health index and a preset threshold, it is determined whether the equipment has entered an abnormal state. One example of how to obtain the preset threshold is as follows: using a trained health assessment model, a batch of historical samples labeled as normal and faulty are predicted to obtain their health index; after plotting the PR curve or ROC curve, a point that can balance the false positive rate and the false negative rate is selected, and the corresponding health index value is the preset threshold.

[0022] S3. When the industrial equipment is not in an abnormal state, continuously calculate the short-term trend of the electrical performance monitoring data; and dynamically adjust the acquisition frequency or scanning intensity of the insulation status monitoring data according to the value of the short-term trend. Feature fusion can be achieved through feature concatenation or weighted averaging.

[0023] The value of the short-term trend is negatively correlated with the acquisition frequency or scanning intensity.

[0024] The pre-trained health assessment model is a machine learning model based on gradient boosting decision trees; The health assessment model is trained using historical data, which includes electrical performance monitoring data and insulation status monitoring data corresponding to the transformer under normal operating conditions and various known fault conditions. The comprehensive health index is a scalar between 0 and 1, used to quantitatively characterize the continuous health status of a transformer from complete health to complete failure.

[0025] In this embodiment, by simultaneously collecting electrical performance monitoring data and insulation status monitoring data of the transformer, multi-source data fusion is achieved, thereby comprehensively capturing the operating status of the equipment. The features of these two types of data are then fused and input into a pre-trained health assessment model to calculate a comprehensive health index. This more accurately reflects the overall health status of the equipment, avoiding misjudgments or omissions that may occur with a single data source. For example, electrical performance data (such as current and voltage) can reflect the instantaneous load and electrical stress of the equipment, while insulation status data (such as gas in oil and partial discharge) can reveal long-term aging trends. The combination of these two data makes the health assessment more comprehensive and reliable. Based on the comparison between the comprehensive health index and a preset threshold, it is possible to promptly determine whether the equipment has entered an abnormal state, facilitating early warning and intervention and reducing the risk of sudden failures.

[0026] Furthermore, when the equipment is not in an abnormal state, by continuously calculating the short-term trend of electrical performance monitoring data and dynamically adjusting the acquisition frequency or scanning intensity of insulation status monitoring data based on this trend, the optimal allocation of monitoring resources is achieved. This reduces unnecessary monitoring when the equipment is in a stable state, thereby lowering energy consumption and data storage costs, while enhancing monitoring and improving detection sensitivity when the state changes. The short-term trend is negatively correlated with the acquisition frequency or scanning intensity, ensuring that monitoring can be strengthened in a timely manner when the equipment condition deteriorates, and that monitoring can be appropriately relaxed when the condition improves, thus improving monitoring efficiency while ensuring safety. The pre-trained health assessment model is based on Gradient Boosting Decision Tree (GBDT). Trained on historical data, including data from normal operation and various known fault states, the model learns the mapping relationship between equipment health status and monitoring data, thus accurately calculating the comprehensive health index. The comprehensive health index is defined as a scalar between 0 and 1, quantifying the continuous state from complete health to complete failure, facilitating understanding and comparison, and providing a clear basis for decision-making. The advantage of the GBDT model lies in its ensemble learning characteristics, which reduce overfitting, improve generalization ability, and are suitable for the variable environment of industrial equipment. Using historical data for training ensures the model is validated and reliably applied to real-world scenarios. The comprehensive health index, as the output, makes the health status assessment objective, avoiding the bias of subjective judgment.

[0027] The specific steps for dynamically adjusting the acquisition frequency or scanning intensity of the insulation condition monitoring data in step S3 are as follows: When the value of the short-term trend is less than the first trend threshold, increase the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor. When the value of the short-term trend exceeds the second trend threshold, reduce the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor.

[0028] The methods for obtaining the first trend threshold and the second trend threshold are as follows: Collect historical data of electrical parameters of equipment during long-term normal operation; Calculate the absolute value of its change trend according to a fixed time window to obtain a list of normal trend values; The first trend threshold (lower limit) is taken as the 10th percentile of the list; if the trend value is lower than it, it indicates that the trend is extremely stable and the monitoring frequency can be reduced. The second trend threshold (upper limit) is taken as the 90th percentile of the list; if the trend value is higher than it, it indicates that the trend is drastic and the monitoring frequency needs to be increased.

