Intelligent detection system for safe electric device

Abstract

The invention relates to the field of intelligent detection, and particularly discloses an intelligent detection system for a safe electrical device, which adopts a neural network model based on deep learning to process real-time operation state parameters of monitored electrical equipment and various state parameters marked as fault states. According to the method, the time domain correlation characteristics of the equipment state multi-modal parameters of the electrical equipment in the real-time operation state and various fault states are extracted respectively, and the real-time state parameter characteristics and the fault state parameter characteristics of the electrical equipment are further inquired, matched and analyzed, so that whether the state of the monitored electrical equipment is abnormal or not is intelligently identified.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection, and more specifically, to an intelligent detection system for safety electrical devices. Background Art

[0002] With the continuous development of power systems, the safe and stable operation of electrical equipment has become one of the key factors to ensure the normal order of social production and life. Existing electrical equipment status monitoring technology mainly relies on traditional threshold comparison method or simple statistical analysis method. Although it can achieve fault warning to a certain extent, it has many limitations.

[0003] Traditional methods often judge the status of equipment based on changes in a single parameter (such as current, voltage or temperature), lacking in-depth exploration of the complex correlation between multiple parameters, resulting in low accuracy and sensitivity of fault detection. The threshold comparison method usually requires pre-setting, but the normal operating range of electrical equipment in actual operation is not fixed due to factors such as environmental changes and load fluctuations. Static thresholds are difficult to adapt to dynamically changing working conditions and are prone to false alarms or missed alarms, affecting the sensitivity and reliability of fault detection.

[0004] Therefore, an intelligent detection system for safe electrical devices is desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an intelligent detection system for safe electrical devices, which respectively processes the real-time operating status parameters of the monitored electrical equipment and various status parameters marked as fault states by using a neural network model based on deep learning, so as to respectively extract the time-domain correlation characteristics of the multimodal parameters of the equipment status under the real-time operating state and various fault states of the electrical equipment, and further query and match the characteristics of the real-time status parameters and the characteristics of the fault status parameters of the electrical equipment, so as to intelligently identify whether there is an abnormality in the status of the monitored electrical equipment. In this way, by comparing and analyzing the real-time data with the historical fault data, it is possible to quickly and accurately determine whether the equipment has potential faults, thereby providing timely warnings to ensure the stable and safe operation of the power system.

[0006] According to one aspect of the present application, there is provided an intelligent detection system for a safety electrical device, comprising: A data acquisition module, used to obtain a data set of state parameters of the monitored electrical equipment collected by the sensor group; A data reconstruction module, used for structuring the data set of the state parameters of the monitored electrical equipment to obtain a real-time state parameter full-time domain full-sample space aggregation matrix; A reference sample extraction module, used to extract a set of full-time domain full-sample spatial aggregate reference matrices of state parameters marked as fault states from a database; A feature extraction module, used to extract a set of real-time state parameter multi-modal time series associated implicit feature vectors and a set of reference state parameter multi-modal time series associated implicit feature vectors from the set of the real-time state parameter full-time domain full-sample space aggregate matrix and the state parameter full-time domain full-sample space aggregate reference matrix marked as a fault state; A query encoding module, configured to perform semantic essential expression difference enhancement modulation on the set of implicit feature vectors associated with the multimodal time series of the reference state parameters, and then perform cross-domain query encoding with the implicit feature vectors associated with the multimodal time series of the real-time state parameters to obtain a real-time state optimization query response representation vector; The anomaly detection module is used to determine whether there is an anomaly in the state of the monitored electrical equipment based on the real-time state optimization query response representation vector.

[0007] Compared with the prior art, the intelligent detection system for safe electrical devices provided by the present application processes the real-time operating state parameters of the monitored electrical equipment and various state parameters marked as fault states by using a neural network model based on deep learning, so as to extract the time domain correlation characteristics of the multimodal parameters of the equipment state under the real-time operating state and various fault states of the electrical equipment, and further query and match the characteristics of the real-time state parameters and the fault state parameters of the electrical equipment, so as to intelligently identify whether the state of the monitored electrical equipment is abnormal. In this way, by comparing and analyzing the real-time data with the historical fault data, it is possible to quickly and accurately determine whether the equipment has potential faults, thereby providing timely warnings to ensure the stable and safe operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 is a block diagram of an intelligent detection system for a safety electrical device according to an embodiment of the present application; Figure 2 A schematic diagram of data flow of an intelligent detection system for a safety electrical device according to an embodiment of the present application; Figure 3 The present invention is a block diagram of a query coding module in an intelligent detection system for a safety electrical device according to an embodiment of the present application. DETAILED DESCRIPTION

[0010] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0011] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0012] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0013] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0014] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0015] In the technical solution of the present application, an intelligent detection system for safe electrical devices is proposed. Figure 1 4 is a block diagram of an intelligent detection system for safe electrical devices according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of an intelligent detection system for a safety electrical device according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to an embodiment of the present application, the intelligent detection system 300 for safety electrical devices includes: a data acquisition module 310, which is used to obtain a data set of state parameters of the monitored electrical equipment collected by a sensor group; a data reconstruction module 320, which is used to structure the data set of state parameters of the monitored electrical equipment to obtain a real-time state parameter full-time domain full-sample space aggregation matrix; a reference sample extraction module 330, which is used to extract a set of full-time domain full-sample space aggregation reference matrices of state parameters marked as fault states from a database; a feature extraction module 340, which is used to extract a set of full-time domain full-sample space aggregation reference matrices of state parameters marked as fault states from the real-time state parameter full-time domain full-sample space aggregation matrix and the marked A set of real-time state parameter multimodal time series associated implicit feature vectors and a set of reference state parameter multimodal time series associated implicit feature vectors are extracted from a set of full-time domain and full-sample space aggregated reference matrices of state parameters of a fault state; a query encoding module 350 is used to perform cross-domain query encoding with the real-time state parameter multimodal time series associated implicit feature vectors after performing semantic essential expression difference enhancement modulation on the set of reference state parameter multimodal time series associated implicit feature vectors to obtain a real-time state optimized query response representation vector; an anomaly detection module 360 ​​is used to determine whether there is an abnormality in the state of the monitored electrical equipment based on the real-time state optimized query response representation vector.

