A power distribution cabinet intelligent monitoring method and system

Abstract

The present application relates to the field of power distribution cabinet monitoring, and more particularly to a power distribution cabinet intelligent monitoring method and system. The method comprises the following steps: first, acquiring an acoustic time sequence signal representing the operating state of the power distribution cabinet and generating a time-frequency energy graph; then, using a variational autoencoder model trained only with normal state data, performing anomaly detection on the time-frequency energy graph to identify an acoustic anomaly event to be verified; subsequently, synchronously acquiring electrical operating parameters near the time point of the event, and calculating an alarm confidence score based on the parameters; then, combining the severity of the acoustic anomaly event with the alarm confidence score to obtain a comprehensive risk score; finally, executing a graded warning or disposal based on the comprehensive risk score. This method can accurately identify power distribution cabinet anomalies and improve the reliability of early warnings, providing effective support for power distribution cabinet operation and maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power distribution cabinet monitoring, and in particular to a power distribution cabinet intelligent monitoring method and system. BACKGROUND

[0002] In order to realize real-time control and fault prediction of the state of the power distribution cabinet, a remote online monitoring system has become an industry standard configuration. At present, the mainstream monitoring scheme mainly relies on the acquisition terminal deployed in the cabinet to periodically collect and report macro electrical parameters such as voltage, current, and temperature. This method is effective for detecting traditional electrical faults such as overload, short circuit, and open phase.

[0003] However, this kind of traditional method has inherent and difficult-to-overcome defects when dealing with two kinds of physical faults with strong concealment and high destructive power. The first is internal arc fault, which occurs extremely short (often in milliseconds) and releases concentrated energy, and when it causes significant current or temperature changes, it often causes irreversible damage to the equipment and even causes fires. For critical application scenarios such as production lines or data centers, downtime caused by such faults will cause huge economic losses. The second is a loose key connection, which is a gradual process, and only shows a slight increase in contact resistance and a small local temperature rise at the beginning. Traditional temperature sensors are difficult to effectively cover and capture this local anomaly due to limited measurement points, and when the temperature rises significantly, the fault has developed to a high-risk stage. Therefore, relying solely on electrical quantity monitoring makes the system's response to these high-risk faults severely lagging, and cannot play a real early warning role.

[0004] Therefore, finding a characteristic signal that can sensitively capture early signs of arc, loose, and other physical faults, and developing an intelligent algorithm that can accurately identify these signals from complex environmental interference, to achieve early and reliable monitoring of physical faults with strong concealment and high destructive power, has become the focus of this scheme. SUMMARY

[0005] In order to solve the problem of how to realize early monitoring of physical faults with strong concealment and high destructive power, the present application provides a power distribution cabinet intelligent monitoring method and system.

[0006] In a first aspect, the present application provides a power distribution cabinet intelligent monitoring method, which adopts the following technical scheme:

[0007] A power distribution cabinet intelligent monitoring method, comprising the steps of:

[0008] a. Acquire an acoustic time sequence signal representing the running state of the power distribution cabinet, and generate a time-frequency energy graph;

[0009] b. performing anomaly detection on the time-frequency energy map based on a variational autoencoder model trained only using normal state data to identify an acoustic abnormal event to be verified;

[0010] c. when the acoustic abnormal event to be verified is identified, synchronously acquiring electrical operation parameters near the event time point, and calculating an alarm credibility score for representing credibility of the acoustic abnormal event based on the electrical operation parameters;

[0011] d. combining the severity of the acoustic abnormal event with the alarm credibility score to calculate a comprehensive risk score;

[0012] e. performing hierarchical early warning or disposal based on the comprehensive risk score.

[0013] The present application firstly realizes early warning by using high-sensitivity acoustic anomaly detection, then cross-verification is introduced by using electrical parameters to ensure the reliability of the alarm, and finally an intelligent monitoring system with high sensitivity and high reliability is formed, which significantly improves the safety of the power distribution cabinet operation.

[0014] Preferably, the method for generating the time-frequency energy map comprises:

[0015] The acoustic time series signal is processed by using continuous wavelet transform.

[0016] By using continuous wavelet transform, the present application enhances the feature extraction capability of the key fault signal.

