A multi-modal feature fusion recognition method and device for partial discharge of a high-voltage cable

Abstract

The application provides a kind of high-voltage cable partial discharge multimodal feature fusion identification method and device, related to high-voltage cable partial discharge multimodal feature fusion technical field, the three-phase current, temperature and the multimodal time series data of ground voltage of high-voltage cable are collected in the application, form multidimensional input matrix after pre-processing and time alignment, then construct physical relationship matrix based on cable physical characteristics, and generate the feature representation after fusion by injecting physical constraints into data through feature modulation, input the feature into the LSTM autoencoder trained with normal data, reconstruct the sequence and calculate the reconstruction error, finally, the comprehensive anomaly score is obtained by exponentially weighted algorithm on error time series accumulation, through comparison with dynamic threshold, realize the accurate identification of insulation state anomaly and partial discharge alarm.

Description

Technical Field

[0001] This invention relates to the field of multimodal feature fusion technology for partial discharge of high-voltage cables, specifically to a multimodal feature fusion identification method and device for partial discharge of high-voltage cables. Background Technology

[0002] High-voltage cables are key infrastructure for urban power grids and long-distance power transmission. The reliability of their insulation status is closely related to the safe operation of the power system. Partial discharge is an important phenomenon characterizing cable insulation degradation. Accurate identification and early warning of partial discharge is of great significance for preventing cable breakdown accidents and improving the level of intelligent operation and maintenance. Existing online monitoring systems usually rely on multi-modal sensor data such as three-phase current, conductor temperature and voltage to ground, aiming to detect insulation anomalies by analyzing the time-series changes of these parameters. Currently, data-driven anomaly detection methods are the mainstream technology in this field. In particular, deep learning models such as long short-term memory networks and autoencoders are used to learn the time-series patterns under normal conditions and identify deviations by reconstructing errors. Although these methods can capture statistical regularities in the data, they are essentially black-box models that learn features entirely from the data. They lack the utilization of the inherent physical laws of high-voltage cable systems. For example, there is a thermal effect relationship between current and temperature determined by Joule's law, the three-phase currents are coupled to each other through electromagnetic induction, and the voltage and current are constrained by Ohm's law. Existing technologies ignore these key physical prior knowledge, causing the features learned by the model to deviate from the real physical process. In complex operating conditions, false alarms or missed alarms are likely to occur. A further shortcoming is that existing methods only perform simple data-level fusion or independent processing on various physical quantities, failing to explicitly model the coupling relationship between different physical quantities governed by physical laws. This limitation makes the model limited in its ability to distinguish between abnormalities caused by real partial discharge and normal load fluctuations or external disturbances, resulting in significant bottlenecks in its generalization performance and interpretability.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-modal feature fusion identification method and device for partial discharge of high-voltage cables, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-modal feature fusion and identification method for partial discharge in high-voltage cables, comprising the following steps: Step 1: Collect multi-mode time series data of the target high-voltage cable equipment. The multi-mode time series data includes three-phase current, three-phase conductor temperature and voltage to ground. Preprocess and time-align the multi-mode time series data to obtain the preprocessed multi-dimensional input matrix. Step 2: Construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, and output the fused feature representation; Step 3: Input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. Step 4: Using an exponential decay weighted recursive algorithm, the reconstruction error is accumulated over time to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with a dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output.

[0006] Furthermore, the multi-mode time series data includes three-phase current, three-phase conductor temperature, and voltage to ground. The logic for preprocessing and time alignment of the multi-mode time series data is as follows: Three-phase current refers to the current flowing in the three-phase conductors A, B, and C. Current transformers are used to collect the current values ​​flowing in the three-phase conductors A, B, and C in real time to reflect the load condition of the line. Three-phase conductor core temperature refers to the temperature measurement value of the corresponding three-phase conductor core. This temperature measurement is directly obtained by fiber optic sensors embedded near the cable core. Voltage to ground refers to the voltage of each of the three-phase conductors A, B, and C relative to ground, measured by voltage transformers. This allows for the collection of nine physical quantities—three-phase current, three-phase conductor core temperature, and voltage to ground—forming a single physical quantity set. The multimodal time series data is based on the current time. As the endpoint, backtrack a length of The time window formed Within, at fixed sampling intervals The collected three-phase current time series, three-phase conductor temperature time series, and voltage-to-ground time series; An outlier detection method based on statistical interquartile range (IIR) is employed to identify and remove data points in multimodal time-series data that exceed 1.5 times the IIR of the upper quartile or fall below 1.5 times the IIR of the lower quartile. For missing values ​​caused by data removal and transmission interruptions, linear interpolation is used to calculate the imputation value by establishing a linear relationship between the two nearest valid data points before and after the missing value. A min-max scaling method is used to linearly transform the values ​​of each physical quantity to [the appropriate scale]. interval; Time alignment with uniform sampling time intervals Using this as a baseline, a corresponding timestamp sequence is generated for all modalities. For data not sampled at the baseline time point, linear interpolation is used to align it to a unified timestamp sequence, ultimately forming a sequence with dimension [missing information]. Multidimensional input matrix ,in This represents the total number of physical quantities, with a value of 9. For time window The total number of sampling time points within, the first Each sampling time point represents the current moment.

[0007] Furthermore, the logic for constructing a predefined physical relationship matrix based on the physical characteristics of high-voltage cables is as follows: Physical Relationship Matrix It is a non-negative symmetric matrix predefined based on Ohm's law, the principle of thermal effect, and the principle of electromagnetic induction for high-voltage cables. The values ​​of the matrix elements are determined by quantifying the coupling strength between physical quantities according to the following principles: The diagonal elements are all 1, representing complete coupling between each physical quantity and itself. The off-diagonal elements are assigned values ​​based on the actual coupling strength between the physical quantities; the stronger the coupling, the larger the element value. The range of element values ​​is... , where 0 indicates no direct physical connection and 1 indicates complete coupling; The specific coupling strength is assigned based on Ohm's law, the principle of thermal effect, and the principle of electromagnetic induction. A strong coupling value is assigned between current and temperature in the same phase, a medium coupling value is assigned between currents in different phases, and a weak coupling value is assigned between voltage and current.

[0008] Furthermore, based on a physical perception attention mechanism, feature modulation of the multidimensional input matrix is ​​achieved through the physical relation matrix, specifically as follows: Using physical relationship matrix For multidimensional input matrices Explicit feature modulation is performed to calculate the feature representation that incorporates physical constraints. The feature modulation is achieved through the following formula: in, for An identity matrix of dimension 1 For the modulated output Feature representation of dimensional fusion is a non-negative modulation coefficient used to control the injection intensity based on physical priors. It degenerates into a multidimensional input matrix. At that time, through The defined physical relationships are explicitly incorporated into the feature representation.