[0029] In this embodiment, the specific conditions for dynamically adjusting the data acquisition frequency or scanning intensity of insulation condition monitoring are further refined, making the adjustment process more precise and controllable. When the value of the short-term trend is less than the first trend threshold, it indicates that the electrical performance data changes little and the equipment condition is relatively stable. At this time, increasing the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor can enhance the ability to capture potential risks and avoid missing subtle anomalies due to insufficient monitoring. For example, when the transformer is running smoothly, increasing the oil chromatograph analysis frequency can detect small changes in dissolved gases in the oil earlier and prevent insulation degradation. Conversely, when the value of the short-term trend is greater than the second trend threshold, it indicates that the electrical performance data changes significantly and the equipment condition may be changing rapidly. At this time, reducing the sampling frequency or detection sensitivity can reduce redundant data acquisition, save computing resources and energy consumption, and avoid equipment interference or data overload caused by excessive monitoring. By dynamically adjusting based on trend thresholds, the monitoring intensity is ensured to match the actual changes in equipment status. This guarantees high-sensitivity monitoring during critical periods while avoiding unnecessary waste during stable periods. It further optimizes the utilization of monitoring resources, improves the system's adaptability and overall efficiency, and provides reliable support for the intelligent maintenance of industrial equipment.

[0030] The method further includes: S4. When the comprehensive health index exceeds the preset alarm threshold, the internal structure of the transformer tank is located; the steps for this location are as follows: S41. Perform the k-th scan on the target area, divide it into grids and calculate the local anomaly index of each grid to identify the k-th hotspot cells. S42. Using the kth hotspot unit as the center, a smaller scanning area is redefined; S43. Within the reduced scanning area, perform the (k+1)th scan, wherein the grid resolution of the (k+1)th scan is higher than that of the kth scan, calculate the local anomaly index of each grid and identify the (k+1)th hot spot cell. S44. Let k = k + 1, iteratively execute steps S42 and S43 until the physical size of the scanned area is less than or equal to the preset accuracy threshold, and output the center coordinates of the hot spot unit identified in the last iteration as the final coordinates of the abnormal part. The accuracy threshold is a fixed physical length, determined according to maintenance requirements.

[0031] This example avoids the problems of excessively large range or insufficient accuracy that may occur in traditional positioning technologies. By using a gradual focusing method, it ensures the accuracy and efficiency of positioning; specifically, First, a preliminary scan of the target area is performed, dividing it into a grid and calculating local anomaly indices to identify hotspot cells. Then, the scan area is narrowed down centered on the hotspot cells, and a next scan is performed at a higher resolution. This process is iterated until the area size is smaller than the accuracy threshold. This effectively handles abnormal signals in complex environments, reduces false alarms and missed alarms, and improves the reliability of fault location. Simultaneously, it allows for the identification of potential fault points at an early stage, facilitating timely and targeted measures by maintenance personnel to prevent the fault from escalating. For example, inside a transformer tank, precise location of partial discharge or overheated hotspots can help maintenance personnel quickly implement repairs, avoiding damage to the entire equipment. Overall, this improves the accuracy and speed of fault location, reduces maintenance costs and time, and enhances the ability to ensure the safe operation of equipment.

[0032] In step S43, during the (k+1)th scan, the following steps are also included: The current scanning area is divided into three grid levels: low, medium, and high resolution. That is, the basic grid size is used as the low resolution, and the grid is scaled proportionally with a scaling factor of 2 to obtain the medium and high resolutions.

[0033] Calculate and identify hotspot units at each level; If a hot spot cell at a high resolution level is contained within a hot spot cell at a low resolution level, it is determined that the focus is consistent, and step S44 is continued to shrink the region. Otherwise, it is determined to be a focus inconsistency, and multi-focus tracking is initiated.

[0034] In this embodiment, multi-level resolution inspection can effectively handle situations with multiple anomaly sources or distributed faults, preventing errors caused by single-focus assumptions. For example, inside a transformer, multiple discharge points or hot zones may exist simultaneously. Through multi-level analysis, primary and secondary anomalies can be distinguished, improving the comprehensiveness of fault location. At the same time, by detecting inconsistencies in focus early on, a timely switch to multi-focus mode can be made, avoiding invalid iterations and saving computational resources. Overall, this improves the fault tolerance and efficiency of fault location.