[0017] Preferably, the method for identifying the acoustic abnormal event to be verified comprises:

[0018] The time-frequency energy map generated in real time is input into the variational autoencoder model to obtain a reconstructed map, and a reconstruction error between the time-frequency energy map and the reconstructed map is calculated.

[0019] A health state index is calculated based on the reconstruction error, and when the health state index is lower than a preset reference threshold, the acoustic abnormal event to be verified is identified.

[0020] In the present application, the variational autoencoder model can sensitively capture any deviation from the normal state by learning the distribution of the normal state, and quantify it as a reconstruction error. This unsupervised anomaly detection mechanism enables the system to capture the deviation from the normal state at a stage where the fault feature is not obvious and far from being recognized by the classifier, and realizes the promotion from fault diagnosis to state prediction.

[0021] Preferably, the method for obtaining the alarm credibility score comprises:

[0022] An electrical disturbance index is calculated based on the electrical operation parameters.

[0023] The electrical disturbance index is mapped by a Sigmoid function to obtain the alarm credibility score.

[0024] Preferably, the electrical disturbance index is calculated according to the weighted sum of the current standard deviation and the temperature change rate in the electrical operation parameters.

[0025] The present application considers that the current standard deviation can effectively reflect electrical faults such as arc and short circuit, and the temperature change rate can reflect thermal faults such as overheating of the connection point. By weighted fusion of the two core indexes, the disturbance degree of the electrical system can be more comprehensively and accurately evaluated.

[0026] Preferably, the method for obtaining the comprehensive risk score is:

[0027] The acoustic anomaly severity derived from the health state index is multiplied by the alarm credibility score.

[0028] The present application can extremely effectively suppress acoustic false positives caused by environmental noise, sensor failure, etc. by combining acoustic detection results with electrical detection results to determine the comprehensive risk score, ensuring that each high-risk alarm has extremely high confidence.

[0029] Preferably, it further comprises:

[0030] When the comprehensive risk score is higher than a preset classification starting threshold, the time-frequency energy map leading to the high comprehensive risk score is input into a pre-trained convolutional neural network classifier to obtain a specific fault classification result.

[0031] Preferably, the execution of hierarchical early warning or disposal is simultaneously based on the comprehensive risk score and the fault classification result.

[0032] Preferably, the acquisition frequency range of the acoustic time sequence signal covers audible and ultrasonic frequency bands.

[0033] The present application has the following technical effects:

[0034] The present application can identify early weak signals and unknown faults deviating from the normal state by introducing unsupervised anomaly detection based on variational autoencoder, realizing predictive maintenance of potential risks, far superior to the traditional post-diagnosis mode relying on fault sample library, and greatly improving the safety of the operation of the power distribution cabinet.

[0035] Further, an acoustic early warning and electrical verification multi-modal fusion architecture is designed, which not only utilizes the high sensitivity of acoustic monitoring, but also effectively suppresses false positives caused by environmental noise and other factors through cross-validation of electrical parameters, thereby ensuring extremely high detection rate while minimizing false alarm rate.

[0036] Further, the verification-first and classification-later process realizes intelligent allocation of computing resources, using the highest-precision diagnostic model for the most critical link. At the same time, the grading early warning and disposal mechanism based on comprehensive risk score and specific fault type makes the final response measures more targeted, more accurate and more reliable, avoiding excessive or insufficient response due to insufficient information. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a flowchart of a power distribution cabinet intelligent monitoring method provided by the embodiment 1 of the present application;

[0038] Figure 2 is a fault early warning capability comparison graph provided by the embodiment 2 of the present application. DETAILED DESCRIPTION

[0039] Embodiment 1:

[0040] The embodiment 1 of the present application discloses a power distribution cabinet intelligent monitoring method, referring to Figure 1 , comprising steps S1-S5:

[0041] S1: Acquire acoustic time series signals representing the running state of the power distribution cabinet, and generate a time-frequency energy graph.

[0042] It should be noted that traditional electrical quantity monitoring is slow to respond to mechanical or high-frequency electrical faults such as arc and loose connection. These faults are accompanied by unique acoustic characteristics during occurrence or development. Therefore, it is necessary to collect acoustic signals as a new monitoring dimension to solve the problem that traditional monitoring methods cannot effectively capture such faults.