[0009] Furthermore, the fused feature representation The data is input to a time-series autoencoder trained offline. Both the encoder and decoder are composed of multi-layer long short-term memory network units. The training process of the time-series autoencoder is carried out offline, and the training data used are all historical data of high-voltage cables under normal insulation conditions. The fused feature representation The input is fed to a temporal autoencoder, which maps the input sequence to a latent spatial feature representation through its encoder, and then reconstructs a reconstructed sequence with the same dimension as the input sequence from the latent spatial feature representation through its decoder. In order to learn the time dependency patterns of data under normal conditions; Computational fusion of feature representations With reconstruction sequence The mean square error between them is taken as the reconstruction error, and its reconstruction error The calculation formula is: in, Feature representation for fusion The value of the nth physical quantity at the tth sampling time point. To reconstruct the values ​​of the corresponding physical quantities in the sequence at the corresponding time points. For physical quantities, This is the index for the sampling time point.

[0010] Furthermore, an exponentially decaying weighted recursive algorithm is used to accumulate the reconstruction error over time to obtain a comprehensive anomaly score, which smooths random fluctuations and highlights persistent anomalies. The calculation formula is as follows: in, For the first Comprehensive anomaly score for each sampling time point, For the first Comprehensive anomaly score for each sampling time point, This is a decay factor used to control the weighting of the current reconstruction error when updating the comprehensive anomaly score.

[0011] Furthermore, the dynamic threshold is based on a time window. The statistical characteristics of the internal comprehensive anomaly scores are calculated for each sampling time point within the window. All are based on having it as the endpoint and a length of Data window The corresponding comprehensive anomaly score is calculated internally, thereby obtaining the time window. A complete historical anomaly sequence; Based on the statistical characteristics of this historical comprehensive anomaly sequence, a dynamic threshold is calculated. Its formula is: in, and Time windows The mean and standard deviation of all composite outlier scores. This is the sensitivity coefficient; The comprehensive anomaly score at the current moment is compared with the dynamic threshold corresponding to that moment. If the comprehensive anomaly score is greater than the dynamic threshold, an instantaneous anomaly flag is generated. If the instantaneous anomaly flag appears continuously for M times, the insulation state of the target high-voltage cable is finally determined to be abnormal, and the discharge identification result and warning signal are output.

