Power distribution equipment on-line monitoring system based on multi-modal data fusion

Abstract

The invention discloses a power distribution equipment on-line monitoring system based on multi-modal data fusion, and relates to the field of power distribution equipment management, and the system comprises a sensing module which is used for collecting an electrical signal, a thermal infrared signal, a mechanical vibration signal and an acoustic signal generated in the operation process of power distribution equipment through a multi-type sensing channel, converting the collected various signals into processable equipment state original data to form an equipment operation state original data set; according to the invention, through accurate acquisition of multi-dimensional signals, combination of space mapping and time delay compensation, data quality is optimized, the reliability of original information is ensured, key features are extracted according to working conditions, subtle state changes are captured by means of temperature field analysis and a variable-resolution spectrum technology, the comprehensiveness, accuracy and response timeliness of equipment operation monitoring are effectively improved, and the real-time performance of equipment operation monitoring is improved. Misjudgment and missed judgment are reduced, and fault risks are avoided in advance.

Description

Technical Field

[0001] This invention relates to the field of power distribution equipment management technology, specifically to an online monitoring system for power distribution equipment based on multimodal data fusion. Background Technology

[0002] Online monitoring of power distribution equipment collects data such as voltage, current, and temperature in real time through sensors, and analyzes the data to provide early warnings of potential faults. This enables visual status monitoring and timely alarms for anomalies, reducing power outage time, improving equipment reliability and grid maintenance efficiency, and ensuring a stable power supply to the power system.

[0003] Patent application number 202110287193.3 discloses an online monitoring method for power distribution equipment based on adaptive intelligent edge agents, comprising the following steps: Step 1, collecting data from remote power distribution equipment in the online monitoring system and preprocessing the data; Step 2, generating early warning information based on the preprocessed data, and sending the early warning information and the preprocessed data corresponding to the early warning information to the alarm or display device of the online monitoring system to display the current early warning; Step 3, if no early warning information is generated in Step 2, analyzing the power distribution equipment scene image based on a deep learning image recognition model and generating early warning information based on the analysis, and sending the early warning information and the power distribution equipment scene image corresponding to the early warning information. The data is transmitted to the alarm or display device of the online monitoring system to display the current warning; Step 4, repeat steps 1-3 to achieve real-time monitoring of the power distribution equipment based on the adaptive intelligent edge agent. This method aims to solve the problem that "the maintenance methods of power distribution equipment in the existing technology are still mainly based on planned maintenance mode. In planned maintenance mode, maintenance personnel will inspect and maintain the power distribution equipment according to a specific time cycle. This mode is not only labor-intensive, but also cannot monitor the power distribution equipment in real time. For power distribution equipment that fails outside the maintenance cycle, it is also difficult for maintenance personnel to respond in a timely manner and to understand the fault information in the first time, resulting in unreliable and unsafe operation of the power distribution equipment and affecting the reliable and stable operation of the power grid."

[0004] Compared to the disclosed technologies, those skilled in the art are eager for more novel and effective online monitoring technologies in the context of online monitoring of power distribution equipment.

[0005] To address this, we propose an online monitoring system for power distribution equipment based on multimodal data fusion. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides an online monitoring system for power distribution equipment based on multimodal data fusion, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0008] This invention discloses an online monitoring system for power distribution equipment based on multimodal data fusion, comprising:

[0009] The sensing module collects electrical signals, thermal infrared signals, mechanical vibration signals, and acoustic signals generated during the operation of power distribution equipment through multiple types of sensing channels. It converts the collected signals into processable raw equipment status data to form a raw equipment operating status dataset. The parsing module receives the raw equipment operating status dataset output by the sensing module and performs feature transformation processing based on preset feature extraction logic, converting the raw dataset into multi-type feature vectors representing the equipment operating status. The fusion module obtains the multi-type feature vectors transformed by the parsing module and performs fusion operations. It eliminates data heterogeneity through a modal adaptation algorithm, generating a comprehensive feature set containing multi-dimensional status information. Diagnosis... The first module receives a comprehensive feature set output from the fusion module, diagnoses the operating status of the power distribution equipment through a preset diagnostic logic model, identifies whether there are any abnormalities and the types of abnormalities, and simultaneously outputs a quantitative assessment result of the degree of abnormality. The second module acquires historical operating feature data of the equipment, combines the historical operating feature data with real-time status feature information, and uses a time-series feature inference model to predict the future operating status of the equipment. When the prediction result exceeds a preset safety threshold, a warning signal is triggered. The third module receives the abnormal assessment results from the diagnostic module and the warning signal from the warning module, and generates appropriate equipment operating parameter adjustment instructions based on the linkage logic constructed by the equipment's rated parameters and operating safety thresholds.

[0010] Furthermore, the sensing module obeys the following during the data acquisition phase:

[0011] For electrical signals and thermal infrared signals, a spatial coordinate mapping relationship is established based on the physical topology of the power distribution equipment, so that the spatial correspondence error between the electrical quantity acquisition point and the thermal infrared imaging area does not exceed 0.3m;

[0012] For mechanical vibration signals and acoustic signals, propagation time delay compensation is performed based on the medium characteristics of the equipment casing. The time delay compensation value is calculated according to the formula. Perform calculations. The straight-line distance between the vibration sensor and the acoustic sensor. The vibration-sound propagation velocity of the equipment casing. The angle between the direction of propagation and the normal to the shell;

[0013] Signal quality factor is calculated in real time during the acquisition process. , Signal-to-noise ratio (SNR) This represents the time synchronization accuracy coefficient. The spatial offset coefficient is used to trigger signal re-acquisition when Q < 15, until Q ≥ 15 after 3 consecutive acquisitions, at which point the data is included in the original dataset of device status.