[0043] It should be further noted that one-dimensional sound signals cannot intuitively reflect fault characteristics, and therefore need to be converted into two-dimensional images.

[0044] Preferably, as an example, high-sensitivity acoustic sensors are installed at multiple key positions inside the power distribution cabinet. The sensors continuously collect sound data to form multi-channel acoustic time series signals. Then, for each channel of the collected acoustic time series signals, continuous wavelet transform (CWT) is used for processing to convert it into a two-dimensional time-frequency energy graph. The horizontal axis of the graph is time, the vertical axis is frequency, and the brightness or color of each point in the graph represents the energy intensity at that time-frequency point. Exemplarily, the key positions can be bus connection points and circuit breaker vicinity, and the acoustic sensors can be microphones.

[0045] It should be noted that, in order to effectively capture various fault signals including partial discharge, the frequency response range of the sensor preferably covers 20Hz to 100kHz to include audible and ultrasonic frequency bands, and its signal-to-noise ratio (SNR) is not less than 60dB to ensure that clear signals are collected in noisy industrial environments.

[0046] It can be understood that, by introducing acoustic monitoring, a new data source more sensitive to mechanical and high-frequency electrical faults is provided for the system. More importantly, by converting a one-dimensional time series signal into a two-dimensional time-frequency energy graph through CWT, not only the frequency components of the signal are preserved, but also the information of the change of these frequency components over time is preserved, providing high-quality feature input for subsequent accurate identification of non-stationary and transient fault signals.

[0047] S2: Based on a variational autoencoder model trained only using normal state data, the time-frequency energy graph is subjected to anomaly detection to identify the acoustic anomaly event to be verified.

[0048] It should be noted that faults, especially progressive faults such as loose screws, have weak acoustic characteristics in the early stage and may never have appeared in historical data. Relying on a supervised learning model for supervised fault identification cannot detect these unknown early anomalies. Therefore, it is necessary to introduce an unsupervised learning method to identify abnormal events deviating from the normal by learning the characteristics of the normal state.

[0049] It should be further noted that the model constructed using normal samples is more representative of normal samples, and its representation of abnormal samples is poor, so the reconstruction error of the model can be used to determine the possibility of the signal input into the model being abnormal.

[0050] Preferably, as an example, based on a variational autoencoder model trained only using normal state data, the time-frequency energy graph is subjected to anomaly detection to identify the acoustic anomaly event to be verified, comprising:

[0051] First, a large number of time-frequency energy graphs generated by the power distribution cabinet when running under various normal loads and environments are collected to form a normal state data set, and the training of the variational autoencoder (VAE) model is completed using the normal state data set.

[0052] Then, the time-frequency energy graph collected at the time to be judged is input into the trained VAE, which will encode and compress the input graph and then decode it to obtain the corresponding reconstruction graph.

[0053] After that, the health status index is calculated according to the reconstruction error of the reconstruction graph.

[0054]

[0055] wherein, represents the health state indicator at the time to be determined, is a normalized value of the reconstruction error of the reconstructed image at the time to be determined, λ is a first sensitivity coefficient, used to adjust the degree of influence of the reconstruction error on the final health state indicator, and e represents a natural constant.

[0056] It can be understood that the VAE has mastered the internal law and high-dimensional distribution of the acoustic mode of the normal state by learning a large number of normal samples. When a tiny, early fault signal appears, the corresponding time-frequency energy image deviates from the known normal distribution, and the VAE cannot effectively reconstruct it using the learned knowledge, resulting in a significant increase in reconstruction error, thereby identifying an abnormal event. The greater the health state indicator value, the smaller the reconstruction error, and the less likely the abnormality at the time to be determined.

[0057] It should be noted that the mean of the absolute value of the difference between the pixel values of all corresponding pixels of the reconstructed image and the time-frequency energy image can be used as the reconstruction error.

[0058] Finally, if the health state indicator is less than a preset reference threshold, it is determined that there is an acoustic abnormal event to be verified at the time to be determined. If the health state indicator is not less than the preset reference threshold, it is determined that the operation is normal at the time to be determined, and no subsequent analysis is required.