[0012] The present invention also provides a multi-modal feature fusion and identification device for partial discharge of high-voltage cables. The device is used to execute the above-described multi-modal feature fusion and identification method for partial discharge of high-voltage cables, comprising: The data acquisition module is used to acquire multi-modal time-series data of the target high-voltage cable equipment. The multi-modal time-series data includes three-phase current, three-phase conductor temperature and voltage to ground. The multi-modal time-series data is preprocessed and time-aligned to obtain a preprocessed multi-dimensional input matrix. The physical construction module is used to construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, and to use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, outputting a fused feature representation. The feature fusion module is used to input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. The result judgment module is used to accumulate the reconstruction error over time using an exponential decay weighted recursive algorithm to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with a dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention explicitly encodes the coupling relationships between different physical quantities determined by Ohm's law, Joule heating effect, and electromagnetic induction principle through the constructed physical relationship matrix, and injects this physical prior knowledge into the feature representation with the help of feature modulation formula. This makes the time-series autoencoder learn and reconstruct no longer the original data itself, but a more interpretable feature representation that integrates physical constraints, thereby greatly enhancing the model's understanding of the balance relationship that should be satisfied between various physical quantities under normal conditions. This invention is more sensitive and accurate in detecting anomalies. When the cable insulation condition is abnormal, such as when partial discharge occurs, it disrupts this inherent physical balance, causing the modulated feature representation to deviate significantly from the normal mode, thus generating significant reconstruction errors. It can effectively distinguish between real insulation faults and normal load fluctuations or external interference, thereby significantly reducing false alarms and missed alarms. This invention further smooths and adaptively judges the reconstructed error sequence through an exponentially weighted recursive algorithm and a dynamic threshold mechanism based on recent historical statistics, ultimately outputting reliable comprehensive anomaly analysis and alarm signals. The synergistic effect of these technical features enables more accurate, reliable, and physically interpretable intelligent identification and early warning of the insulation status of high-voltage cables, improving the intelligence level and practical value of the condition monitoring system. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a bar-dot graph of the comprehensive anomaly score-attenuation factor at the current moment of this invention; Figure 3 This is a comparison chart of the comprehensive anomaly score at the previous moment and the comprehensive anomaly score at the current moment in this invention. Figure 4 This is the fitted curve of the reconstruction error versus the current time-time comprehensive anomaly score of the present invention; Figure 5 This is a curve showing the fitting of the comprehensive anomaly score at the previous time step and the comprehensive anomaly score at the current time step. Figure 6 This is a flowchart of the overall device structure of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] Example: Please see Figures 1-5 The present invention provides a technical solution: A multi-modal feature fusion and identification method for partial discharge in high-voltage cables, comprising the following steps: Step 1: Collect multi-mode time series data of the target high-voltage cable equipment. The multi-mode time series data includes three-phase current, three-phase conductor temperature and voltage to ground. Preprocess and time-align the multi-mode time series data to obtain the preprocessed multi-dimensional input matrix. Multimodal time series data includes three-phase current, three-phase conductor temperature, and voltage to ground. The logic for preprocessing and time alignment of the multimodal time series data is as follows: Three-phase current refers to the current flowing in the three-phase conductors A, B, and C. Current transformers are used to collect the current values ​​flowing in the three-phase conductors A, B, and C in real time to reflect the load condition of the line. Three-phase conductor core temperature refers to the temperature measurement value of the corresponding three-phase conductor core. This temperature measurement is directly obtained by fiber optic sensors embedded near the cable core. Voltage to ground refers to the voltage of each of the three-phase conductors A, B, and C relative to ground, measured by voltage transformers. This allows for the collection of nine physical quantities—three-phase current, three-phase conductor core temperature, and voltage to ground—forming a single physical quantity set. Three-phase current directly reflects the load condition of the cable and the health of the conductive core. Insulation defects such as partial discharge can cause rapid charge migration, resulting in subtle distortions or high-frequency components in the current waveform. The three-phase core temperature is a key parameter reflecting the thermal state of the cable. Insulation aging and partial discharge can lead to increased dielectric loss, heat generation, and abnormal temperature rise. Monitoring the three-phase temperature helps to locate the overheated phase. The voltage to ground reflects the electrical stress borne by the main insulation of the cable. Voltage fluctuations and harmonic components are related to the excitation of partial discharge. Selecting data from these three modes constitutes a complete monitoring perspective from electrical (current, voltage) to thermodynamic (temperature), which can cross-verify the insulation status of the cable and provide a rich source of information for subsequent fusion analysis; The multimodal time series data is generated by backtracking a length of [length missing] with the current time as the endpoint. The time window formed Within, at fixed sampling intervals The collected three-phase current time series, three-phase conductor temperature time series, and voltage-to-ground time series, among which... For 1-60 seconds; Sampling interval The choice is based on a comprehensive consideration of the physical characteristics of partial discharge phenomena and the thermodynamic inertia of cable systems. Partial discharge itself is a rapid electromagnetic transient process on the nanosecond to microsecond scale, requiring an extremely high sampling rate to directly capture its pulse waveform. However, this invention achieves detection by analyzing the secondary effects caused by the discharge, namely the distortion of current and voltage and the cumulative effect of core temperature. Partial discharge will cause weak distortion and harmonics in the power frequency current and voltage waveforms. These changes occur on the time scale of the power frequency period (20ms) or its multiples. The energy dissipation at the partial discharge point will lead to local overheating, and the core temperature will show a slow upward trend through heat conduction. Temperature change is a slow process with extremely high inertia, and its time constant is usually on the order of seconds to minutes. The lower limit of 1 second allows for effective learning of waveform anomalies and short-term fluctuations in current and voltage through subsequent time-series models (LSTM). The accuracy of 1 second is also sensitive enough to track the initial temperature change phase measured in minutes. If the interval is shorter (e.g., 0.1 seconds), although more details can be captured, it will significantly increase the amount of data and the computational complexity of the model, while the marginal benefit to the improvement of detection performance is limited. The upper limit of 60 seconds is to ensure effective monitoring of the trend of slowly changing parameters such as temperature. If the sampling interval is too long (e.g., several minutes), the rapid temperature rise phase caused by violent discharge will be missed, resulting in detection delay. The 60-second interval can track temperature changes with reasonable resolution and matches the data recording cycle of most industrial monitoring systems, which has a good engineering practice basis. Time window Defines the range of data the analysis relies on, if Too short (less than 30 minutes) is insufficient to capture enough process dynamics and timing patterns, and is easily affected by transient noise. If the time is too long (more than 2-3 hours), it will introduce a large amount of outdated and irrelevant historical information, reduce the system's response speed to recent changes, and incur high computational costs. Starting from minutes, the time is fine-tuned based on the specific cable's operating characteristics and monitoring requirements: for scenarios with frequent load changes and requiring rapid response, it is appropriately shortened to 30-45 minutes; for scenarios with relatively stable conditions and requiring higher detection reliability, it is extended to 90-120 minutes. A fixed sampling interval defines the temporal resolution of the data. It needs to be determined based on the physical characteristics of the monitored phenomenon, such as the pulse width of partial discharge and the Nyquist sampling theorem, to ensure that key high-frequency information can be captured. A fixed interval is a prerequisite for subsequent time alignment and construction of a regular matrix. An outlier detection method based on statistical interquartile range (IIR) is employed to identify and remove data points in multimodal time-series data that exceed 1.5 times the IIR of the upper quartile or fall below 1.5 times the IIR of the lower quartile. For missing values ​​caused by data removal and transmission interruptions, linear interpolation is used to calculate the imputation value by establishing a linear relationship between the two nearest valid data points before and after the missing value. A min-max scaling method is used to linearly transform the values ​​of each physical quantity to [the appropriate scale]. interval; Data cleaning ensures data quality. Inevitably, outliers will be mixed into the raw sensor data, such as abrupt changes that do not conform to physical laws due to electromagnetic interference, sensor interruption, transmission errors, etc. The method is a robust statistical approach that calculates the upper and lower quartiles of the data. , ) and interquartile range ( The range of normal data is defined as follows: Points falling outside this range are considered outliers. Compared to methods based on the mean and standard deviation, the IQR method is less sensitive to extreme values ​​and less affected by outliers themselves, thus making it more robust. After outliers are removed or data packets are lost, gaps will appear in the time series. Many models (especially time series models) require the input to be a continuous and complete data sequence. Linear interpolation is a simple and effective method for filling missing values. For any missing data point, the arithmetic mean of the values ​​of the two nearest valid sampling points before and after it is taken as the filling value of the missing point. This method is based on the reasonable assumption that the data changes continuously and gradually between adjacent time points. It is suitable for physical quantities with relatively slow changes, such as temperature. It is not the optimal choice for high-frequency abrupt signals, but it is widely used in engineering practice because of its simplicity and effectiveness. For missing points within a time window, linear interpolation is used to fill them using the two nearest valid data points before and after the missing point. For consecutive missing values ​​at the beginning of the time window, forward filling is used, filling them with the value of the first valid data point after the missing point. For consecutive missing values ​​at the end of the time window (i.e., near the current time), backward filling is used, filling them with the value of the last valid data point before the missing point. If all data for a physical quantity is missing within the entire time window, the sequence of that physical quantity within this window is filled with 0 (or a specific invalid flag), and this physical quantity is considered for exclusion from the current analysis in subsequent processing. Mini-max scaling linearly maps all data for each physical quantity to The interval can accelerate model convergence, improve model performance and ensure numerical stability. Optimization algorithms such as gradient descent converge faster on normalized data, preventing overflow or instability caused by excessively large or small values ​​during the calculation process. Time alignment with uniform sampling time intervals Using this as a baseline, a corresponding timestamp sequence is generated for all modalities. For data not sampled at the baseline time point, linear interpolation is used to align it to a unified timestamp sequence, ultimately forming a sequence with dimension [missing information]. Multidimensional input matrix ,in This represents the total number of physical quantities, with a value of 9. For time window The total number of sampling time points within, the first Each sampling time point represents the current moment. In practical systems, different types of sensors have different sampling rates or non-strictly synchronized sampling clocks, which means that at any given time point, data from different modes are not collected simultaneously. Constructing a regular multidimensional input matrix, where each column strictly corresponds to the same time point and each row represents a complete time series of a physical quantity, is the basic requirement for matrix operations and time series model processing. For each modality's data, check if it has a sampled value at a standard time point. If not, use linear interpolation again to calculate its approximate value at the standard time point based on the nearest sampled values ​​before and after that time point. After alignment, all N physical quantities have values ​​at T unified time points, thus forming a regular... 3D matrix; The N rows represent a complete sequence of physical quantities, such as phase A current and phase B temperature, within the entire time window. The T columns represent a vector composed of the observation values ​​of all N physical quantities at the same point in time. This vector is a multimodal state snapshot of the cable at that moment.