[0014] Among them, the time synchronization accuracy coefficient The value follows the condition: synchronization error ≤ 5 The value is 1.0 at the beginning, and increases by 5 every time. Attenuation 0.1, spatial offset coefficient The value follows the following rules: 1.0 when the spatial error is ≤0.3m, and 0.2 for every additional 0.1m.

[0015] Furthermore, the preset feature extraction logic of the parsing module is as follows:

[0016] When the equipment is under light load conditions, i.e., load rate < 30%, harmonic characteristics are extracted first from the electrical signal, and the characteristic frequency is: , Indicates the fundamental frequency;

[0017] When the equipment is under heavy load, i.e., the load rate is ≥30%, transient features are extracted first, and the feature quantity is the transient pulse amplitude. ;

[0018] For thermal infrared signals, the temperature field entropy value is used. Distinguishing between uniform heating and localized overheating characteristics <0.6 indicates a tendency for localized overheating;

[0019] In the formula: t represents the real-time sampled value of the electrical signal under heavy load conditions, where t is time. This represents the steady-state mean of the electrical signal. The number of temperature range divisions; This represents the area percentage of the i-th temperature range.

[0020] Furthermore, the modal adaptation algorithm of the fusion module achieves fusion operation through modal complementarity quantization:

[0021] Define the modal complementarity index , This represents the joint information entropy of modes m and n. The information entropy of mode m is represented by the information entropy. The information entropy of mode n is represented by... Represents cross-information entropy, when Strengthen the correlation between two modal features;

[0022] A correlation enhancement factor is generated based on the covariance value and the modal complementarity index. The correlation enhancement factor increases linearly with the increase of both the covariance value and the modal complementarity index.

[0023] Comprehensive feature set The generating formula is , This indicates the total number of modal types participating in the fusion. This represents the dynamic weighting coefficient of the m-th mode at time t. This represents the feature vector obtained after the m-th mode is processed by the parsing module;

[0024] in, , This represents the sum of complementarity of the m-th mode at time t. This represents the information entropy of the m-th mode. This represents the sum of complementarity of the nth mode at time t. The information entropy of the nth mode is represented.

[0025] Furthermore, the preset diagnostic logic model of the diagnostic module is a causal attention deep belief network, which sequentially includes an input layer, a causal inference layer, an attention weighting layer, and an output layer:

[0026] The causal inference layer is used to mine the causal relationship between various features and equipment anomaly results. It quantifies the contribution of features to anomalies through causal strength, and the formula for calculating causal strength is:

[0027] ;

[0028] In the formula: Indicates the existence of features And exclude other features The probability of an anomaly occurring at that time; Indicates the absence of features And the probability of anomalies when other features are retained;

[0029] The attention weighting layer assigns dynamic weights to features based on causal strength, with weight coefficients... L is the number of features in the comprehensive feature set, and only applies to... Causal strength features with a value ≥0.2 are retained in the weighted allocation.

[0030] The abnormality level assessment results are represented by the abnormality confidence level S. When S≥1.2, it is judged as a serious abnormality, when 0.8≤S<1.2, it is judged as a mild abnormality, and when S<0.8, it is judged as normal.

[0031] ;

[0032] In the formula: is the normalized value of the i-th feature; Preset security feature thresholds.

[0033] Furthermore, the time-series feature inference model of the early warning module follows the following:

[0034] Short-term trend forecast:

[0035] ;

[0036] The formula for dynamically adjusting the safety threshold is:

[0037] ;

[0038] In the formula: The predicted value of the device state characteristics at time t+k in the future; It is a bidirectional long short-term memory network; Let w be the sequence of device state characteristics from the current time t to the historical time t−w, where w represents the length of the sliding time window; The environmental parameter vector at time t includes at least temperature and humidity;

[0039] The device security threshold is dynamically updated at time t. This is the initial safety threshold for the device; For aging coefficient and environmental impact coefficient; This refers to the cumulative operating time of the device. The baseline runtime is fixed at 1000 hours. The deviation rate of environmental parameters from the baseline value at time t;

[0040] in, The value ranges are 0.02~0.06 and 0.01~0.03 respectively, and the warning trigger condition is... And it continues to predict for 3 cycles.

[0041] Furthermore, when the analysis module extracts the frequency domain features of the mechanical vibration signal, it does so based on variable resolution spectrum analysis:

[0042] High-resolution analysis, i.e., frequency resolution, is used in the resonant frequency range of the device. The characteristic frequency is calculated as follows: , Indicates the frequency point index. Indicates the signal sampling frequency. Indicates the number of sampling points used in high-resolution analysis;

[0043] Low-resolution analysis, i.e., frequency resolution, is used in the non-resonance region. The characteristic frequency is calculated as follows: , Indicates one-quarter ;

[0044] Through spectral energy concentration Screening key frequency domain features;

[0045] In the formula: This represents the total energy value within the resonant frequency range of the device. This represents the total energy value across the entire frequency range.

[0046] The resonant frequency range of the equipment is set to 500-2000Hz. Resonance characteristics are preserved when the value is greater than 0.4.