[0016] In particular, the data acquisition module 310 is used to obtain a data set of state parameters of the monitored electrical equipment collected by the sensor group. Among them, the state parameters include current value, voltage value and temperature value. It should be understood that the key state parameters during the operation of the equipment can reflect the operation of the equipment within a predetermined time period. Specifically, the current value directly reflects the workload of the electrical equipment. Too high a current may cause the equipment to be overloaded, overheated or even damaged, while too low a current may indicate that there is a circuit break or other faults in the circuit. The voltage value is a key parameter for measuring the power supply quality of electrical equipment. Too high or too low a voltage may cause damage to the equipment. For example, too low a voltage may cause the equipment to fail to start normally or reduce working efficiency; too high a voltage may burn circuit components. In addition, the temperature value reflects the thermal state inside the electrical equipment. Since the equipment generates heat during operation, poor heat dissipation may cause the equipment temperature to rise, thereby affecting the performance and safety of the equipment. Therefore, the current value, voltage value and temperature value are the three core parameters for evaluating the operating status of electrical equipment. By comprehensively analyzing these parameters, the operating status of the equipment can be fully understood, and hidden faults can be discovered and eliminated in a timely manner to ensure the safe and stable operation of the power system.

[0017] In order to comprehensively monitor the operating status of electrical equipment, it is first necessary to select appropriate sensors to collect key status parameters. Common sensor types include current sensors, voltage sensors, and temperature sensors. Current sensors are used to monitor the current value of electrical equipment. Common ones include Hall effect sensors and current transformers. Hall effect sensors measure current by detecting changes in magnetic fields and are suitable for DC and AC circuits; current transformers measure current by the principle of electromagnetic induction and are suitable for high-voltage and high-current circuits. Voltage sensors are used to monitor the voltage value of electrical equipment. Common ones include voltage transformers and resistor dividers. Voltage transformers measure voltage by the principle of electromagnetic induction and are suitable for high-voltage circuits; resistor dividers measure voltage by the principle of resistor divider and are suitable for low-voltage circuits. Temperature sensors are used to monitor the temperature value of electrical equipment. Common ones include thermocouples, thermistors, and infrared temperature sensors. Thermocouples measure temperature by measuring the electromotive force generated by temperature difference and are suitable for high-temperature environments; thermistors measure temperature by measuring the change in resistance with temperature and are suitable for medium and low-temperature environments; infrared temperature sensors measure temperature by detecting the infrared energy radiated by objects and are suitable for non-contact measurement.

[0018] The arrangement of sensors needs to select appropriate installation locations and quantities according to the structure and operating characteristics of electrical equipment. Reasonable sensor arrangement can improve the accuracy and reliability of monitoring. In terms of location selection, current sensors can be installed in the main circuit, voltage sensors can be installed at the power input end, and temperature sensors can be installed near key components. For example, for transformers, temperature sensors can be installed at key locations such as windings, cores, and oil tanks; for circuit breakers, temperature sensors can be installed at key locations such as contacts and arc extinguishing chambers. In terms of quantity configuration, the number of sensors should be reasonably configured according to monitoring requirements. For important key components, the number of sensors can be increased to improve monitoring accuracy and reliability. For example, for high-voltage switchgear, multiple current sensors and temperature sensors can be installed at different phases and different locations to comprehensively monitor the operating status of the equipment.

[0019] The data acquisition system is the entrance for sensor data to enter the system. It is responsible for converting sensor signals into digital signals and performing preliminary processing. The data acquisition module includes hardware and software parts. The hardware part includes sensor interface, signal conditioning circuit and analog-to-digital converter (ADC). The sensor interface is used to connect various sensors and provide standard interfaces (such as analog input, digital input, pulse input, etc.). The signal conditioning circuit is used to amplify and filter sensor signals to ensure the stability and reliability of the signal. Common signal conditioning circuits include amplifiers, filters and voltage regulators. The analog-to-digital converter (ADC) is used to convert analog signals into digital signals. Common ADCs include successive approximation type, Σ-Δ type and pipeline type. The software part is to write data acquisition programs to realize the reading, storage and transmission of sensor data. Common programming languages ​​are C / C++ and Python. The data acquisition program needs to have the functions of data reading, data storage and data transmission, and can read sensor data regularly or in real time, and store the data in local storage devices (such as SD cards, hard disks), or transmit the data to the central processing unit or cloud server.

[0020] Data transmission is the process of transmitting collected data from sensor nodes to a central processing unit or cloud server. Common data transmission methods include wired transmission and wireless transmission. Wired transmission transmits data to a central processing unit through a serial communication interface (such as UART, SPI, I2C) or an Ethernet interface. The advantages of wired transmission are fast transmission speed and high reliability, but the wiring is complex and the cost is high.

[0021] Data cleaning is to remove noise and outliers from sensor data to ensure the quality and reliability of the data. Methods for removing noise include using filtering algorithms such as low-pass filtering, high-pass filtering, and band-pass filtering. Low-pass filtering is used to remove high-frequency noise, high-pass filtering is used to remove low-frequency noise, and band-pass filtering is used to retain signals within a specific frequency range. Methods for handling outliers include threshold methods and statistical methods. The threshold method sets a reasonable threshold and treats data that exceeds the threshold as outliers and removes them. The statistical method detects and removes outliers by using statistical methods such as mean, median, and standard deviation. For example, the 3σ principle can be used to treat data that exceeds the mean ±3 times the standard deviation as outliers and remove them.