[0018] Step 2: Construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, and output the fused feature representation; The physical relationship matrix is ​​essentially a coupling relationship model. It quantitatively expresses the mutual influence relationships between various physical quantities in a complex, multi-dimensional physical system (three-phase cable) in the form of a matrix. The logic for constructing a predefined physical relationship matrix based on the physical characteristics of high-voltage cables is as follows: Physical Relationship Matrix It is a non-negative symmetric matrix predefined based on Ohm's law, the principle of thermal effect, and the principle of electromagnetic induction for high-voltage cables. The values ​​of the matrix elements are determined by quantifying the coupling strength between physical quantities according to the following principles: The size of the matrix is ​​determined by the number of physical quantities. For example, if there are 3-phase currents, 3-phase temperatures, and 3 voltages to ground, then... This is a 9x9 square matrix that fully describes the pairwise relationships between these 9 quantities; the coupling strength is a concept of promotion or influence, not inhibition. For example, an increase in current leads to an increase in temperature, which is a positive and non-negative influence, so it is reasonable for the matrix element values ​​to be non-negative. Symmetry is a very important physical constraint. It is based on fundamental principles such as Newton's third law, which means that the coupling strength of physical quantity i to j is equivalent to the coupling strength of j to i. For example, the electromagnetic influence of phase A current on phase B current should be similar in strength to the influence of phase B current on phase A current. Symmetry greatly reduces the number of parameters that need to be defined and enhances the physical rationality of the model. i and j are the indices of physical quantities. The diagonal elements are all 1, representing complete coupling between each physical quantity and itself. The off-diagonal elements are assigned values ​​based on the actual coupling strength between the physical quantities; the stronger the coupling, the larger the element value. , where 0 indicates no direct physical connection and 1 indicates complete coupling; 0 indicates that the two physical quantities have no direct interaction path in the physical model under consideration. For example, the temperature of phase A has no direct and immediate significant effect on the current of phase C, so it can be set to 0. 1 indicates an ideal and extreme coupling situation, such as two completely identical physical quantities, which rarely occurs in actual systems. It provides a reference point for the upper bound. Values ​​between these values, such as 0.2, 0.5, 0.9, etc., are used to relatively represent the strength differences of different relationship pairs. The specific coupling strength assignments are based on Ohm's law, the principle of thermal effect, and the principle of electromagnetic induction. Strong coupling values ​​are assigned to current and temperature within the same phase, medium coupling values ​​to currents in different phases, and weak coupling values ​​to voltage and current. 1) Strong coupling values ​​are imparted between the current and temperature of the same phase, for example... Its physical principles are Joule's law and heat conduction. When the current in phase A ( When the conductor resistance increases, Joule heating will occur, causing the temperature of phase A conductor core to rise. A direct increase indicates a very strong and direct causal relationship. Because this relationship is so direct and strong, it is assigned a high value close to 1, such as 0.9. 2) Assigning moderate coupling values ​​between different phase currents, for example... Its physical principle is the principle of electromagnetic induction. The magnetic fields generated by the three-phase alternating current are intertwined. According to Faraday's law of electromagnetic induction, the changing magnetic field generated by the A-phase current will induce an electromotive force in the B-phase conductor, thereby affecting the B-phase current. However, this coupling is achieved indirectly through the magnetic field and is weaker than the direct electro-thermal coupling in the same phase. A moderate value, such as 0.5, is assigned to it to indicate that there is a significant but indirect coupling. 3) Assign a weak coupling value between voltage and current, for example... Its physical principle is the generalized form of Ohm's law. Voltage is the driving force for generating current. However, in high-voltage cable systems, the current is greatly affected by the load, and the system impedance is not purely resistive, but includes inductive reactance and capacitive reactance. Therefore, at a certain moment, the instantaneous control relationship between voltage and current is weakened by the complex system conditions, and is given a relatively weak value, such as 0.2, indicating that there is a basic physical connection between them, but the instantaneous correlation is not strong. The resulting fused feature representation is no longer a bunch of isolated data points, but features with richer information containing physical constraints. When an anomaly, such as partial discharge, occurs, it will disrupt this normal physical coupling relationship, for example, causing a mismatch between temperature changes and current changes, which increases the reconstruction error and makes it easier to detect. Based on a physical perception attention mechanism, feature modulation of a multidimensional input matrix is ​​achieved through a physical relation matrix, specifically as follows: Using physical relationship matrix For multidimensional input matrices Explicit feature modulation is performed to calculate the feature representation that incorporates physical constraints. The feature modulation is achieved through the following formula: in, for An identity matrix of dimension 1 For the modulated output Feature representation of dimensional fusion is a non-negative modulation coefficient used to control the injection intensity based on physical priors. At that time, the feature representation degenerates into a multidimensional input matrix. At that time, through The defined physical relationships are explicitly incorporated into the feature representation; It reflects the theoretical, interrelated state value that each physical quantity should exhibit after considering the mutual influence between physical quantities under the given physical laws. It is no longer an isolated observation value, but a context-enhanced feature that contains the inherent physical correlation of the system. Each element in This indicates the characteristic value of the i-th physical quantity at the t-th sampling time point. This value consists of two parts: the first part is the observed value of the physical quantity itself. The second part is based on the physical relationship matrix. The combined influence of all other physical quantity observations on this physical quantity; The absolute numerical value itself has no independent physical meaning, because it is determined by the multidimensional input matrix. Physical relationship The derived features are calculated together; their magnitude depends on the original data. In comparison, if A certain value is significantly greater than The corresponding original observations, which represent the values ​​based on the physical model. The state of other physical quantities strongly indicates that this physical quantity should be at a higher level, and vice versa; when the system is operating normally, and The difference (i.e., the basis of the reconstruction error) should be small. If an anomaly occurs, such as partial discharge, this normal physical coupling relationship is disrupted, leading to... The calculated theoretical value and The observed values ​​deviated significantly, thus being detected. This is the basic data; if all observations... If both are zero, then no matter how strong the physical relationship is, the output will be zero. It must also be zero. Provides the cardinality of features; physical relation matrix. The larger the element in the matrix, the larger the corresponding output feature. The larger the portion affected by it, the greater the impact. Modulation coefficient The larger the output feature Physical relationship The larger the incremental part introduced, the better. It is a global amplification factor that controls the volume of prior physical knowledge in the final features; when hour, Physical relationships have no effect whatsoever; when hour, The value is directly multiplied by Incremental Features , The larger the increment, the more The greater the proportion of physical constraints, the stronger their effect on shaping the characteristic form. It is the master switch for regulating the influence of physical a priori principles.