[0047] Furthermore, the linkage logic of the response module is dynamically adjusted based on the device's health status:

[0048] ;

[0049] When the diagnostic abnormality level is severe, and When <0.6, a load reduction instruction is generated, and the load reduction amount is: ;

[0050] When the anomaly is mild, a step-wise adjustment command is generated, with the adjustment amount being: ;

[0051] After adjustment, the rate of health recovery is high. To verify the effect, Upgrade adjustment when <0;

[0052] In the formula: The device health assessment value at time t; The number of equipment loss types, including electrical loss, mechanical wear loss, and thermal aging loss; Let be the weighting coefficient for the i-th type of loss. The greater the impact of loss on equipment life, the larger the value, and it follows the principle of weighting coefficients. ; Let be the cumulative amount of the i-th type of loss at time t; The load shedding ratio for emergency load reduction commands; For the step-wise adjustment of equipment operating parameters; The health association coefficient; For abnormal confidence levels; To adjust the observation time interval after the command is executed;

[0053] in, The sensitivity setting should be adjusted according to the equipment type and parameters. For distribution transformers, the value should be 5-10, and for switchgear, it should be 3-6. The higher the parameter adjustment sensitivity, the larger the value should be.

[0054] Furthermore, during the operation phase of the fusion module, the feature vectors of each modality are processed... Make corrections:

[0055] ;

[0056] In the formula: is the feature vector of the m-th mode after outlier correction; These are the characteristic correction coefficients for the m-th mode; Let be the mean of the eigenvectors of the m-th mode; Let be the standard deviation of the eigenvector of the m-th mode; This is a sign function; it outputs 1 when the input value is positive, -1 when it is negative, and 0 when it is 0.

[0057] in, according to The signal stability adaptively takes a value in the range of 0.8-1.2. The higher the signal stability, the closer the value is to 1.2, and vice versa. The signal stability is characterized by a combination of characteristic fluctuation variance and signal-to-noise ratio.

[0058] Furthermore, the sensing module is interconnected with the parsing module and the fusion module via a wireless network. The fusion module is interconnected with the diagnostic module and the early warning module via a wireless network. The diagnostic module and the early warning module are interconnected with the response module via a wireless network.

[0059] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0060] This invention provides an online monitoring system for power distribution equipment based on multimodal data fusion. During operation, the system accurately acquires multi-dimensional signals, optimizes data quality through spatial mapping and time delay compensation to ensure the reliability of raw information, extracts key features according to different operating conditions, captures subtle state changes using temperature field analysis and variable resolution spectrum technology, eliminates data heterogeneity through modal adaptation, strengthens feature correlation to achieve comprehensive state characterization, quantifies anomaly contribution based on causal correlation, accurately identifies anomaly types and degrees, dynamically infers trends by combining historical data and environmental parameters, dynamically adjusts safety thresholds to achieve early warning, and dynamically generates adjustment commands based on health status and verifies the effects. This effectively improves the comprehensiveness, accuracy, and timeliness of equipment operation monitoring, reduces misjudgments and omissions, avoids fault risks in advance, ensures the safe and stable operation of the power distribution system, and reduces maintenance costs and downtime losses. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0062] Figure 1 This is a schematic diagram of the structure of an online monitoring system for power distribution equipment based on multimodal data fusion. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0064] The present invention will be further described below with reference to embodiments.

[0065] Example:

[0066] This embodiment presents an online monitoring system for power distribution equipment based on multimodal data fusion, such as... Figure 1 As shown, it includes:

[0067] The sensing module is used to collect electrical signals, thermal infrared signals, mechanical vibration signals and acoustic signals generated during the operation of power distribution equipment through multiple types of sensing channels, and convert the collected signals into raw equipment status data that can be processed to form a raw dataset of equipment operating status.

[0068] The sensing module follows the following rules during the data acquisition phase:

[0069] For electrical signals and thermal infrared signals, a spatial coordinate mapping relationship is established based on the physical topology of the power distribution equipment, so that the spatial correspondence error between the electrical quantity acquisition point and the thermal infrared imaging area does not exceed 0.3m;

[0070] For mechanical vibration signals and acoustic signals, propagation time delay compensation is performed based on the medium characteristics of the equipment casing. The time delay compensation value is calculated according to the formula. Perform calculations. The straight-line distance between the vibration sensor and the acoustic sensor. The vibration-sound propagation velocity of the equipment casing. The angle between the direction of propagation and the normal to the shell;

[0071] Signal quality factor is calculated in real time during the acquisition process. , Signal-to-noise ratio (SNR) This represents the time synchronization accuracy coefficient. The spatial offset coefficient is used to trigger signal re-acquisition when Q < 15, until Q ≥ 15 after 3 consecutive acquisitions, at which point the data is included in the original dataset of device status.

[0072] Among them, the time synchronization accuracy coefficient The value follows the condition: synchronization error ≤ 5 The value is 1.0 at the beginning, and increases by 5 every time. Attenuation 0.1, spatial offset coefficient The value follows this rule: 1.0 is taken when the spatial error is ≤0.3m, and 0.2 is added for every additional 0.1m.

[0073] The parsing module receives the raw dataset of device operating status output by the perception module, performs feature transformation processing based on the preset feature extraction logic, and converts the raw dataset into multi-type feature vectors representing the device operating status.