[0059] S3: When the acoustic abnormal event to be verified is identified, the electrical operation parameters near the event time point are synchronously acquired, and an alarm credibility score representing the credibility of the acoustic abnormal event is calculated based on the electrical operation parameters.

[0060] It should be noted that a high-sensitivity acoustic detection model may produce false positives due to changes in environmental noise or sensor faults, so it is necessary to cross-verify with other key physical quantities to improve the credibility of the alarm.

[0061] It should be further noted that most acoustic abnormalities in the power distribution cabinet will inevitably be accompanied by synchronous abnormal fluctuations in their electrical parameters, so it is necessary to judge the abnormality of the electrical parameters when an acoustic abnormal event occurs, further verify the credibility of the abnormality, and reduce the occurrence of false judgments.

[0062] Preferably, as an example, when the acoustic abnormal event to be verified is identified, the electrical operation parameters near the event time point are synchronously acquired, and an alarm credibility score representing the credibility of the acoustic abnormal event is calculated based on the electrical operation parameters, comprising:

[0063] If there is an acoustic anomaly to be verified at the time to be determined, a preset number of times are obtained before and after the time to be determined as analysis times, and electrical operation data at the analysis times are obtained. For example, the types of electrical operation data include, but are not limited to, the following: current and temperature.

[0064] Calculate the electrical disturbance index:

[0065]

[0066] in, The weighting parameter represents the current disturbance index. The weighting parameter represents the temperature disturbance index. This represents the standard deviation of the current at all analysis times. This represents the mean rate of temperature change at the time point being analyzed. This represents the normalization method; for example, the normalization method could be the Z-score standardization method. This indicates the electrical disturbance index.

[0067] Calculate the alert confidence score:

[0068]

[0069] The Sigmoid function formula is used here. Normalize, This represents the steepness coefficient in the Sigmoid function formula, used to control the steepness of the Sigmoid function curve. This represents the center standard coefficient, used to adjust the position of the curve's inflection point. This indicates the confidence score of the alert.

[0070] It is understandable that the larger the current standard deviation and the greater the temperature change rate, the greater the overall electrical fluctuation, and therefore the greater the credibility of the abnormal events obtained based on acoustic information.

[0071] S4: Calculate a comprehensive risk score by combining the severity of the acoustic anomaly with the confidence score of the alarm.

[0072] It should be noted that, in order to achieve accurate anomaly warning, the indicators obtained from acoustics and electrical measurements need to be combined to determine the abnormal situation.

[0073] Preferably, as an example, a comprehensive risk score is calculated by combining the severity of the acoustic anomaly with the alarm confidence score, including:

[0074]

[0075] in, represents the severity of the acoustic anomaly, represents the comprehensive risk score.

[0076] It can be understood that, reflects the health of the operating state obtained by the acoustic information, and the larger the value is, the less likely the operating state is abnormal, reflects the abnormality of the operating state obtained by the acoustic information, reflects the operating state obtained by the electrical information, and the operating state obtained by the acoustic information and the electrical information is combined to determine the final operating state.

[0077] S5: based on the comprehensive risk score, performing hierarchical warning or disposal.

[0078] Optionally, as an example, based on the comprehensive risk score, performing hierarchical warning or disposal, including:

[0079] Attention level: Record "verified minor anomaly" in the background, and prompt the operation and maintenance personnel to pay attention.

[0080] Warning level: Send "high-confidence warning" to the operation and maintenance platform, and suggest arranging maintenance.

[0081] Alarm level: The system immediately issues the highest level of audible and visual alarm.

[0082] Preferably, as an example, based on the comprehensive risk score, performing hierarchical warning or disposal, including:

[0083] The risk classification method and the warning method are the same as the optional method, and the only difference is that if the comprehensive risk score is higher than the preset classification starting threshold, a pre-trained convolutional neural network (CNN) classifier is called to classify the fault of the time-frequency energy graph causing the event, and the fault classification result is attached to the alarm information. In particular, when the fault classification result is "arc fault", no matter the level, the system will raise the alarm level to the highest "alarm level", and link the protection device to perform emergency power-off disposal.