[0022] Data synchronization ensures that the data collected by different sensors have the same timestamp to facilitate subsequent time series analysis. Timestamp synchronization ensures the time consistency of all sensor nodes by using a synchronized clock (such as NTP, PTP) and adding a timestamp to each data during data collection to ensure the time consistency of the data. Data alignment ensures the consistency and integrity of the data by aligning the data collected by different sensors to the same timeline. For example, interpolation methods can be used to align data with different sampling frequencies to the same sampling frequency.

[0023] Data storage is to save the collected data to a local storage device or cloud server for subsequent data processing and analysis. Local storage uses local storage devices such as SD cards and hard disks to save data and convert the data into standardized formats (such as CSV, JSON, XML) for subsequent data processing and analysis. Cloud storage transmits data to a cloud server via wired or wireless transmission, and performs data management on the cloud server, including data storage, data backup, and data query.

[0024] Data formatting is the process of converting collected data into a standardized format to facilitate subsequent data processing and analysis. Standardized formats include CSV (comma-separated value format, suitable for tabular data), JSON (JavaScript object notation, suitable for structured data), and XML (eXtensible Markup Language, suitable for complex structured data). Data annotation includes recording the source of the data (such as sensor ID, device number, etc.), time (such as timestamp), and type (such as current value, voltage value, temperature value, etc.).

[0025] Data backup is an important measure to ensure data security and prevent data loss or damage. Regular backup should be based on the importance and frequency of data changes, and a reasonable backup cycle (such as daily, weekly, and monthly) should be set. Physical media such as tapes, optical disks, and external hard disks should be used for backup, or cloud storage services should be used for backup. Version management ensures data traceability and recovery by recording the time and content of each data update, and uses version control systems (such as Git) to manage data versions to ensure data integrity and consistency.

[0026] Data encryption is an important means of protecting data security and preventing data from being illegally accessed and tampered with during transmission and storage. Transmission encryption ensures the security of data during transmission by using encryption protocols such as SSL / TLS, and encrypts the transmitted data using encryption algorithms such as AES and RSA. Storage encryption encrypts stored data by using file encryption software (such as VeraCrypt) and database encryption functions (such as MySQL's TDE).

[0027] Access control is an important measure to ensure that only authorized users can access and operate data. User authentication is performed by using user names and passwords, digital certificates, biometrics, etc., and different data access rights are assigned according to the user's role and responsibilities. Audit logs record data access and operation logs, including access time, access users, access content, etc., and regularly review log records to discover and handle abnormal access behaviors.

[0028] In summary, obtaining a data set of state parameters of monitored electrical equipment collected by a sensor group is a systematic process, involving multiple links such as sensor selection and layout, data acquisition system design, data preprocessing, data management and data security. Through reasonable sensor configuration, efficient data acquisition and transmission, strict data management and security assurance, the collected data can be ensured to be accurate, complete and reliable, providing a solid foundation for subsequent fault detection and state assessment. The specific implementation method includes selecting suitable sensors and reasonably arranging them at key locations, designing efficient data acquisition modules and selecting appropriate data transmission methods, ensuring data quality and integrity through data cleaning, data synchronization and data storage, converting data into a standardized format and performing data annotation and backup, ensuring data management and traceability, and ensuring data security and privacy through data encryption and access control. Through the above steps, efficient collection and management of electrical equipment state parameters can be achieved, providing high-quality data support for subsequent processing and analysis of the intelligent detection system.

[0029] In particular, the data reconstruction module 320 is used to structure the data set of the state parameters of the monitored electrical equipment to obtain a real-time state parameter full-time domain full-sample space aggregation matrix. It should be understood that during the operation of the electrical equipment, there are complex correlation interactions between different state parameters. For example, fluctuations in current and voltage may affect the temperature change of the equipment. Therefore, in order to better understand the interaction between the various state parameters, in the technical solution of the present application, the data set of the state parameters of the monitored electrical equipment is structured to obtain a real-time state parameter full-time domain full-sample space aggregation matrix; by arranging the data set of the state parameters into a matrix form according to the time dimension and the parameter sample dimension, a time series correlation relationship between different parameters is established to facilitate subsequent comprehensive time series correlation analysis.

[0030] Data structuring refers to converting the collected raw data into a structured form to facilitate subsequent data processing and analysis. Data format conversion is to convert raw data into a standardized format, such as CSV, JSON (JavaScript Object Notation, suitable for structured data) or XML (Extensible Markup Language, suitable for complex structured data). These formats facilitate subsequent data processing and analysis. Data annotation is to add necessary annotation information to each piece of data, such as sensor ID, device number, collection time, etc., to facilitate data management and query.

[0031] The construction of data matrix is ​​to arrange the data in chronological order to form the time dimension, which reflects the trend of data changes over time. Different types of parameters (such as current value, voltage value, temperature value) are used as different feature columns to form a multi-parameter data matrix, and the parameter dimension reflects the correlation between different parameters. The data at each time point is taken as a sample to form the sample dimension, which reflects the distribution of data at different time points.

[0032] The real-time state parameter full-time domain full-sample spatial aggregation matrix is ​​the final form of the structured data. Time window division is to select the appropriate time window size according to the monitoring needs. For example, 1 minute, 5 minutes or 10 minutes can be selected as the time window. The time window is slid along the time axis to form multiple time windows, and the data in each time window is regarded as an independent sample.

[0033] Feature extraction is to extract the statistical features of data in each time window, such as mean, variance, maximum value, minimum value, etc. Extract the time series features of data, such as autocorrelation coefficient, partial autocorrelation coefficient, Fourier transform coefficient, etc. Extract the correlation features between different parameters, such as mutual correlation coefficient, covariance matrix, etc.

[0034] Matrix construction is to arrange the data in each time window in chronological order to form the time dimension. Different types of parameters are used as different feature columns to form the parameter dimension. Each time window is used as a sample to form the sample dimension. The data of the time dimension, parameter dimension and sample dimension are combined to form a real-time state parameter full-time domain full-sample spatial aggregation matrix.

[0035] Through the above steps, the collected raw data can be converted into a structured form, providing high-quality data support for subsequent feature extraction, fault detection and status assessment.