[0074] The default feature extraction logic of the parsing module is as follows:

[0075] When the equipment is under light load conditions, i.e., load rate < 30%, harmonic characteristics are extracted first from the electrical signal, and the characteristic frequency is: , Indicates the fundamental frequency;

[0076] When the equipment is under heavy load, i.e., the load rate is ≥30%, transient features are extracted first, and the feature quantity is the transient pulse amplitude. ;

[0077] For thermal infrared signals, the temperature field entropy value is used. Distinguishing between uniform heating and localized overheating characteristics <0.6 indicates a tendency for localized overheating;

[0078] In the formula: t represents the real-time sampled value of the electrical signal under heavy load conditions, where t is time. This represents the steady-state mean of the electrical signal. The number of temperature range divisions; This represents the area percentage of the i-th temperature range;

[0079] When the parsing module extracts the frequency domain features of the mechanical vibration signal, it does so based on variable resolution spectrum analysis:

[0080] High-resolution analysis, i.e., frequency resolution, is used in the resonant frequency range of the device. The characteristic frequency is calculated as follows: , Indicates the frequency point index. Indicates the signal sampling frequency. Indicates the number of sampling points used in high-resolution analysis;

[0081] Low-resolution analysis, i.e., frequency resolution, is used in the non-resonance region. The characteristic frequency is calculated as follows: , Indicates one-quarter ;

[0082] Through spectral energy concentration Screening key frequency domain features;

[0083] In the formula: This represents the total energy value within the resonant frequency range of the device. This represents the total energy value across the entire frequency range.

[0084] The resonant frequency range of the equipment is set to 500-2000Hz. Resonance characteristics are preserved when >0.4;

[0085] The fusion module is used to obtain the multi-type feature vectors obtained by the parsing module and perform fusion operations. It eliminates data heterogeneity through modal adaptation algorithm and generates a comprehensive feature set containing multi-dimensional state information.

[0086] The modal adaptation algorithm of the fusion module achieves fusion operation through modal complementarity quantization:

[0087] Define the modal complementarity index , This represents the joint information entropy of modes m and n. The information entropy of mode m is represented by the information entropy. The information entropy of mode n is represented by... Represents cross-information entropy, when Strengthen the correlation between two modal features;

[0088] The above-mentioned process for enhancing the correlation between two modal features includes:

[0089] In the fusion operation, the initial dynamic weights are first calculated based on the information entropy and complementarity of each modality;

[0090] The correlation enhancement factor is generated by performing element-wise multiplication on the feature vectors of mode m and mode n to produce a set of related features and calculating their covariance to characterize the linear correlation strength. The correlation enhancement factor is generated in the following way:

[0091] Using covariance value and The difference exceeding 0.3 is used as input to establish a linear mapping relationship between the two and the enhancement factor (the mapping coefficient makes the enhancement factor have a minimum value of 0 and a maximum value of 0.5), that is, the enhancement factor increases with the increase of the covariance value and The difference exceeding the threshold increases linearly, and the value range of the enhancement factor is limited to 0 to 50% of the initial dynamic weight value. The initial dynamic weight of mode m is superimposed and corrected using the enhancement factor. At the same time, the above-mentioned associated feature set is integrated into the feature vector of mode m at a preset ratio not exceeding 20% ​​of the original feature vector weight of mode m. Then, the corrected dynamic weight is combined to participate in the fusion operation of the comprehensive feature set.

[0092] A correlation enhancement factor is generated based on the covariance value and the modal complementarity index. The correlation enhancement factor increases linearly with the increase of both the covariance value and the modal complementarity index.

[0093] Comprehensive feature set The generating formula is , This indicates the total number of modal types participating in the fusion. This represents the dynamic weighting coefficient of the m-th mode at time t. This represents the feature vector obtained after the m-th mode is processed by the parsing module;

[0094] in, , This represents the sum of complementarity of the m-th mode at time t. This represents the information entropy of the m-th mode. This represents the sum of complementarity of the nth mode at time t. The information entropy represents the nth mode;

[0095] During the fusion module's operation phase, the feature vectors of each modality are processed. Make corrections:

[0096] ;

[0097] In the formula: is the feature vector of the m-th mode after outlier correction; These are the characteristic correction coefficients for the m-th mode; Let be the mean of the eigenvectors of the m-th mode; Let be the standard deviation of the eigenvector of the m-th mode; This is a sign function; it outputs 1 when the input value is positive, -1 when it is negative, and 0 when it is 0.

[0098] in, according to The signal stability adaptively takes a value in the range of 0.8-1.2. The higher the signal stability, the closer the value is to 1.2, and vice versa. The signal stability is characterized by a combination of characteristic fluctuation variance and signal-to-noise ratio.

[0099] The above formula corrects for outliers in the original eigenvector using the mean, standard deviation, and sign function, with the correction coefficient β. m The signal stability (characterized by a combination of characteristic fluctuation variance and signal-to-noise ratio) is adaptively selected in the range of 0.8-1.2. The higher the stability, the closer the value is to 1.2, and vice versa. This optimizes the quality of the feature vector and reduces the interference of outliers on subsequent analysis.

[0100] The diagnostic module receives a comprehensive feature set output by the fusion module, diagnoses the operating status of the power distribution equipment through a preset diagnostic logic model, identifies whether the equipment has any abnormalities and the types of abnormalities, and simultaneously outputs a quantitative assessment result of the degree of abnormality.

[0101] The pre-defined diagnostic logic model for the diagnostic module is a causal attention deep belief network, which consists of an input layer, a causal inference layer, an attention weighting layer, and an output layer.

[0102] The causal inference layer is used to uncover the causal relationships between various features and equipment anomalies. It quantifies the contribution of features to anomalies through causal strength, calculated using the formula:

[0103] ;

[0104] In the formula: Indicates the existence of features And exclude other features The probability of an anomaly occurring at that time; Indicates the absence of features And the probability of anomalies when other features are retained;

[0105] The above formula quantifies the contribution of features to abnormal results by calculating the difference between the probability of anomalies occurring when a certain feature exists and other features are excluded, and the probability of anomalies occurring when the feature does not exist and other features are retained.