[0035] The multi-focus tracking method is as follows: Sub-scan regions are created for each of the currently identified high-abnormality index units, and step S4 is executed in parallel for each sub-scan region.

[0036] In this embodiment, a multi-focus tracking approach is proposed to address the possibility of multiple anomalies. By creating sub-scanning regions for each of the currently identified high-anomaly index units and performing localization steps in parallel for each sub-scanning region, the ability to simultaneously handle multiple potential fault points is achieved. This avoids the limitations of traditional single-focus localization when facing distributed faults and enables comprehensive capture of multiple abnormal locations in the equipment. For example, in a large transformer, insulation defects may be distributed in multiple locations. Multi-focus tracking ensures that each point is scanned and evaluated independently, preventing the omission of minor but important anomalies.

[0037] In step S1, the electrical performance monitoring data includes high-frequency current signals, voltage harmonics, and reactive power curves. The insulation condition monitoring data includes dissolved gas content in the oil, partial discharge signal, and top oil temperature. The dissolved gas content in the oil is monitored by an online oil chromatograph, and the partial discharge signal is monitored by a partial discharge sensor.

[0038] High-frequency current signals and voltage harmonics can reflect transient changes and distortions in electrical systems. Reactive power curves indicate the power factor and operating efficiency of equipment. The dissolved gas content in oil is an important indicator of transformer insulation degradation. Partial discharge signals are directly related to insulation defects, while the top oil temperature reflects the heat dissipation status of the equipment.

[0039] In step S41, the specific process of calculating the local anomaly index of each grid and identifying the k-th hotspot cell is as follows: S441. Based on the spatial positioning monitoring data obtained from the k-th scan, the spatial positioning monitoring data includes one or more of ultrasonic signal amplitude, ultra-high frequency signal energy, and infrared thermometry data; the spatial positioning monitoring data is acquired based on a partial discharge sensor array, an ultrasonic sensor array, an ultra-high frequency sensor array, and an infrared thermal imager. S442. For each grid cell within the current scanning area, a comprehensive local anomaly index is calculated through weighted summation. The calculation formula is as follows: ; in, This is a local anomaly index. This is the normalized value of the ultrasonic signal amplitude. This is the normalized value of the UHF signal energy. This is the normalized value of the temperature. , , These are the weighting coefficients, and + + =1; weighting coefficients were determined using historical data and principal component analysis (PCA). S443. Compare the local anomaly indices of all grids, and identify the grid cell with the highest local anomaly index that exceeds the preset activity threshold as the k-th hotspot cell.

[0040] Example of a preset activity threshold: When the device is running normally, perform multiple scans to calculate the local anomaly index of all grids; calculate the mean μ and standard deviation σ of these indices; set the activity threshold to μ+3σ; only when the anomaly index of a grid significantly exceeds the background noise level is it identified as a valid hotspot.

[0041] This embodiment provides a specific method for calculating the local anomaly index. By using weighted summation, the amplitude of the ultrasonic signal, the energy of the ultra-high frequency signal, and the infrared thermometry data are fused into a comprehensive index, thereby accurately identifying hotspot units. The weighted summation formula of the local anomaly index allows for the allocation of weights according to the importance of different signals, enabling the local anomaly index to balance multiple factors and more accurately reflect the degree of anomaly.

[0042] For example, ultrasonic signal amplitude may be more sensitive to partial discharge, ultra-high frequency signal energy can indicate electromagnetic interference, and infrared thermometry data directly shows temperature anomalies, through a weighted system. , , Adjustments can optimize the index's adaptability to different fault types. The calculated local anomaly index is used to compare all grids, identifying the cells with the highest activity levels exceeding the activity threshold as hotspot cells. This ensures that only significant anomalies are further processed, reducing the impact of noise and thus improving the accuracy and reliability of localization, while avoiding misjudgments that may result from a single signal source.

[0043] The steps for obtaining the short-term trend value are as follows: S31. Obtain the historical sequence of each parameter in the electrical performance monitoring data within the most recent time window T, and calculate the standard deviation of the historical sequence respectively; S32. Select the parameter with the largest standard deviation as the key electrical parameter; S33. Perform linear regression analysis on the historical sequences corresponding to key electrical parameters and calculate the slope values ​​of the key electrical parameters; S34. The absolute value of the slope value is taken as the numerical value of the short-term trend.