[0019] Step 3: Input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. Enhanced feature representation obtained through physical modulation The data is fed into a neural network model specifically designed for processing time series data for analysis. The time-series autoencoder plays the role of a normal pattern learner. Its task is not to classify or predict, but to learn the time-series variation patterns and relationship patterns between features that the multimodal data of high-voltage cables should have under healthy and fault-free conditions. An encoder is a neural network model specifically designed to learn the temporal patterns of multimodal data under normal conditions. Its core idea is to enable the model to master the data characteristics and patterns under healthy conditions through the process of compression and reconstruction. That is, through the compression-reconstruction process, the neural network learns the standard pattern when a high-voltage cable is operating normally, and any state that deviates from this pattern will be identified as abnormal. LSTM is a variant of recurrent neural networks. It has a gating mechanism (input gate, forget gate, output gate) that allows it to selectively remember important long-term information and forget irrelevant information. This makes it particularly good at learning long-term dependencies in time series. For example, if an overheating phenomenon in a cable is caused by a sudden increase in load a few minutes ago, LSTM can capture this causal relationship across time points. The encoder is responsible for learning the compressed representation of the input sequence. The encoder processes the input time series... For example, a sequence containing 9 physical quantities and 100 time points is progressively compressed, and temporal features are extracted using a two-layer LSTM network: the first LSTM layer contains 64 hidden units, receives the original input, and outputs 64-dimensional features; the second LSTM layer contains 32 hidden units, further compressing the features; finally, the encoder compresses it into a 16-dimensional, fixed-length context vector. This vector can be understood as a summary or latent representation of the data features within the entire time window, containing global information of the entire 100-time-point sequence. The decoder is responsible for reconstructing the original sequence from the compressed latent representation. Starting with the context vector, the decoder receives the 16-dimensional latent vector output by the encoder and generates the sequence through two layers of inverse LSTM networks: the first LSTM layer contains 32 hidden units, and the second LSTM layer contains 64 hidden units. Finally, a fully connected layer maps the 64-dimensional features back to the original input. Sequences with identical dimensions Ideally, Should with Almost identical; The gating mechanism of LSTM plays a key role in this process: the input gate determines which new information needs to be stored, the forget gate controls which historical information needs to be retained, and the output gate adjusts the impact of the current state on the next time step. This mechanism enables the network to effectively capture cross-time point causal relationships such as the increase in temperature delay after the current increases. The decoder adopts a structure symmetrical to the encoder, also containing two layers of LSTM. The initial state of the decoder is set to the context vector generated by the encoder, and the sequence is reconstructed step by step. At each time step, the decoder outputs a 9-dimensional vector, corresponding to the reconstructed values ​​of 9 physical quantities at that time point. Through 100 consecutive time steps, the complete reconstructed sequence R is finally generated. During training, a teacher-forced strategy is adopted, so that the decoder uses the true value of the previous time step as input at each step, which helps to accelerate the convergence of the model. In the inference phase, the decoder uses its own output from the previous time step as the current input to achieve a fully autoregressive reconstruction process. The autoencoder is trained entirely on historical normal data and adopts an unsupervised learning approach. The goal is to minimize the reconstruction error. After training, the model parameters are fixed and deployed to the online monitoring system. This design ensures that the model only has a clear understanding of normal conditions and remains sensitive to anomalies. Any state that deviates from the normal pattern will lead to an increase in reconstruction error, thereby triggering an alarm mechanism. The fused feature representation The data is input to a time-series autoencoder trained offline. Both the encoder and decoder are composed of multi-layer long short-term memory network units. The training process of the time-series autoencoder is carried out offline, and the training data used are all historical data of high-voltage cables under normal insulation conditions. Offline training shows that before deployment, the model is trained using a large amount of historical normal data, and its parameters are fixed. During online monitoring, the model no longer learns but only infers, which ensures the stability of the model. The most critical point is that the model has never seen abnormal data such as partial discharge. It has only learned what normal should look like. Therefore, when an anomaly occurs, the model's reconstruction ability will be significantly reduced due to its unfamiliarity with this pattern, resulting in huge errors. This avoids the difficulty of collecting difficult-to-obtain abnormal samples for supervised learning. The fused feature representation The input is fed to a temporal autoencoder, which maps the input sequence to a latent spatial feature representation through its encoder, and then reconstructs a reconstructed sequence with the same dimension as the input sequence from the latent spatial feature representation through its decoder. In order to learn the time dependency patterns of data under normal conditions; The patterns that the model needs to learn include: the normal variation patterns of each physical quantity itself, such as the normal fluctuation range of current with load changes; the cooperative relationship of different physical quantities over time, such as how the temperature will rise with a delay after the current increases; and the overall morphological characteristics of the sequence. Computational fusion of feature representations With reconstruction sequence The mean square error between them is taken as the reconstruction error, and its reconstruction error The calculation formula is: in, Feature representation for fusion The value of the nth physical quantity at the tth sampling time point. To reconstruct the values ​​of the corresponding physical quantities in the sequence at the corresponding time points. For physical quantities, Index of sampling time points; It reflects the autoencoder's ability to reconstruct the input data at the t-th sampling time point, and it quantifies the model output. Compared to real input The degree of difference between them; Under normal conditions, autoencoders excel at reconstruction, therefore and Very close. It will remain at a low level, and in abnormal conditions, such as partial discharge, the input data will be processed. The pattern deviates from the normal pattern learned by the model during training. The model cannot understand this new pattern, thus producing a reconstructed sequence. Will be with the real This results in a large deviation, leading to Significantly increased; therefore It directly indicates the degree of deviation of the observed data at the t-th sampling time point from the normal baseline, and is an instantaneous measure of anomaly; The larger the value, the higher the probability of an anomaly and the stronger the anomaly; a huge reconstruction error means that the current data pattern is extremely abnormal and cannot be understood or reproduced by the model at all, which strongly suggests that the system operating state has changed significantly, such as the insulating material is undergoing violent partial discharge. The value of is the average of the sum of the squared errors of all physical quantities at the t-th sampling time point, therefore With each They are all positively correlated; this is a "bucket effect," where a significant deviation in the reconstruction of any one physical quantity leads to an increase in the sum, thus... Increase It is sensitive to anomalies in any physical quantity; if only one physical quantity, such as the temperature of phase A, has a large reconstruction error, while other physical quantities are reconstructed well, It will also increase, but the magnitude will be limited. If multiple physical quantities, such as phase A current, phase B temperature, and voltage, simultaneously exhibit large reconstruction errors, The value will be significantly increased, indicating that the anomaly has affected multiple system parameters and is more severe. The formula uses square terms. ,so It is particularly sensitive to large absolute errors, and squaring has an amplifying effect; an error of magnitude 2 contributes 4, while an error of magnitude 4 contributes 16, which is 4 times the former, even though the absolute error is only 2 times; this nonlinear relationship makes... It is better at capturing significant anomalies that deviate far from the normal pattern, and relatively insensitive to small, random noise fluctuations. This meets the needs of anomaly detection, namely, focusing on significant anomalies and ignoring small perturbations. Reconstruction error It is an indicator that measures confusion or anomaly, and this indicator provides the most direct input signal for subsequent time series accumulation and threshold judgment.