[0106] Attention-weighted layers assign dynamic weights to features based on causal strength, with weight coefficients... L is the number of features in the comprehensive feature set, and only applies to... Causal strength features with a value ≥0.2 are retained in the weighted allocation.

[0107] The abnormality level assessment results are represented by the abnormality confidence level S. When S≥1.2, it is judged as a serious abnormality, when 0.8≤S<1.2, it is judged as a mild abnormality, and when S<0.8, it is judged as normal.

[0108] ;

[0109] In the formula: is the normalized value of the i-th feature; Preset security feature thresholds;

[0110] The above formula combines the feature normalization value and the safety feature threshold to calculate the anomaly confidence S, and the S value is used to achieve a precise quantitative assessment of the degree of anomaly.

[0111] The early warning module is used to acquire historical operating characteristic data of the equipment, combine the historical operating characteristic data with real-time status characteristic information, and predict the future operating status of the equipment through a time-series characteristic inference model. When the prediction result exceeds the preset safety threshold, an early warning signal is triggered and output.

[0112] The time-series feature inference model of the early warning module follows:

[0113] Short-term trend forecast:

[0114] ;

[0115] The formula for dynamically adjusting the safety threshold is:

[0116] ;

[0117] In the formula: The predicted value of the device state characteristics at time t+k in the future; It is a bidirectional long short-term memory network; Let w be the sequence of device state characteristics from the current time t to the historical time t−w, where w represents the length of the sliding time window; The environmental parameter vector at time t includes at least temperature and humidity;

[0118] The device security threshold is dynamically updated at time t. This is the initial safety threshold for the device; For aging coefficient and environmental impact coefficient; This refers to the cumulative operating time of the device. The baseline runtime is fixed at 1000 hours. The deviation rate of environmental parameters from the baseline value at time t;

[0119] in, The value ranges are 0.02~0.06 and 0.01~0.03 respectively, and the warning trigger condition is... And it continues to predict for 3 cycles;

[0120] In the above formula, short-term trend prediction uses a bidirectional long short-term memory network (BiLSTM) to infer the device state at future moments based on the state feature sequence and environmental parameter vector within the current and historical sliding windows;

[0121] The safety threshold is dynamically adjusted by combining the initial threshold, aging coefficient, environmental impact coefficient, cumulative running time and environmental parameter deviation rate. When the predicted value exceeds the dynamic threshold for three consecutive cycles, an early warning is triggered, realizing a forward-looking state warning that takes into account equipment aging and environmental changes.

[0122] The response module receives the abnormal assessment results from the diagnostic module and the warning signals from the early warning module, and generates appropriate equipment operating parameter adjustment instructions based on the linkage logic constructed by the equipment's rated parameters and operating safety thresholds.

[0123] The response module's linkage logic is dynamically adjusted based on device health:

[0124] ;

[0125] When the diagnostic abnormality level is severe, and When <0.6, a load reduction instruction is generated, and the load reduction amount is: ;

[0126] When the anomaly is mild, a step-wise adjustment command is generated, with the adjustment amount being: ;

[0127] After adjustment, the rate of health recovery is high. To verify the effect, Upgrade adjustment when <0;

[0128] In the formula: The device health assessment value at time t; The number of equipment loss types, including electrical loss, mechanical wear loss, and thermal aging loss; Let be the weighting coefficient for the i-th type of loss. The greater the impact of loss on equipment life, the larger the value, and it follows the principle of weighting coefficients. ; Let be the cumulative amount of the i-th type of loss at time t; The load shedding ratio for emergency load reduction commands; For the step-wise adjustment of equipment operating parameters; The health association coefficient; For abnormal confidence levels; To adjust the observation time interval after the command is executed;

[0129] in, The sensitivity setting should be adjusted according to the equipment type and parameters. For distribution transformers, the value should be 5-10, and for switchgear, it should be 3-6. The higher the parameter adjustment sensitivity, the larger the value should be.

[0130] The above formula calculates equipment health by comprehensively considering the cumulative amount of various losses, including electrical and mechanical wear, and thermal aging, and assigning weights according to the degree of impact of these losses on lifespan; serious abnormalities and When the value is <0.6, the burden reduction ratio is determined according to the health correlation coefficient and the abnormality confidence level; when the abnormality is mild, the adjustment instruction is generated by the step adjustment formula, and the effect is verified according to the health recovery rate after adjustment. If the expected result is not achieved, the adjustment is upgraded to realize dynamic adaptive adjustment based on health status.

[0131] The sensing module interacts with the parsing module and the fusion module via a wireless network. The fusion module interacts with the diagnostic module and the early warning module via a wireless network. The diagnostic module and the early warning module interact with the response module via a wireless network.

[0132] In this embodiment, the sensing module collects electrical signals, thermal infrared signals, mechanical vibration signals, and acoustic signals generated during the operation of the power distribution equipment through multiple types of sensing channels. It converts the collected signals into processable raw equipment status data to form a raw equipment operating status dataset. The parsing module then receives this raw dataset from the sensing module and performs feature transformation processing based on preset feature extraction logic, converting the raw dataset into multi-type feature vectors representing the equipment operating status. The fusion module then obtains these multi-type feature vectors from the parsing module and performs a fusion operation. A modal adaptation algorithm eliminates data heterogeneity, generating a comprehensive feature set containing multi-dimensional status information. The diagnostic module further receives the comprehensive feature set output by the fusion module, diagnoses the operating status of the power distribution equipment through a preset diagnostic logic model, identifies whether there are any abnormalities in the equipment and the type of abnormality, and outputs the quantitative assessment result of the degree of abnormality simultaneously. It also uploads the historical operating feature data of the equipment through the early warning module, combines the historical operating feature data of the equipment with the real-time status feature information, and uses a time-series feature inference model to predict the future operating status of the equipment. When the prediction result exceeds the preset safety threshold, an early warning signal is triggered. Finally, the response module receives the abnormality assessment result from the diagnostic module and the early warning signal from the early warning module, and generates an appropriate equipment operating parameter adjustment instruction based on the linkage logic constructed by the rated parameters of the equipment and the operating safety threshold.