[0044] This embodiment provides a process for obtaining short-term trend values. By analyzing the historical sequence of each parameter in the electrical performance monitoring data within a time window T, the standard deviation is calculated, and the parameter with the largest standard deviation is selected as the key electrical parameter. Then, linear regression analysis is performed on this parameter to obtain the slope value, and the absolute value of the slope is used as the short-term trend value. This ensures that the trend calculation is based on the most easily changing parameter, thereby sensitively capturing the dynamic changes in equipment status. The parameter with the largest standard deviation is selected as the key parameter because its changes are the most significant and may indicate potential problems. Linear regression analysis provides a quantitative measure of the trend, and the absolute value of the slope reflects the rate of change, avoiding directional interference. Through the above calculations, the short-term trend value can reliably represent the stability of the equipment status, providing a basis for dynamically adjusting the monitoring intensity.

[0045] For example: Based on historical data analysis and engineering settings, the following thresholds are preset: First trend threshold (lower limit): 0.1; Second trend threshold (upper limit): 0.2; Preset threshold: 0.85; Positioning accuracy threshold: 0.5 meters (side length of the scanning area); Hotspot activity threshold: 0.3; Local anomaly index weighting , , : 0.5, 0.3, 0.2; Assuming ultrasound localization is the most reliable; First, data is collected synchronously and health assessments are conducted. Data collected synchronously: Electrical performance data: Reactive power = 65kVar; Insulation status data: Partial discharge amplitude = 58dB, acetylene content in oil = 4ppm; Feature fusion and health assessment: These data, along with other unlisted data such as voltage harmonics and oil temperature, are input into a pre-trained Gradient Boosting Decision Tree (GBDT) model; the model outputs a comprehensive health index of 0.82.

[0046] judge: The overall health index is 0.82, which is less than the preset threshold of 0.85; therefore, the device has not entered an abnormal state.

[0047] Then: Calculate short-term trends and dynamic adjustments; Obtain short-term trend values; Time window T: the most recent 6 hours; Historical sequence: Obtain reactive power data over a 6-hour period: [50,52,48,55,60,65](kVar).

[0048] Standard deviation calculation: The standard deviation of this series is calculated to be 6.45; similarly, the standard deviations of other electrical parameters (such as current harmonics) are calculated, and it is found that the standard deviation of reactive power is the largest.

[0049] Determine the key electrical parameter: reactive power.

[0050] Linear regression: A linear regression was performed on the sequence [50,52,48,55,60,65], and the slope value k=2.3 was obtained.

[0051] Calculate the trend value: Trend value = |k| = 2.3.

[0052] Dynamic adjustment of insulation monitoring: The trend value 2.3 is greater than the second trend threshold 0.2; therefore, it is necessary to increase the intensity of insulation monitoring; the sampling frequency of the online oil chromatograph should be increased from once / 4 hours to once / 30 minutes. At the same time, the detection sensitivity of the partial discharge sensor should be increased by 10dB, and the scanning density should be doubled.

[0053] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting the health status of industrial equipment based on multi-source data fusion, characterized in that, Includes the following steps: S1. Synchronously collect electrical performance monitoring data and insulation status monitoring data of the transformer; S2. The features of the electrical performance monitoring data and insulation status monitoring data are fused and input into a pre-trained health assessment model to calculate a comprehensive health index; based on the comparison result of the comprehensive health index and a preset threshold, it is determined whether the equipment has entered an abnormal state. S3. When the industrial equipment is not in an abnormal state, continuously calculate the short-term trend of the electrical performance monitoring data; Based on the numerical value of the short-term trend, the acquisition frequency or scanning intensity of the insulation condition monitoring data is dynamically adjusted. The value of the short-term trend is negatively correlated with the acquisition frequency or scanning intensity.

2. The industrial equipment health status prediction method based on multi-source data fusion according to claim 1, characterized in that, The specific steps for dynamically adjusting the acquisition frequency or scanning intensity of the insulation condition monitoring data in step S3 are as follows: When the value of the short-term trend is less than the first trend threshold, increase the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor. When the value of the short-term trend exceeds the second trend threshold, reduce the sampling frequency of the online oil chromatograph or the detection sensitivity and scanning density of the partial discharge sensor.