[0020] Step 4: Using the exponential decay weighted recursive algorithm, the reconstruction error is accumulated over time to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with the dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output. Reconstruction error is instantaneous and contains noise. Using it directly for judgment will cause the system to be oversensitive and generate a large number of false alarms. The core purpose of this step is to smooth the data and extract trends. Instantaneous high errors are random noise or transient interference, while real insulation faults (such as partial discharge) usually occur continuously or intermittently. Through time-series accumulation, transient interference can be suppressed, while continuous abnormal signals can be amplified to improve the reliability of detection. The normal operating status of a cable can drift slowly due to changes in ambient temperature and load, which in turn changes the normal fluctuation range of the comprehensive anomaly score. Fixed thresholds cannot adapt to this change, and will either generate false alarms when the load is high or miss alarms when the load is low. Dynamic thresholds can adjust with changes in the normal state of the system and always maintain the accuracy of the judgment. An exponentially decaying weighted recursive algorithm is used to accumulate the reconstruction error over time to obtain a comprehensive anomaly score, which smooths random fluctuations and highlights persistent anomalies. The calculation formula is as follows: in, For the first Comprehensive anomaly score for each sampling time point, For the first Comprehensive anomaly score for each sampling time point, This is a decay factor used to control the weighting of the current reconstruction error when updating the comprehensive anomaly score; It does not reflect a state at a moment, but rather a weighted average trend of recent anomalies. It includes the anomaly information of the current moment, and also includes the anomaly information of all moments in history in a decaying manner. It indicates the persistence and severity of the abnormal state, a brief period of high... It will only cause A small, brief bulge, and a sustained high This will lead to This accumulation and amplification creates a stable platform at a high level, therefore It is better at indicating persistent faults than transient disturbances; The higher the value, the greater the likelihood of a persistent anomaly in the near future, and the higher the overall energy level of the anomaly. This means that the system has been consistently or frequently deviating from its normal pattern over a recent period, which is more significant than a single, transient high error. This further proves that a real system failure has occurred, rather than random noise; The value is positively correlated with The value, It is a term in the formula, representing the instantaneous error at the current moment if other conditions remain unchanged. The larger it is, the better it is for comprehensive anomaly analysis. The greater the direct contribution, the more direct the impact. This represents historical memory. It already contains Historical information, therefore Inheriting the trend of historical anomalies, if the system has been running abnormally for some time ( (very high), then even if currently Slightly decreased It will remain at a high level, which provides unusual inertia; The larger the value (closer to 1), the more the formula approximates. This means It is almost entirely determined by the current error, has a very short historical memory (forgetfulness), and the system responds quickly to anomalies, but is also easily affected by transient noise (high sensitivity, poor stability). The smaller (closer to 0), the more the formula approximates. This means The changes are very slow, with a long historical memory (stubbornness). The system responds slowly to anomalies and can effectively smooth out transient noise, but it will only be confirmed when the anomaly persists for a long enough time (high stability, low sensitivity). It is a knob that balances detection sensitivity and anti-interference capability, and needs to be adjusted according to the specific application scenario; Table 1 shows the specific data of some data numbers and the comprehensive anomaly score at the current time.

[0021] Table 1 Through data analysis, it can be observed that there is a clear temporal cumulative relationship between the comprehensive anomaly score and the reconstruction error. As the sample number increases, the reconstruction error shows a fluctuating upward trend, while the comprehensive anomaly score shows a more gradual but continuous growth trend. This indicates that the system has smoothed out the instantaneous abnormal signal and pays more attention to the persistence characteristics of the abnormal state. When analyzing the relationship between the overall anomaly score and historical states, the overall anomaly score at the current moment is always significantly affected by the state at the previous moment. When the reconstruction error fluctuates greatly, the change in the overall anomaly score is relatively mild, which reflects the system's memory and accumulation effect on anomaly signals. For example, the reconstruction error of sample number 5 is 0.25, and the corresponding overall anomaly score increases from 0.10 to 0.13; while when the reconstruction error of sample number 8 reaches 0.32, the overall anomaly score increases from 0.13 to 0.18, showing that the system's attention to persistent anomalies is gradually increasing. The variation of the attenuation factor also reflects the adaptive characteristics of the system's detection strategy. When the abnormal signal is weak, the attenuation factor remains at a small value, and the system pays more attention to the continuity of historical states. When the abnormal signal is significantly enhanced, the attenuation factor increases accordingly, and the system will improve its response speed to the current abnormality. This dynamic adjustment mechanism enables the system to maintain its sensitivity to sudden abnormalities while effectively suppressing misjudgments caused by transient interference.

[0022] Figure 2 The data number represents a sequence of consecutive sampling points arranged in chronological order. Each data number corresponds to a specific sampling time. The increment of the number reflects the progress of time. Therefore, the data number is essentially a discretized time coordinate.