[0133] The system in the above embodiments accurately collects multiple types of signals through multiple sensor channels and ensures data quality. After feature fusion to eliminate heterogeneity, it can accurately diagnose the type and degree of equipment anomalies. Combined with time-series deduction, it can provide early warning of potential risks. Then, it can dynamically generate adjustment instructions based on health status, so as to keep track of equipment status in real time, avoid faults in advance, reduce downtime losses, and intelligently optimize operating parameters, effectively improving the reliability of power distribution equipment and the safety of power supply.

[0134] The following example illustrates an application instance of the system described in the above embodiments:

[0135] A 110kV substation is equipped with an online monitoring system for power distribution equipment based on multi-modal data fusion. This system enables real-time status monitoring and intelligent control of the 10kV switchgear cluster within the substation. The specific application process is as follows:

[0136] Data acquisition from the sensing module:

[0137] The system's sensing module acquires data through multiple types of sensing channels. For both electrical and thermal infrared signals, a spatial coordinate mapping relationship is established based on the physical topology of the switchgear. After calibration, the spatial correspondence error between the electrical quantity acquisition points and the thermal infrared imaging area is controlled within 0.25m, meeting the requirement of not exceeding 0.3m.

[0138] For both mechanical vibration and acoustic signals, propagation delay compensation is performed considering the medium characteristics of the equipment housing. The straight-line distance between the vibration sensor and the acoustic sensor of the switchgear at this station is 1.5m. The vibration-sound propagation velocity of the equipment housing is 5000m / s, and the angle between the propagation direction and the normal to the housing is 30 degrees. The calculated delay compensation value is 0.000327 seconds.

[0139] During the acquisition process, the signal quality factor is calculated in real time. In a certain acquisition, the signal-to-noise ratio is 20, the time synchronization error is 3 (corresponding to a time synchronization accuracy coefficient of 1.0), the spatial offset is 0.2m (corresponding to a spatial offset coefficient of 1.0), and the final signal quality factor is 20. Since it is ≥15, it is directly included in the original dataset of device status.

[0140] Feature extraction from the parsing module:

[0141] The analysis module dynamically adjusts the feature extraction logic based on the equipment operating conditions. When the switchgear is under light load (load rate 25% < 30%), harmonic features are extracted first from the electrical signals, with a focus on capturing the 2nd, 3rd, and 5th harmonics at the fundamental frequency of 50Hz. When some switchgear enters heavy load condition (load rate 35% ≥ 30%) due to load fluctuations, transient features are extracted first, and key characteristic quantities such as transient pulse amplitude are recorded.

[0142] For thermal infrared signals, the type of heat generation is distinguished by calculating the temperature field entropy value. The temperature field entropy value of phase A of switch cabinet #2 is 0.5, which is determined to be less than 0.6 and thus indicates a tendency for local overheating.

[0143] When performing variable-resolution spectral analysis on mechanical vibration signals, high-resolution analysis is used in the resonant frequency range of 500-2000Hz, with the frequency resolution calculated based on the sampling frequency and the number of high-resolution sampling points. Low-resolution analysis is used in the non-resonant range (the number of low-resolution sampling points is one-quarter of the number of high-resolution sampling points). Based on the calculated spectral energy concentration, the total energy within the resonant frequency range accounts for 0.45% of the total energy across the entire frequency range; since this value is greater than 0.4, the resonant characteristics are preserved.

[0144] Feature fusion of the fusion module:

[0145] The fusion module eliminates data heterogeneity through a modal adaptation algorithm. The complementarity index between each mode is calculated. After analysis, the joint information entropy, individual information entropy, and cross-information entropy of electrical signals and thermal infrared signals show a high complementarity index, and the system strengthens the correlation between their characteristics. Due to the strong physical correlation between vibration signals and acoustic signals, the complementarity index also meets the condition for strengthening the correlation.

[0146] A correlation enhancement factor is generated based on the covariance value and modal complementarity index, and the weights of each mode are dynamically allocated. Finally, the dynamic weight coefficients of the electrical mode at time t are determined to be 0.3, the thermal infrared mode to be 0.25, the vibration mode to be 0.2, and the acoustic mode to be 0.25. A comprehensive feature set containing multi-dimensional state information is generated through weighted fusion.

[0147] Diagnostic module status assessment:

[0148] The diagnostic module employs a causal attention deep belief network for state diagnosis. The causal inference layer mines the correlation between features and anomalies, calculating a causal strength of 0.7 for local overheating and 0.5 for transient pulse features, both ≥0.2, and is included in the weight allocation. The attention-weighted layer, after assigning dynamic weights based on causal strength, calculates an anomaly confidence level of 1.3. Since this confidence level is ≥1.2, a serious anomaly is determined in phase A of switchgear #2.

[0149] Early warning module trend prediction:

[0150] The early warning module combines historical data with real-time information to perform trend prediction. Through a bidirectional long short-term memory network, based on the state feature sequence within the current and past sliding windows, and combined with environmental parameter vectors of 35℃ temperature and 60% humidity, it predicts the state feature values ​​for the next three periods.