3. The industrial equipment health status prediction method based on multi-source data fusion according to claim 2, characterized in that, The method further includes: S4. When the comprehensive health index exceeds the preset alarm threshold, the internal structure of the transformer tank is located; the steps for this location are as follows: S41. Perform the k-th scan on the target area, divide it into grids and calculate the local anomaly index of each grid to identify the k-th hotspot cells. S42. Using the kth hotspot unit as the center, a smaller scanning area is redefined; S43. Within the reduced scanning area, perform the (k+1)th scan, wherein the grid resolution of the (k+1)th scan is higher than that of the kth scan, calculate the local anomaly index of each grid and identify the (k+1)th hot spot cell. S44. Let k = k + 1, iteratively execute steps S42 and S43 until the physical size of the scanned area is less than or equal to the preset accuracy threshold, and output the center coordinates of the hot spot unit identified in the last iteration as the final coordinates of the abnormal part.

4. The industrial equipment health status prediction method based on multi-source data fusion according to claim 3, characterized in that, In step S43, during the (k+1)th scan, the following steps are also included: The current scanning area is simultaneously divided into three grid levels: low, medium, and high resolution. Calculate and identify hotspot units at each level; If a hot spot cell at a high resolution level is contained within a hot spot cell at a low resolution level, it is determined that the focus is consistent, and step S44 is continued to shrink the region. Otherwise, it is determined to be a focus inconsistency, and multi-focus tracking is initiated.

5. The industrial equipment health status prediction method based on multi-source data fusion according to claim 4, characterized in that, The multi-focus tracking method is as follows: Sub-scan regions are created for each of the currently identified high-abnormality index units, and step S4 is executed in parallel for each sub-scan region.

6. The method for predicting the health status of industrial equipment based on multi-source data fusion according to claim 1 or 5, characterized in that, In step S1, the electrical performance monitoring data includes high-frequency current signals, voltage harmonics, and reactive power curves. The insulation condition monitoring data includes dissolved gas content in the oil, partial discharge signal, and top oil temperature. The dissolved gas content in the oil is monitored by an online oil chromatograph, and the partial discharge signal is monitored by a partial discharge sensor.

7. The industrial equipment health status prediction method based on multi-source data fusion according to claim 3, characterized in that, In step S41, the specific process of calculating the local anomaly index of each grid and identifying the k-th hotspot cell is as follows: S441. Based on the spatial positioning monitoring data obtained from the kth scan, the spatial positioning monitoring data includes one or more of the following: ultrasonic signal amplitude, ultra-high frequency signal energy, and infrared thermometry data; S442. For each grid cell within the current scanning area, a comprehensive local anomaly index is calculated through weighted summation. The calculation formula is as follows: ； in, This is a local anomaly index. This is the normalized value of the ultrasonic signal amplitude. This is the normalized value of the UHF signal energy. This is the normalized value of the temperature. , , These are the weighting coefficients, and + + =1; S443. Compare the local anomaly indices of all grids, and identify the grid cell with the highest local anomaly index that exceeds the preset activity threshold as the k-th hotspot cell.

8. The industrial equipment health status prediction method based on multi-source data fusion according to claim 1, characterized in that, The steps for obtaining the short-term trend value are as follows: S31. Obtain the historical sequence of each parameter in the electrical performance monitoring data within the most recent time window T, and calculate the standard deviation of the historical sequence respectively; S32. Select the parameter with the largest standard deviation as the key electrical parameter; S33. Perform linear regression analysis on the historical sequences corresponding to key electrical parameters and calculate the slope values ​​of the key electrical parameters; S34. The absolute value of the slope value is taken as the numerical value of the short-term trend.

9. The industrial equipment health status prediction method based on multi-source data fusion according to claim 1, characterized in that, The pre-trained health assessment model is a machine learning model based on gradient boosting decision trees; The health assessment model is trained using historical data, which includes electrical performance monitoring data and insulation status monitoring data corresponding to the transformer under normal operating conditions and various known fault conditions. The comprehensive health index is a scalar between 0 and 1, used to quantitatively characterize the continuous health status of a transformer from complete health to complete failure.

Images