[0036] In particular, the reference sample extraction module 330 is used to extract a set of state parameter full-time domain full-sample space aggregated reference matrices marked as fault states from the database. The state parameter full-time domain full-sample space aggregated reference matrix is ​​a data sample formed by the above-mentioned structured processing of the labeled and verified real fault data, which provides a benchmark for known fault modes and can provide an important reference for fault detection of monitored electrical equipment. Therefore, in the technical solution of the present application, it is expected to discover potential fault signs by comparing the parameter characteristics of the real-time state and the known fault state, so as to improve the sensitivity of fault detection.

[0037] In particular, the feature extraction module 340 is used to extract a set of real-time state parameter multimodal temporal correlation implicit feature vectors and reference state parameter multimodal temporal correlation implicit feature vectors from the set of the real-time state parameter full-time domain full-sample space aggregation matrix and the state parameter full-time domain full-sample space aggregation reference matrix marked as a fault state. Since the operating state of the electrical equipment involves multiple parameters (such as current, voltage, temperature, etc.), there are complex time domain correlation patterns between these parameters. By extracting features from the operating state parameters, the correlation relationship between these multimodal parameters can be captured, thereby more comprehensively reflecting the operating state of the equipment. Considering that the traditional convolutional neural network is limited by its structural characteristics, its computational cost will increase significantly as the network depth increases. The hole convolution can expand the receptive field without increasing the number of parameters by inserting holes (i.e., zero filling) in the convolution kernel, thereby capturing the parameter association characteristics on a long time scale. Therefore, in the technical solution of the present application, a device state multimodal parameter time domain correlation pattern feature extractor based on a hole convolutional neural network model is used to perform hole convolution coding on each state parameter full time domain full sample space aggregation reference matrix in the set of the real-time state parameter full time domain full sample space aggregation matrix and the state parameter full time domain full sample space aggregation reference matrix marked as a fault state, so as to mine the time series correlation implicit features of the multimodal parameters at different time points, and obtain a set of real-time state parameter multimodal time series correlation implicit feature vectors and reference state parameter multimodal time series correlation implicit feature vectors. Among them, the real-time state parameter multimodal time series correlation implicit feature vector contains the operating status of the monitored device and its change trend, which can comprehensively reflect the operating status of the device.

[0038] In particular, the query encoding module 350 is used to perform semantic essential expression difference enhancement modulation on the set of implicit feature vectors associated with the multimodal time series of the reference state parameters, and then perform cross-domain query encoding with the implicit feature vectors associated with the multimodal time series of the real-time state parameters to obtain a real-time state optimized query response representation vector. Since different equipment failure modes usually have different parameter distribution characteristics, the above-mentioned feature encoding processing may cause the parameter distribution features of different failure modes to overlap in the process of extracting the time domain associated mode features, and directly querying them may not accurately distinguish between normal states and fault states. Therefore, in the technical solution of the present application, in order to reduce the confusion between different fault mode features, the set of implicit feature vectors associated with the multimodal time series of the reference state parameters is subjected to semantic essential expression difference enhancement modulation, and then cross-domain query encoding is performed with the implicit feature vectors associated with the multimodal time series of the real-time state parameters to obtain a real-time state optimized query response representation vector. Among them, the cross-domain query encoding enhances the distinguishability of the fault state features and improves the query capability of the fault by learning the essential features of the reference state features and modulating them. In particular, the cross-domain query encoding module based on hierarchical modulation of essential features is a new type of query encoding processing architecture based on deep learning, which aims to enhance the information interaction and expression ability of the model between different fields through multi-stage data parsing and reorganization. By introducing a cross-domain attention mechanism, it enables the model to cross different data types and structures to achieve deeper information fusion, thereby improving the performance of multimodal tasks. At the same time, in complex real-world applications, raw data usually contains a lot of redundancy and noise. The essential feature capture and hierarchical modulation process in this module helps to identify and highlight key features, while reducing unnecessary information interference and increasing the model's attention to important signals. Specifically, in a specific example of the present application, such as Figure 3 As shown, the query encoding module 350 includes: an essential feature extraction unit 351, which is used to input the set of reference state parameter multimodal temporal association implicit feature vectors into an essential feature capture network to obtain a reference state parameter essential feature representation vector; a feature weight modulation unit 352, which is used to perform semantic expression hierarchical modulation on each reference state parameter multimodal temporal association implicit feature vector in the set of reference state parameter multimodal temporal association implicit feature vectors based on the reference state parameter essential feature representation vector to obtain a set of modulated reference state parameter multimodal temporal association implicit feature vectors; a cross-domain query encoding unit 353, which is used to perform cross-domain query encoding on the real-time state parameter multimodal temporal association implicit feature vector and the set of modulated reference state parameter multimodal temporal association implicit feature vectors to obtain the real-time state optimized query response representation vector.

[0039] Specifically, the essential feature extraction unit 351 is used to input the set of implicit feature vectors of the multimodal temporal association of the reference state parameters into an essential feature capture network to obtain a reference state parameter essential feature representation vector. That is, the set of implicit feature vectors of the multimodal temporal association of the reference state parameters is input into an essential feature capture network designed to extract intrinsic attributes or core structures from data. The essential feature capture network is suitable for learning and extracting complex reference state parameter multimodal temporal association patterns from the set of implicit feature vectors of the multimodal temporal association of the reference state parameters, and compressing it into a fixed-length reference state parameter essential feature representation vector, so as to retain important information, reduce noise, and provide a reference anchor point for subsequent processing.