[0023] Reconstruction error represents the degree of deviation between the original features and the reconstructed features at the current moment, and is an instantaneous indicator for anomaly detection. The attenuation factor, as a weight adjustment parameter, controls the proportion of the current observation and historical memory in the comprehensive evaluation. When the reconstruction error spikes, the attenuation factor will be adjusted accordingly to balance the weight of instantaneous anomalies and historical trends. Finally, a comprehensive anomaly score that is stable but reflects the true anomaly trend is generated through a recursive formula. This design ensures that the detection system is sensitive to real faults and can effectively suppress false alarms caused by random interference. It is worth noting that the growth trend of the comprehensive anomaly score is more robust than that of the reconstruction error. When the reconstruction error declines, the comprehensive anomaly score can still be maintained at a certain level due to the historical accumulation effect. This characteristic helps to identify true persistent faults and avoid overreacting to transient fluctuations. Through this time-series accumulation mechanism, the system achieves a more accurate and reliable assessment of the cable insulation status.

[0024] Dynamic threshold based on time window The statistical characteristics of the internal comprehensive anomaly scores are calculated for each sampling time point within the window. All are based on a length of [missing information] as the endpoint. Data Window The corresponding comprehensive anomaly score is calculated internally, thereby obtaining the time window. A complete historical anomaly sequence; Based on the statistical characteristics of this historical comprehensive anomaly sequence, a dynamic threshold is calculated. Its formula is: in, and Time windows The mean and standard deviation of all composite outlier scores. This is the sensitivity coefficient; The comprehensive anomaly score at the current moment is compared with the dynamic threshold corresponding to that moment. If the comprehensive anomaly score is greater than the dynamic threshold, an instantaneous anomaly flag is generated. If the instantaneous anomaly flag appears continuously for M times, the insulation status of the target high-voltage cable is finally determined to be abnormal, and the discharge identification result and warning signal are output. It reflects the upper limit of the normal fluctuation range of the comprehensive anomaly score when the system was in normal operation in the recent past. It indicates how much of the comprehensive anomaly score can still be considered normal random fluctuation under the current operating conditions. Any comprehensive anomaly score that exceeds this threshold is considered a potential anomaly that deviates significantly from the normal baseline. The larger the value, the greater the fluctuation in the system's recent normal operating status within the window period. Therefore, a higher threshold is needed to determine the abnormality. For example, when the cable is under heavy load, the normal fluctuation range of its current and temperature will be greater than that under light load, resulting in a higher normal value for the comprehensive abnormality score. At this time, the dynamic threshold will automatically increase to avoid misjudging normal fluctuations under heavy load as abnormal. Represents a time window The historical average level of all comprehensive anomaly scores within the system. If the system is operating normally at a high level of comprehensive anomaly score recently, for example due to continuous high load, then the baseline of the dynamic threshold will also be raised accordingly. This ensures that the threshold can adapt to the slow drift of the system operating conditions. Represents a time window The dispersion of all comprehensive anomaly scores, i.e., volatility, is considered. If the recent comprehensive anomaly scores fluctuate greatly, for example due to frequent load changes, it means that the boundary of the normal state is wider. Therefore, the dynamic threshold should leave a wider margin based on the mean. (larger), to avoid false alarms; It is an adjustable safety factor that determines how lenient the threshold is relative to the mean. Smaller values ​​(e.g., 2) have stricter thresholds, are close to the mean, and are easy to detect weak anomalies, but the false alarm rate increases. A larger threshold (e.g., 3) results in a more lenient threshold, is farther from the mean, makes the system more stable, and has a lower false alarm rate, but may miss some early or weak abnormal signals. It is another knob for balancing false alarm rate and false negative rate; Even after smoothing, a single moment's overall anomaly score exceeding the dynamic threshold may still originate from noise that has not been fully filtered. The requirement that the threshold be exceeded M times consecutively greatly reduces the possibility of false alarms and ensures that alarms are triggered by stable, continuous fault states rather than transient disturbances. Parameter M is the last line of defense for system robustness. The output discharge identification results contain richer diagnostic information, such as: anomaly level: based on the magnitude of the comprehensive anomaly score exceeding the dynamic threshold, it is divided into different levels such as early warning, alarm, and critical; occurrence time: the timestamp of the start of the anomaly duration; core characteristics: the main physical quantities that cause the anomaly, such as which phase current or temperature anomaly is most significant. Outputting warning signals is the interface that triggers external actions. Warning signals can drive various operations: local alarms, which alert operators on the monitoring interface by flashing red or popping up dialog boxes; remote notifications, which notify relevant maintenance engineers via SMS, email, audible and visual alarms; and log recording, which records event details to a database for post-event analysis and tracing.

[0025] Please see Figure 6 The present invention also provides a multi-modal feature fusion and identification device for partial discharge of high-voltage cables. The device is used to execute the above-described multi-modal feature fusion and identification method for partial discharge of high-voltage cables, comprising: The data acquisition module is used to acquire multi-modal time-series data of the target high-voltage cable equipment. The multi-modal time-series data includes three-phase current, three-phase conductor temperature and voltage to ground. The multi-modal time-series data is preprocessed and time-aligned to obtain a preprocessed multi-dimensional input matrix. The physical construction module is used to construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, and to use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, outputting a fused feature representation. The feature fusion module is used to input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. The result judgment module is used to accumulate the reconstruction error over time using an exponential decay weighted recursive algorithm to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with a dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output.