[0084] Embodiment 1 of the present application also discloses an intelligent monitoring system for power distribution cabinet, comprising a processor and a memory, and the memory stores computer program instructions, which realize the intelligent monitoring method for power distribution cabinet according to the present application when executed by the processor.

[0085] The above system also includes a communication bus and a communication interface and other components well known to those skilled in the art, and their settings and functions are known in the art, so they will not be described here.

[0086] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as resistive random access memory, dynamic random access memory, static random access memory, enhanced dynamic random access memory, high bandwidth memory, hybrid memory cube, or the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto.

[0087] Embodiment 2

[0088] This embodiment is a specific embodiment of a power distribution cabinet intelligent monitoring method.

[0089] In order to illustrate the technical effect in embodiment 1 of the application, the technical effect of the technical scheme will be illustrated below in combination with some experimental process data.

[0090] Experimental goal: verify the early warning capability of the algorithm.

[0091] Figure 2 The early warning capability comparison chart of the fault is shown, the orange dotted curve in the chart is the abnormal level curve of the traditional method, and the blue solid curve is the comprehensive risk score curve of the method in embodiment 1. It can be seen from the image that the method in embodiment 1 can identify risk information earlier, thereby giving a risk prompt as soon as possible.

[0092] In summary, compared with the traditional method, the method in embodiment 1 can identify the dangerous situation of the power distribution cabinet earlier, thereby reserving more time to respond, and thus effectively reducing the harm of the risk.

Claims

1. A power distribution cabinet intelligent monitoring method, characterized in that, The method comprises the steps of: a. acquiring an acoustic time series signal representing the running state of the power distribution cabinet and generating a time-frequency energy map; b. based on a variational autoencoder model trained only using normal state data; The time-frequency energy graph is subjected to anomaly detection to identify an acoustic anomaly event to be verified, including: inputting the real-time generated time-frequency energy graph into a variational autoencoder model to obtain a reconstructed graph; calculating the mean value of the absolute value of the pixel value difference of all corresponding pixels of the reconstructed graph and the time-frequency energy graph as a reconstruction error; calculating a health status indicator based on the reconstruction error, satisfying: , is the health status indicator of the to-be-judged moment, is the normalized value of the reconstruction error of the reconstructed graph of the to-be-judged moment, is a first sensitivity coefficient, is a natural constant; when the health status indicator is lower than a preset reference threshold, an acoustic anomaly event to be verified is identified; c. When an acoustic abnormal event to be verified is identified, electrical operation parameters near the event time point are synchronously acquired; an alarm credibility score for representing credibility of the acoustic abnormal event is calculated based on the electrical operation parameters, including: , is the alarm credibility score, is a steepness coefficient in a Sigmoid function, is an electrical disturbance index, is a central standard coefficient; the electrical disturbance index is calculated according to a weighted sum of a current standard deviation and a temperature change rate in the electrical operation parameters; d. combining the severity of the acoustic anomaly event with the alert confidence score to calculate a composite risk score, including: , is the composite risk score, is the severity of the acoustic anomaly event; e. based on the comprehensive risk score, performing a hierarchical early warning or disposal.

2. The intelligent monitoring method for power distribution cabinet according to claim 1, characterized in that, The method for generating a time-frequency energy map is to process the acoustic time series signal using continuous wavelet transform.

3. The intelligent monitoring method of a power distribution cabinet according to claim 1, characterized in that, Further comprising: When the comprehensive risk score is higher than a preset classification starting threshold, input the time-frequency energy map leading to the high comprehensive risk score into a pre-trained convolutional neural network classifier to obtain a specific fault classification result.

4. The intelligent monitoring method for power distribution cabinet according to claim 3, characterized in that, The performance of hierarchical early warning or disposal is simultaneously based on the comprehensive risk score and the fault classification result.

5. The intelligent monitoring method of a power distribution cabinet according to claim 1, characterized in that, The acquisition frequency range of the acoustic time series signal covers audible and ultrasonic frequency bands.

6. A power distribution cabinet intelligent monitoring system, characterized in that, It comprises: a processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement a power distribution cabinet intelligent monitoring method according to any one of claims 1-5.

Images