[0040] In an embodiment of the present application, the specific process of inputting the set of reference state parameter multimodal temporal association implicit feature vectors into the essential feature capture network includes: calculating the semantic difference coefficient of each reference state parameter multimodal temporal association implicit feature vector in the set of reference state parameter multimodal temporal association implicit feature vectors relative to other reference state parameter multimodal temporal association implicit feature vectors to obtain a set of reference state parameter semantic difference coefficients; taking the opposite of each reference state parameter semantic difference coefficient in the set of reference state parameter semantic difference coefficients and performing normalization processing based on the Softmax function to obtain a set of normalized reference state parameter semantic difference coefficients; using the set of normalized reference state parameter semantic difference coefficients as a set of weights, calculating the position-weighted sum of the set of reference state parameter multimodal temporal association implicit feature vectors to obtain the reference state parameter essential feature representation vector.

[0041] Among them, the specific process of calculating the semantic difference coefficient of each reference state parameter multimodal time series association implicit feature vector in the set of reference state parameter multimodal time series association implicit feature vectors relative to other reference state parameter multimodal time series association implicit feature vectors includes: calculating the mean of the norm of the positional difference vector between the reference state parameter multimodal time series association implicit feature vector and other reference state parameter multimodal time series association implicit feature vectors in the set of reference state parameter multimodal time series association implicit feature vectors as the reference state parameter semantic difference coefficient of the reference state parameter multimodal time series association implicit feature vector.

[0042] In summary, in the above embodiment, the set of implicit feature vectors associated with the multimodal time series of the reference state parameters is input into the essential feature capture network to obtain the essential feature representation vector of the reference state parameters, including: extracting the essential features of the set of implicit feature vectors associated with the multimodal time series of the reference state parameters using the following essential feature extraction formula to obtain the essential feature representation vector of the reference state parameters; wherein the essential feature extraction formula is:

[0043] in, represents the set of implicit feature vectors of multimodal temporal association of the reference state parameters, , , and respectively represent the first, second, and third in the set of multimodal temporal association implicit feature vectors of the reference state parameter and The multi-modal temporal association implicit eigenvector of reference state parameters, is the number of implicit eigenvectors associated with the multimodal time series of the reference state parameters, represents exponential operation, represents the one-norm of a vector, Indicates The semantic difference coefficient of the reference state parameter corresponding to the implicit feature vector of the multimodal temporal association of the reference state parameter, The reference state parameter essential characteristic represents a vector.

[0044] Specifically, the feature weight modulation unit 352 is used to perform semantic expression hierarchical modulation on each reference state parameter multimodal temporal association implicit feature vector in the set of the reference state parameter multimodal temporal association implicit feature vector based on the reference state parameter essential feature representation vector to obtain a set of modulated reference state parameter multimodal temporal association implicit feature vectors. That is, in the technical solution of the present application, by calculating the semantic contribution coefficient between each reference state parameter multimodal temporal association implicit feature vector in the set of the reference state parameter multimodal temporal association implicit feature vector and the reference state parameter essential feature representation vector, the specific contribution degree of each reference state parameter multimodal temporal association implicit feature vector to the overall semantic expression is determined to obtain a set of reference state semantic contribution coefficients. Subsequently, based on the reference state semantic contribution coefficient, the set of the reference state parameter multimodal temporal association implicit feature vector is hierarchically modulated to widen the semantic essential expression difference between each reference state parameter multimodal temporal association implicit feature vector in the set of the reference state parameter multimodal temporal association implicit feature vector. Through hierarchical modulation, the expressiveness of important features can be enhanced while suppressing the influence of minor features. In this way, the model can better focus on those features that are critical to the final decision, thereby improving robustness and generalization capabilities.

[0045] In an embodiment of the present application, based on the reference state parameter essential feature representation vector, the specific process of performing semantic expression hierarchical modulation on each reference state parameter multimodal timing associated implicit feature vector in the set of reference state parameter multimodal timing associated implicit feature vectors includes: calculating the reference state semantic contribution coefficient between each reference state parameter multimodal timing associated implicit feature vector in the set of reference state parameter multimodal timing associated implicit feature vectors and the reference state parameter essential feature representation vector to obtain a set of reference state semantic contribution coefficients; based on the set of reference state semantic contribution coefficients, performing hierarchical modulation on the set of reference state parameter multimodal timing associated implicit feature vectors to widen the semantic essential expression differences between each reference state parameter multimodal timing associated implicit feature vector in the set of reference state parameter multimodal timing associated implicit feature vectors to obtain a set of modulated reference state parameter multimodal timing associated implicit feature vectors.

[0046] Among them, the specific process of calculating the reference state semantic contribution coefficient between each reference state parameter multimodal time series associated implicit feature vector in the set of reference state parameter multimodal time series associated implicit feature vectors and the reference state parameter essential feature representation vector includes: calculating the covariance matrix between each reference state parameter multimodal time series associated implicit feature vector in the set of reference state parameter multimodal time series associated implicit feature vectors and the reference state parameter essential feature representation vector; calculating the differential feature vector between each reference state parameter multimodal time series associated implicit feature vector in the set of reference state parameter multimodal time series associated implicit feature vectors and the reference state parameter essential feature representation vector; calculating the product of the transposed vector of the differential feature vector, the inverse matrix of the covariance matrix and the differential feature vector, and performing a quadratic root operation on the calculated result to obtain the set of reference state semantic contribution coefficients.

[0047] More specifically, based on the set of reference state semantic contribution coefficients, the specific process of hierarchically modulating the set of implicit feature vectors of the multimodal temporal association of the reference state parameters includes: setting a first mask threshold and a second mask threshold, wherein the second mask threshold is twice the first mask threshold; if the eigenvalue in the implicit feature vector of the multimodal temporal association of the reference state parameters is greater than the second mask threshold, amplifying it by a factor of two; if the eigenvalue in the implicit feature vector of the multimodal temporal association of the reference state parameters is greater than the first mask threshold and is less than or equal to the second mask threshold, keeping it unchanged; if the eigenvalue in the implicit feature vector of the multimodal temporal association of the reference state parameters is less than or equal to the first mask threshold, reducing it by a factor of two, thereby obtaining the set of implicit feature vectors of the multimodal temporal association of the reference state parameters after modulation.