[0026] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0027] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0028] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0029] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-modal feature fusion and identification method for partial discharge in high-voltage cables, characterized in that, The specific steps include: Step 1: Collect multi-mode time series data of the target high-voltage cable equipment. The multi-mode time series data includes three-phase current, three-phase conductor temperature and voltage to ground. Preprocess and time-align the multi-mode time series data to obtain the preprocessed multi-dimensional input matrix. Step 2: Construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, and output the fused feature representation; Step 3: Input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. Step 4: Using the exponential decay weighted recursive algorithm, the reconstruction error is accumulated over time to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with the dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output. Multimodal time series data includes three-phase current, three-phase conductor temperature, and voltage to ground. The logic for preprocessing and time alignment of the multimodal time series data is as follows: Three-phase current refers to the current flowing in the three-phase conductors A, B, and C. Current transformers are used to collect the current values ​​flowing in the three-phase conductors A, B, and C in real time to reflect the load condition of the line. Three-phase conductor core temperature refers to the temperature measurement value of the corresponding three-phase conductor core. This temperature measurement is directly obtained by fiber optic sensors embedded near the cable core. Voltage to ground refers to the voltage of each of the three-phase conductors A, B, and C relative to ground, measured by voltage transformers. This allows for the collection of nine physical quantities—three-phase current, three-phase conductor core temperature, and voltage to ground—forming a single physical quantity set. The multimodal time series data is based on the current time. As the endpoint, backtrack a length of The time window formed Within, at fixed sampling intervals The collected three-phase current time series, three-phase conductor temperature time series, and voltage-to-ground time series; An outlier detection method based on statistical interquartile range (IIR) is employed to identify and remove data points in multimodal time-series data that exceed 1.5 times the IIR of the upper quartile or fall below 1.5 times the IIR of the lower quartile. For missing values ​​caused by data removal and transmission interruptions, linear interpolation is used to calculate the imputation value by establishing a linear relationship between the two nearest valid data points before and after the missing value. A min-max scaling method is used to linearly transform the values ​​of each physical quantity to [the appropriate scale]. interval; Time alignment with uniform sampling time intervals Using this as a baseline, a corresponding timestamp sequence is generated for all modalities. For data not sampled at the baseline time point, linear interpolation is used to align it to a unified timestamp sequence, ultimately forming a sequence with dimension [missing information]. Multidimensional input matrix ,in This represents the total number of physical quantities, with a value of 9. For time window The total number of sampling time points within, the first Each sampling time point represents the current moment. The logic for constructing a predefined physical relationship matrix based on the physical characteristics of high-voltage cables is as follows: Physical Relationship Matrix It is a non-negative symmetric matrix predefined based on Ohm's law, the principle of thermal effect, and the principle of electromagnetic induction for high-voltage cables. The values ​​of the matrix elements are determined by quantifying the coupling strength between physical quantities according to the following principles: The diagonal elements are all 1, representing complete coupling between each physical quantity and itself. The off-diagonal elements are assigned values ​​based on the actual coupling strength between the physical quantities; the stronger the coupling, the larger the element value. The range of element values ​​is... , where 0 indicates no direct physical connection and 1 indicates complete coupling; The specific coupling strength is assigned based on Ohm's law, the principle of thermal effect and the principle of electromagnetic induction. Strong coupling value is assigned between current and temperature in the same phase, medium coupling value is assigned between currents in different phases, and weak coupling value is assigned between voltage and current. Based on a physical perception attention mechanism, feature modulation of a multidimensional input matrix is ​​achieved through a physical relation matrix, specifically as follows: Using physical relationship matrix For multidimensional input matrices Explicit feature modulation is performed to calculate the feature representation that incorporates physical constraints. The feature modulation is achieved through the following formula: in, for An identity matrix of dimension 1 For the modulated output Feature representation of dimensional fusion is a non-negative modulation coefficient used to control the injection intensity based on physical priors. It degenerates into a multidimensional input matrix. At that time, through The defined physical relationships are explicitly incorporated into the feature representation.

2. The multi-modal feature fusion and identification method for partial discharge of high-voltage cables according to claim 1, characterized in that: The fused feature representation The data is input to a time-series autoencoder trained offline. Both the encoder and decoder are composed of multi-layer long short-term memory network units. The training process of the time-series autoencoder is carried out offline, and the training data used are all historical data of high-voltage cables under normal insulation conditions. The fused feature representation The input is fed to a temporal autoencoder, which maps the input sequence to a latent spatial feature representation through its encoder, and then reconstructs a reconstructed sequence with the same dimension as the input sequence from the latent spatial feature representation through its decoder. In order to learn the time dependency patterns of data under normal conditions; Computational fusion of feature representations With reconstruction sequence The mean square error between them is taken as the reconstruction error, and its reconstruction error The calculation formula is: in, Feature representation for fusion The value of the nth physical quantity at the tth sampling time point. To reconstruct the values ​​of the corresponding physical quantities in the sequence at the corresponding time points. For physical quantities, This is the index for the sampling time point.

3. The multi-modal feature fusion and identification method for partial discharge of high-voltage cables according to claim 2, characterized in that: An exponentially decaying weighted recursive algorithm is used to accumulate the reconstruction error over time to obtain a comprehensive anomaly score, which smooths random fluctuations and highlights persistent anomalies. The calculation formula is as follows: in, For the first Comprehensive anomaly score for each sampling time point, For the first Comprehensive anomaly score for each sampling time point, This is a decay factor used to control the weighting of the current reconstruction error when updating the comprehensive anomaly score.

4. The multi-modal feature fusion and identification method for partial discharge of high-voltage cables according to claim 3, characterized in that: Dynamic threshold based on time window The statistical characteristics of the internal comprehensive anomaly scores are calculated for each sampling time point within the window. All are based on having it as the endpoint and a length of Data window The corresponding comprehensive anomaly score is calculated internally, thereby obtaining the time window. A complete historical anomaly sequence; Based on the statistical characteristics of this historical comprehensive anomaly sequence, a dynamic threshold is calculated. Its formula is: in, and Time windows The mean and standard deviation of all composite outlier scores. This is the sensitivity coefficient; The comprehensive anomaly score at the current moment is compared with the dynamic threshold corresponding to that moment. If the comprehensive anomaly score is greater than the dynamic threshold, an instantaneous anomaly flag is generated. If the instantaneous anomaly flag appears continuously for M times, the insulation state of the target high-voltage cable is finally determined to be abnormal, and the discharge identification result and warning signal are output.

5. A multi-modal feature fusion and identification device for partial discharge of high-voltage cables, characterized in that: The device is used to perform a multi-modal feature fusion identification method for partial discharge of high-voltage cables as described in any one of claims 1-4, including: The data acquisition module is used to acquire multi-modal time-series data of the target high-voltage cable equipment. The multi-modal time-series data includes three-phase current, three-phase conductor temperature and voltage to ground. The multi-modal time-series data is preprocessed and time-aligned to obtain a preprocessed multi-dimensional input matrix. The physical construction module is used to construct a predefined physical relationship matrix based on the physical characteristics of high-voltage cables, and to use the physical relationship matrix to perform feature modulation on the multidimensional input matrix, outputting a fused feature representation. The feature fusion module is used to input the fused feature representation into the temporal autoencoder, which is composed of long short-term memory network units to learn its temporal dependencies and perform feature reconstruction, output the reconstruction sequence, and calculate the mean square error between the fused feature representation and the reconstruction sequence as the reconstruction error. The result judgment module is used to accumulate the reconstruction error over time using an exponential decay weighted recursive algorithm to obtain a comprehensive anomaly score. The comprehensive anomaly score is compared with a dynamic threshold. If the comprehensive anomaly score is greater than the dynamic threshold, the insulation state of the target high-voltage cable is determined to be abnormal, and the discharge identification result is output.

Images