[0151] Meanwhile, based on the cumulative operating time of the equipment and the deviation rate of environmental parameters from the baseline value, the initial safety threshold is dynamically adjusted to 1.1. Because the predicted value of 1.4 exceeded the safety threshold of 1.1 for three consecutive cycles, the system triggered an early warning signal.

[0152] Intelligent control of response module:

[0153] The response module generates adjustment instructions based on the diagnostic results and early warning signals. Calculations show that the current health assessment value of switchgear #2 is 0.5. Since this value is less than 0.6 and the system is diagnosed as severely abnormal, it generates a load reduction instruction of 15%.

[0154] After the adjustment command was executed, the system monitored the equipment at the set observation time interval and found that the health recovery rate was negative, indicating that the adjustment effect did not meet expectations. The system then upgraded the adjustment strategy and further optimized the equipment operating parameters.

[0155] Through the system's full-process monitoring and control, the substation successfully achieved early detection and rapid handling of power distribution equipment anomalies, effectively improving the safety and reliability of equipment operation.

[0156] In summary, during operation, the system in the above embodiments accurately acquires multi-dimensional signals, optimizes data quality through spatial mapping and time delay compensation to ensure the reliability of original information, extracts key features according to different operating conditions, captures subtle state changes with the help of temperature field analysis and variable resolution spectrum technology, eliminates data heterogeneity through modal adaptation, strengthens feature correlation to achieve comprehensive state characterization, quantifies the contribution of anomalies based on causal correlation, accurately identifies the type and degree of anomalies, dynamically infers trends by combining historical data and environmental parameters, dynamically adjusts safety thresholds to achieve early warning, and dynamically generates adjustment instructions based on health status and verifies the effect. This effectively improves the comprehensiveness, accuracy and timeliness of equipment operation monitoring, reduces misjudgments and omissions, avoids fault risks in advance, ensures the safe and stable operation of the power distribution system, and reduces maintenance costs and downtime losses.

[0157] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power distribution equipment online monitoring system based on multi-modal data fusion, characterized in that, The method comprises the following steps: A perception module is used to collect electrical signals, thermal infrared signals, mechanical vibration signals and acoustic signals generated during the operation of power distribution equipment through multiple types of sensing channels, and convert the collected signals into device state raw data for forming a device operation state raw data set; An analysis module is used to receive the device operation state raw data set output by the perception module, and perform feature conversion processing based on a preset feature extraction logic to convert the raw data set into a multi-type feature vector representing the device operation state; A fusion module is used to obtain the multi-type feature vector converted by the analysis module and perform fusion operation, eliminate data heterogeneity through a modal adaptation algorithm, and generate a comprehensive feature set containing multi-dimensional state information; A diagnosis module is used to receive the comprehensive feature set output by the fusion module, diagnose the device operation state through a preset diagnosis logic model, identify whether the device has an abnormality and the type of abnormality, and synchronously output an abnormality degree quantitative evaluation result; An early warning module is used to obtain device historical operation feature data, combine the device historical operation feature data with real-time state feature information, and perform trend prediction on the future operation state of the device through a time series feature deduction model, and trigger an early warning signal output when the prediction result exceeds a preset safety threshold; A response module is used to receive the abnormality evaluation result in the diagnosis module and the early warning signal in the early warning module, and generate an adaptive device operation parameter adjustment instruction based on the linkage logic constructed based on the device rated parameters and the operation safety threshold.

2. The power distribution equipment online monitoring system based on multi-modal data fusion according to claim 1, characterized in that, In the data collection stage, the perception module is subject to: For electrical signals and thermal infrared signals, a spatial coordinate mapping relationship is established based on the physical topology of the power distribution equipment, so that the spatial correspondence error between the electrical quantity collection point and the thermal infrared imaging area is not more than 0.3m; For mechanical vibration signal and acoustic signal, based on the medium characteristics of the equipment shell to carry out delay compensation, the delay compensation value is calculated according to the formula , is the straight-line distance of the vibration sensor and the acoustic sensor, is the vibration-acoustic propagation speed of the equipment shell, is the angle between the propagation direction and the normal of the shell; Real-time calculation of signal quality factor during acquisition , S is the signal-to-noise ratio of the signal, T is the time synchronization accuracy coefficient, S is the spatial offset coefficient, triggering signal reacquisition when Q < 15, until Q ≥ 15 for 3 consecutive acquisitions, the data is included in the device state raw data set; Wherein, the time synchronization accuracy coefficient The value is subject to: synchronization error ≤5 Take 1.0, increase 0.2 every 5 Decay 0.1, spatial offset coefficient The value is subject to: spatial error ≤0.3m take 1.0, increase 0.2 every 0.1m.

3. The power distribution equipment online monitoring system based on multi-modal data fusion according to claim 1, characterized in that, The preset feature extraction logic of the analysis module is: When the device is in light load condition, i.e. load ratio < 30%, the harmonic characteristics are extracted preferentially for the electrical signal, and the characteristic frequency is , represents the fundamental frequency; When the device is in heavy load working condition, i.e. the load rate ≥ 30%, the transient feature is extracted preferentially, and the characteristic quantity thereof is the transient pulse amplitude ; For thermal infrared signals, the entropy value of the temperature field is used to distinguish between uniform heating and local overheating characteristics, <0.6 is determined as a local overheating tendency; In the formula: is the real-time sampling value of the electrical signal under heavy load conditions, t is time; is the steady-state average value of the electrical signal; is the number of temperature intervals; is the area ratio of the i th temperature interval.

4. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The modal adaptation algorithm of the fusion module realizes fusion operation through modal complementarity quantification: Definition of the modality complementarity index , denotes the joint information entropy of modalities m and n, denotes the information entropy of modality m, denotes the information entropy of modality n, denotes the cross information entropy, when the correlation of the features of the two modalities is emphasized. An association enhancement factor is generated based on the covariance value and the modal complementarity index, and the association enhancement factor linearly increases with the increase of the covariance value and the modal complementarity index Comprehensive feature set The generation formula is , represents the total number of modal types participating in fusion, represents the dynamic weight coefficient of the mth modal at time t, represents the feature vector obtained after the mth modal is processed by the analysis module. wherein, , denotes the sum of complementarity of the mth modality at time t, denotes the information entropy of the mth modality, denotes the sum of complementarity of the nth modality at time t, denotes the information entropy of the nth modality.

5. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The preset diagnosis logic model of the diagnosis module is a causal attention deep confidence network, which includes an input layer, a causal reasoning layer, an attention weighting layer and an output layer in sequence: The causal reasoning layer is used to mine the causal relationship between each feature and the device abnormality result, and the contribution relationship of the feature to the abnormality is quantified through causal strength, and the calculation formula of the causal strength is: ； wherein: represents the probability of an anomaly occurring when a feature is present and other features are excluded represents the probability of an anomaly occurring when a feature is not present and other features are retained​ The attention weighting layer assigns dynamic weights to the features based on the causal strength, with weight coefficients L is the number of features in the comprehensive feature set, and only for causal strength features with ≥ 0.2, weight assignments are retained; The abnormality degree quantitative evaluation result is represented by an abnormality confidence S, when S≥1.2, it is determined as a serious abnormality, when 0.8≤S<1.2, it is determined as a mild abnormality, and when S<0.8, it is determined as normal; ； In the formula: is a normalized value of the i-th feature; is a preset security feature threshold value.

6. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The time series feature deduction model of the early warning module is subject to: Short-term trend prediction value: ； The dynamic adjustment formula of the safety threshold is: ； In the formula: is the device state feature prediction value at future time t+k; is a bidirectional long short-term memory network; is a device state feature sequence from the current time t to the historical time t−w, and w represents the length of the sliding time window; is an environmental parameter vector at time t, at least including temperature and humidity; is a dynamic updated equipment safety threshold at time t; is an initial equipment safety threshold; is an aging coefficient, an environmental influence coefficient; is a cumulative running time of the equipment; is a reference running time, and is fixed at 1000 hours; is a deviation rate of the environmental parameter from the reference value at time t; Wherein, The value range is 0.02-0.06, 0.01-0.03 respectively, and the early warning triggering condition is And the continuous prediction is 3 periods.

7. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, When the analysis module extracts the frequency domain features of the mechanical vibration signals, it is based on variable resolution spectrum analysis: In the device resonance frequency interval, high resolution analysis is adopted, i.e. frequency resolution , the characteristic frequency is calculated as , represents the frequency point index, represents the signal sampling frequency, represents the sampling point number in high resolution analysis; In the non-resonance interval a low resolution analysis is applied, i.e. a frequency resolution , the characteristic frequency is calculated as , represents a quarter ; By spectral energy concentration Screening key frequency domain features; In the formula: is the total energy value in the device resonance frequency interval; is the total energy value in the full frequency interval; wherein the device resonance frequency interval is set to 500-2000 Hz, > 0.4 retains the resonance characteristic.

8. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The linkage logic of the response module is based on dynamic adjustment of device health degree: ； When the diagnosis abnormality level is a serious abnormality, and When the diagnosis abnormality level is a serious abnormality, and ; When the abnormality is slight, a stepwise adjustment command is generated, whose adjustment amount is ; Adjustment by health recovery rate Verify effect, in Adjustment upgrade when <0; In the formula, is the equipment health assessment value at time t; is the number of equipment loss types, including electrical loss, mechanical wear loss, and thermal aging loss; is the weight coefficient of the ith loss, and the greater the impact of the loss on the equipment life, the greater the value, and the weight coefficient is subject to ; is the cumulative amount of the ith loss at time t; is the load shedding ratio of the emergency load shedding instruction; is the step adjustment amount of the equipment operating parameter; is the health correlation coefficient; is the abnormal confidence; is the observation time interval after the adjustment instruction is executed; Wherein, According to the device type and parameter adjustment sensitivity setting, the power distribution transformer takes 5-10, the switch cabinet takes 3-6, and the higher the parameter adjustment sensitivity, the greater the value.

9. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The fusion module running stage corrects each modality feature vector Amendment: ； In the formula, is the feature vector of the mth modality after outlier correction; is the feature correction coefficient of the mth modality; is the mean of the feature vector of the mth modality; is the standard deviation of the feature vector of the mth modality; is a sign function, outputting 1 when the input value is positive, -1 when the input value is negative, and 0 when the input value is 0; wherein, According to The signal stability is adaptively valued in the interval of 0.8-1.2, the higher the signal stability is, the closer the value is to 1.2, and vice versa, the closer the value is to 0.8, and the signal stability is comprehensively represented by the characteristic fluctuation variance and the signal-to-noise ratio.

10. The power distribution equipment online monitoring system based on multi-modal data fusion of claim 1, wherein, The perception module is connected with an analysis module and a fusion module through a wireless network, the fusion module is connected with a diagnosis module and an early warning module through a wireless network, and the diagnosis module and the early warning module are connected with a response module through a wireless network.

Images