[0048] In summary, in the above embodiment, based on the reference state parameter essential feature representation vector, each reference state parameter multimodal time series associated implicit feature vector in the set of the reference state parameter multimodal time series associated implicit feature vectors is subjected to semantic expression hierarchical modulation to obtain a set of modulated reference state parameter multimodal time series associated implicit feature vectors, including: based on the reference state parameter essential feature representation vector, each reference state parameter multimodal time series associated implicit feature vector is subjected to semantic expression hierarchical modulation using the following hierarchical modulation formula to obtain a set of modulated reference state parameter multimodal time series associated implicit feature vectors; wherein the hierarchical modulation formula is:

[0049] in, represents the covariance matrix, Indicates The reference state semantic contribution coefficient corresponding to the implicit feature vector of the multimodal temporal association of the reference state parameters, For masking, is a predetermined threshold, Represents a set of implicit feature vectors of multi-modal temporal association of the modulated reference state parameters.

[0050] Specifically, the cross-domain query encoding unit 353 is used to perform cross-domain query encoding on the set of the multi-modal temporal association implicit feature vector of the real-time state parameter and the multi-modal temporal association implicit feature vector of the modulated reference state parameter to obtain the real-time state optimization query response representation vector. That is, the set of the multi-modal temporal association implicit feature vector of the real-time state parameter and the multi-modal temporal association implicit feature vector of the modulated reference state parameter is input into the cross-domain query encoding attention network to generate an attention score according to the interaction between the multi-modal temporal association implicit feature vector of the real-time state parameter and the multi-modal temporal association implicit feature vector of the modulated reference state parameter, reflecting the correlation of different feature combinations, so as to obtain a set of real-time state cross-domain query attention coefficients. In order to ensure the numerical stability of the attention weight and make it a smooth and interpretable probability distribution, the set of the real-time state cross-domain query attention coefficient is further input into the Sigmoid activation function as a nonlinear mapping tool to convert it into a weight value within the normalized range. In this way, it can be ensured that the features in the multi-modal temporal association implicit feature vectors of the modulated reference state parameters are properly taken into account, while also avoiding the gradient vanishing or exploding problems in extreme cases.

[0051] In an embodiment of the present application, the specific steps of cross-domain query encoding for the set of the real-time state parameter multimodal temporal association implicit feature vector and the modulated reference state parameter multimodal temporal association implicit feature vector include: inputting each modulated reference state parameter multimodal temporal association implicit feature vector in the set of the real-time state parameter multimodal temporal association implicit feature vector and the modulated reference state parameter multimodal temporal association implicit feature vector into the cross-domain query coding attention network to obtain a set of real-time state cross-domain query attention coefficients; inputting the set of real-time state cross-domain query attention coefficients into the weight conversion network based on the Sigmoid activation function to obtain a set of real-time state cross-domain query attention weights; finally, taking the set of real-time state cross-domain query attention weights as the set of weights, performing element-by-element multiplication and then accumulation operation on the set of reference state parameter multimodal temporal association implicit feature vectors to form the final real-time state optimization query response representation vector. Through the attention weight, the correlation between the real-time state feature and the reference state feature can be enhanced, the real-time state optimization query feature representation can be optimized, and the accuracy of subsequent classifier classification can be improved.

[0052] Among them, the specific steps of inputting the implicit feature vector of the multimodal time series association of the real-time state parameter and each post-modulation reference state parameter multimodal time series association implicit feature vector in the set of the multimodal time series association implicit feature vector of the modulated reference state parameter into the cross-domain query coding attention network include: calculating the product between the transposed vectors of the multimodal time series association implicit feature vector of the real-time state parameter and each post-modulation reference state parameter multimodal time series association implicit feature vector in the set of the multimodal time series association implicit feature vector of the modulated reference state parameter to obtain a set of global fine-grained correlations; calculating the product between a norm of the multimodal time series association implicit feature vector of the real-time state parameter and a norm of each post-modulation reference state parameter multimodal time series association implicit feature vector to obtain a set of scale modulation factors; dividing each global fine-grained correlation in the set of global fine-grained correlations by the corresponding scale modulation factor in the set of scale modulation factors to obtain a set of real-time state cross-domain query attention coefficients.

[0053] In summary, in the above embodiment, the set of the real-time state parameter multimodal timing associated implicit feature vector and the modulated reference state parameter multimodal timing associated implicit feature vector is cross-domain queried and encoded to obtain the real-time state optimization query response representation vector, including: the set of the real-time state parameter multimodal timing associated implicit feature vector and the modulated reference state parameter multimodal timing associated implicit feature vector is cross-domain queried and encoded to obtain the real-time state optimization query response representation vector according to the following cross-domain query encoding formula; wherein the cross-domain query encoding formula is:

[0054] in, represents the multi-modal temporal association implicit feature vector of the real-time state parameter, Indicates the real-time state cross-domain query attention coefficient corresponding to the multi-modal temporal association implicit feature vector of the modulated reference state parameter, yes function, The first one in the set of attention weights of the real-time state cross-domain query Real-time state cross-domain query attention weights, The real-time status optimization query response representation vector is represented.

[0055] In particular, the abnormality detection module 360 ​​is used to determine whether the state of the monitored electrical device is abnormal based on the real-time state optimization query response representation vector. In a specific example of the present application, the real-time state optimization query response representation vector is input into a detection module based on a classifier to obtain a detection result, and the detection result is used to indicate whether the state of the monitored electrical device is abnormal. In particular, the classifier is mainly based on machine learning and statistical methods, and learns the features of different category labels through training data to classify the real-time state optimization query response representation vector. Specifically, in the technical solution of the present application, the labels of the classifier include the state of the monitored electrical device is abnormal (first label), and the state of the monitored electrical device is not abnormal (second label), wherein the classifier determines which classification label the real-time state optimization query response representation vector belongs to by a soft maximum function. It is worth noting that the first label p1 and the second label p2 here do not contain artificially set concepts. In fact, in the training process, the computer model does not have the concept of "whether the state of the monitored electrical device is abnormal". It only has two classification labels and the probability of the output feature under these two classification labels, that is, the sum of p1 and p2 is one. Therefore, the classification result of whether the state of the monitored electrical equipment is abnormal is actually converted into a binary classification probability distribution that conforms to the laws of nature through the classification label. In essence, the physical meaning of the natural probability distribution of the label is used, rather than the linguistic text meaning of "whether the state of the monitored electrical equipment is abnormal". In this way, the system can quickly and accurately determine whether the equipment has potential faults, thereby providing timely warnings to ensure the stable and safe operation of the power system.

[0056] Preferably, inputting the real-time state optimization query response representation vector into a classifier-based detection module to obtain a detection result comprises: Determine the number of zero eigenvalues ​​in the real-time state optimization query response representation vector, and respectively calculate the reciprocal of the logarithm of the number of zero eigenvalues ​​with base two and the exponential value of the reciprocal of the number of zero eigenvalues ​​with base natural constants to obtain a first real-time state optimization query response representation value and a second real-time state optimization query response representation value; Calculate a power function of each eigenvalue of the real-time state optimization query response representation vector with the inverse of the number of zero eigenvalues ​​as an exponent to obtain a real-time state optimization query response structural state vector; Performing a dot multiplication of the real-time state optimization query response structure state vector and the first real-time state optimization query response representation value to obtain a real-time state optimization query response information representation vector; Performing dot multiplication of the autocorrelation matrix of the real-time state optimization query response structure state vector with the second real-time state optimization query response representation value and weight hyperparameter respectively to obtain a real-time state optimization query response regression understanding matrix; Performing matrix multiplication of the real-time state optimization query response information representation vector and the real-time state optimization query response regression understanding matrix to obtain an optimized real-time state optimization query response representation vector; and The optimized real-time state optimized query response representation vector is input into a classifier-based detection module to obtain a detection result.

[0057] That is, the first real-time state optimization query response representation value and the second real-time state optimization query response representation value are represented as:

[0058] in is the number of zero eigenvalues ​​in the real-time state optimization query response representation vector, is the first real-time status optimization query response representation value, is the second real-time status optimization query response representation value, represents the logarithm operation with base 2, Represents exponential operation; And the optimized real-time state optimized query response representation vector is expressed as:

[0059] The real-time state optimization query response representation vector is a row vector, represents calculating a power function value of each eigenvalue of the real-time state optimization query response representation vector with the inverse of the number of zero eigenvalues ​​as the exponent, is the real-time status optimization query response information representation vector, is the weight hyperparameter, is the real-time state optimization query response regression understanding matrix, represents matrix multiplication, It means point multiplication by position. is the optimized real-time state optimized query response representation vector.

[0060] Here, when the set of the real-time state parameter multimodal time series associated implicit feature vector and the reference state parameter multimodal time series associated implicit feature vector respectively represent the state parameter time series pattern characteristics of the monitored electrical equipment and the state parameter multimodal time series pattern characteristics marked as a fault state, when encoding the cross-domain query based on the hierarchical modulation of the essential features, it is considered that there is noise in the state parameters during collection, which makes the existence of noise-based interference selective state space during cross-domain query encoding, so that the real-time state optimization query response representation vector has a remote class regression understanding representation imbalance, thereby deviating from the source time series semantic feature representation class understanding, affecting the accuracy of the classification result.

[0061] Based on this, the present application realizes effective long-range modeling of the real-time state optimization query response representation vector under the field linear complexity by modeling the vector field zero dimension of the real-time state optimization query response representation vector as a global field redundant dependency. In this way, in the feature regression process of the real-time state optimization query response representation vector, a structured correlation self-distillation between information representation and class understanding is performed in a selective state space based on feature distribution, thereby maintaining the overall information complexity of the feature distribution of the real-time state optimization query response representation vector while promoting the long-range dynamic class understanding balance of its feature convergence, so as to improve the class understanding consistency between the classification regression and feature extraction processes, and improve the accuracy of the detection results obtained by the detection module based on the classifier when the real-time state optimization query response representation vector is input.

[0062] As described above, the intelligent detection system 300 for safety electrical devices according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an intelligent detection algorithm for safety electrical devices. In a possible implementation, the intelligent detection system 300 for safety electrical devices according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent detection system 300 for safety electrical devices can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the intelligent detection system 300 for safety electrical devices can also be one of the many hardware modules of the wireless terminal.

[0063] Alternatively, in another example, the intelligent detection system 300 for safety electrical devices and the wireless terminal may also be separate devices, and the intelligent detection system 300 for safety electrical devices may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0064] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent detection system for safety electrical devices, characterized in that: include: A data acquisition module, used to obtain a data set of status parameters of the monitored electrical equipment collected by the sensor group; A data reconstruction module, used for structuring the data set of the state parameters of the monitored electrical equipment to obtain a real-time state parameter full-time domain full-sample space aggregation matrix; A reference sample extraction module, used to extract a set of full-time domain full-sample spatial aggregate reference matrices of state parameters marked as fault states from a database; A feature extraction module, used to extract a set of real-time state parameter multi-modal time series associated implicit feature vectors and a set of reference state parameter multi-modal time series associated implicit feature vectors from the set of the real-time state parameter full-time domain full-sample space aggregate matrix and the state parameter full-time domain full-sample space aggregate reference matrix marked as a fault state; A query encoding module, configured to perform semantic essential expression difference enhancement modulation on the set of implicit feature vectors associated with the multimodal time series of the reference state parameters, and then perform cross-domain query encoding with the implicit feature vectors associated with the multimodal time series of the real-time state parameters to obtain a real-time state optimization query response representation vector; The anomaly detection module is used to determine whether there is an anomaly in the state of the monitored electrical equipment based on the real-time state optimization query response representation vector.

2. The intelligent detection system for safety electrical devices according to claim 1, characterized in that: The state parameters include current value, voltage value and temperature value.

3. The intelligent detection system for safety electrical devices according to claim 2, characterized in that: The feature extraction module is used to: A device state multimodal parameter time domain correlation pattern feature extractor based on a void convolutional neural network model is used to perform void convolution encoding on each state parameter full time domain full sample space aggregation reference matrix in the set of the real-time state parameter full time domain full sample space aggregation reference matrix of the state parameter marked as a fault state, so as to obtain the real-time state parameter multimodal time series correlation implicit feature vector and the reference state parameter multimodal time series correlation implicit feature vector.

4. The intelligent detection system for safety electrical devices according to claim 3, characterized in that: The query encoding module comprises: An essential feature extraction unit is used to input the set of implicit feature vectors associated with the multi-modal time series of the reference state parameters into an essential feature capture network to obtain an essential feature representation vector of the reference state parameters; A feature weight modulation unit, configured to perform semantic expression hierarchical modulation on each reference state parameter multimodal temporal associated implicit feature vector in the set of reference state parameter multimodal temporal associated implicit feature vectors based on the reference state parameter essential feature representation vector to obtain a set of modulated reference state parameter multimodal temporal associated implicit feature vectors; The cross-domain query encoding unit is used to perform cross-domain query encoding on the set of the real-time state parameter multi-modal temporal association implicit feature vector and the modulated reference state parameter multi-modal temporal association implicit feature vector to obtain the real-time state optimization query response representation vector.

5. The intelligent detection system for safety electrical devices according to claim 4, characterized in that: The essential feature extraction unit comprises: A reference state parameter semantic difference coefficient calculation subunit, used to calculate the semantic difference coefficient of each reference state parameter multimodal temporal association implicit feature vector in the set of reference state parameter multimodal temporal association implicit feature vectors relative to other reference state parameter multimodal temporal association implicit feature vectors to obtain a set of reference state parameter semantic difference coefficients; A normalization subunit, used for taking the inverse of each reference state parameter semantic difference coefficient in the set of reference state parameter semantic difference coefficients and performing normalization processing based on a Softmax function to obtain a set of normalized reference state parameter semantic difference coefficients; The weighting subunit is used to calculate the position-weighted sum of the set of multimodal temporal association implicit feature vectors of the reference state parameters using the set of normalized reference state parameter semantic difference coefficients as a set of weights to obtain the essential feature representation vector of the reference state parameters.

6. The intelligent detection system for safety electrical devices according to claim 5, characterized in that: The feature weight modulation unit comprises: A semantic contribution coefficient calculation subunit, used to calculate the reference state semantic contribution coefficient between each reference state parameter multimodal temporal association implicit feature vector in the set of reference state parameter multimodal temporal association implicit feature vectors and the reference state parameter essential feature representation vector to obtain a set of reference state semantic contribution coefficients; A modulation subunit is used to hierarchically modulate the set of reference state parameter multimodal timing associated implicit feature vectors based on the set of reference state semantic contribution coefficients so as to widen the semantic essential expression differences between the individual reference state parameter multimodal timing associated implicit feature vectors in the set of reference state parameter multimodal timing associated implicit feature vectors so as to obtain a set of modulated reference state parameter multimodal timing associated implicit feature vectors.

7. The intelligent detection system for safety electrical devices according to claim 6, characterized in that: The semantic contribution coefficient calculation subunit is used to: Calculating the covariance matrix between each reference state parameter multimodal time series associated implicit feature vector in the set of reference state parameter multimodal time series associated implicit feature vectors and the reference state parameter essential feature representation vector; Calculating a differential feature vector between each reference state parameter multimodal time series associated implicit feature vector in the set of reference state parameter multimodal time series associated implicit feature vectors and the reference state parameter essential feature representation vector; The product of the transposed vector of the differential eigenvector, the inverse matrix of the covariance matrix, and the differential eigenvector is calculated, and a quadratic root operation is performed on the calculated result to obtain a set of reference state semantic contribution coefficients.

8. The intelligent detection system for safety electrical devices according to claim 7, characterized in that: The cross-domain query encoding unit includes: A cross-domain query attention coefficient calculation subunit, used for inputting each post-modulation reference state parameter multimodal time series associated implicit feature vector in the set of the real-time state parameter multimodal time series associated implicit feature vector and the post-modulation reference state parameter multimodal time series associated implicit feature vector into a cross-domain query encoding attention network to obtain a set of real-time state cross-domain query attention coefficients; A weight conversion subunit, used for inputting the set of real-time state cross-domain query attention coefficients into a weight conversion network based on a Sigmoid activation function to obtain a set of real-time state cross-domain query attention weights; The optimized query subunit is used to calculate the position-weighted sum of the set of multimodal temporal association implicit feature vectors of the reference state parameters using the set of real-time state cross-domain query attention weights as a set of weights to obtain the real-time state optimized query response representation vector.

9. The intelligent detection system for safety electrical devices according to claim 8, characterized in that: The cross-domain query attention coefficient calculation subunit is used to: Calculating the product between the transposed vectors of each post-modulation reference state parameter multimodal time series associated implicit feature vector in the set of the real-time state parameter multimodal time series associated implicit feature vector and the post-modulation reference state parameter multimodal time series associated implicit feature vector to obtain a set of global fine-grained correlation degrees; Calculating the product of a norm of the real-time state parameter multi-modal time series associated implicit feature vector and a norm of each modulated reference state parameter multi-modal time series associated implicit feature vector to obtain a set of scale modulation factors; Each global fine-grained relevance in the set of global fine-grained relevance is divided by a corresponding scale modulation factor in the set of scale modulation factors to obtain a set of real-time state cross-domain query attention coefficients.

10. The intelligent detection system for safety electrical devices according to claim 9, characterized in that: The abnormality detection module is used to: input the real-time state optimization query response representation vector into a classifier-based detection module to obtain a detection result, and the detection result is used to indicate whether the state of the monitored electrical equipment is abnormal.

Images