A power communication network fault prediction method, system, device and medium

Abstract

The application discloses a power communication network fault prediction method, system, device and medium, relates to the technical field of power communication network fault prediction, and comprises the following steps: generating a fault evolution characteristic vector according to the deterioration direction characteristics of equipment; performing synchronous block processing to obtain original data patches and patch evolution state vectors; generating basic patch encoding, modulating the basic patch encoding; establishing a time series prediction model, dynamically adjusting the historical visible range of local attention, performing mask attention calculation on the modulated basic patch encoding, determining the fusion weight of each monitoring channel according to the local attention coding characteristics, and obtaining patch-level fusion features; outputting fault category probability and fault advance according to the patch-level fusion features, and obtaining power communication network fault prediction results. In the application, the patch represents the degradation direction information, and the attention range and channel weight are dynamically adjusted according to the evolution state, so that the matching degree of the prediction result and the gradual degradation process of the equipment is improved.

Description

Technical Field

[0001] This invention relates to the field of power communication network fault prediction technology, specifically to a power communication network fault prediction method, system, equipment, and medium. Background Technology

[0002] The power communication network is the core carrier network for power system dispatching, protection and control services. Its equipment failures usually involve a gradual degradation process, such as optical power attenuation and bit error rate increase.

[0003] Existing fault prediction methods typically use deep learning models to perform time-series patching on communication network monitoring data and combine local attention and channel attention to extract time-series features to achieve fault prediction.

[0004] However, existing time-series patch modeling methods only describe the numerical form of patches and fail to distinguish the consistency of degradation directions across different monitored quantities. This forces the model to relearn domain patterns from a limited number of fault samples, making it difficult to reflect the directional characteristics of gradual equipment degradation. Furthermore, the fixed local attention mechanism used to reduce computational overhead employs a uniform neighborhood range. Setting the neighborhood range too wide increases computational redundancy in stable segments, while setting it too narrow truncates the slow decay process, making it difficult to consider the differentiated historical information requirements of the stable period and the gradual failure period. Moreover, existing local attention channel fusion weights are typically set in a fixed manner without dynamically adjusting them based on the patch's evolution, resulting in insufficient matching between the predicted results and the actual gradual degradation process of the equipment. Summary of the Invention

[0005] In view of the above-mentioned existing problems, the present invention provides a fault prediction method, system, device and medium for power communication networks, which solves the problems in the prior art where patch representation does not distinguish the consistency of the degradation direction of monitoring quantities, and the historical visibility range of local attention and channel fusion weight are not dynamically adjusted in conjunction with the patch evolution state, resulting in insufficient matching degree between fault prediction results and actual gradual degradation process of equipment.

[0006] To address the aforementioned technical problems, a fault prediction method for power communication networks is proposed, including: A basic time-series matrix is ​​generated based on the acquired multi-monitoring channel data of the target power communication network, and a fault evolution feature vector is generated according to the characteristics of equipment deterioration direction. Synchronous block processing is performed on the basic time-series matrix and the fault evolution feature vector to obtain original data patches and patch evolution state vectors. A basic patch code is generated based on the original data patches, and the basic patch code is modulated using the patch evolution state vector to obtain the modulated basic patch code. A time-series prediction model is established, and the historical visibility range of local attention is dynamically adjusted according to the patch evolution state vector. Masked attention calculation is performed on the modulated basic patch code to obtain local attention coding features. The fusion weights of each monitoring channel are determined based on the local attention coding features, and a weighted summation is performed on the local attention coding features to obtain patch-level fusion features. The fault category probability and fault lead time are output based on the patch-level fusion features to obtain the fault prediction results of the power communication network.

[0007] As a preferred embodiment of the power communication network fault prediction method described in this invention, the obtained modulated basic patch code includes, Perform linear mapping and positional encoding on the original data patch to obtain the basic patch code; Generate a modulation vector based on the patch evolution state vector; The basic patch code is modulated according to the modulation vector to obtain the modulated basic patch code.

[0008] As a preferred embodiment of the power communication network fault prediction method described in this invention, the obtained local attention coding features include: The overall evolutionary feature intensity is determined based on the patch evolutionary state vector, and a fixed local neighborhood is set for the current patch; An intensity judgment is performed on the overall evolutionary feature intensity, and an evolutionary neighborhood is generated based on the intensity judgment result.

[0009] The fixed local neighborhood is merged with the evolved neighborhood to obtain a set of visible locations. Mask attention calculation is then performed on the modulated base patch code according to the set of visible locations to obtain the local attention coding features.

[0010] The beneficial effects of this preferred technical solution are as follows: This invention determines whether the current patch is in a progressive failure stage by the intensity of the overall evolutionary features. In the stable stage where the intensity does not exceed the threshold, attention calculation is performed only in a fixed local neighborhood to avoid redundant computational overhead. When the intensity exceeds the threshold, historical patches are selected to form an evolutionary neighborhood based on the consistency of the degradation direction and the similarity of the patch's evolutionary state vector. This expands the visible range of attention to historical periods that carry the same degradation trend, taking into account both computational efficiency and the ability to capture long-range degradation features.

[0011] As a preferred embodiment of the power communication network fault prediction method described in this invention, the obtained patch-level fusion features include: Pooling and activation processing are performed on the local attention-encoded features to obtain the basic weights corresponding to each monitoring channel; The channel evolution priors corresponding to each monitoring channel are determined based on the patch evolution state vector; The basic weights are corrected using channel evolution priors to obtain the fusion weights of each monitoring channel; The local attention encoding features of each monitoring channel are weighted and summed according to the fusion weights to obtain patch-level fusion features.

[0012] The beneficial effects of this preferred technical solution are as follows: The present invention introduces channel evolution prior to perform correction on the basic weights, so that the fusion weights of each monitoring channel not only reflect the hidden layer distribution of local attention coding features, but also combine the degree of degradation contribution of the channel in the evolution dimension, avoiding the channel weights being pulled by the instantaneous amplitude fluctuations caused by environmental disturbances, thereby preserving the expression of cross-channel fault evolution relationships such as continuous decrease in optical power and synchronous increase in bit error rate in the fusion features.

[0013] As a preferred embodiment of the power communication network fault prediction method described in this invention, the step of outputting fault category probabilities and fault lead times based on patch-level fusion features to obtain power communication network fault prediction results includes: Global pooling is performed on the patch-level fused features to obtain observation window-level features; Establish a fault classification head and an advance prediction head. Use the fault classification head to perform classification mapping processing on the observation window-level features and output the fault category probability. The lead prediction head is used to perform regression mapping on the observation window-level features to output the fault lead. The fault risk score of the current observation window is determined based on the fault category probability. If the subsequent observation window is closer to the fault occurrence time and the overall evolution characteristic intensity is not lower than that of the previous observation window, then the fault risk score of the adjacent observation window is subject to monotonicity constraint processing.

[0014] As a preferred embodiment of the power communication network fault prediction method described in this invention, the generation of the fault evolution feature vector includes, The directional coefficient is determined based on the direction of equipment deterioration in each monitoring channel, and the health baseline value for each monitoring channel is determined based on the health history interval. The directional degradation amount is obtained by mapping the basic time series matrix with the directional coefficient and the health baseline value. The variable acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount are determined based on the directional degradation amount. The directional degradation amount, variable acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount are then concatenated to obtain the fault evolution feature vector.

[0015] As a preferred embodiment of the power communication network fault prediction method described in this invention, the step of obtaining the original data patch and the patch evolution state vector includes, The base time series matrix and fault evolution feature vector are processed in blocks using the same patch length and patch step size to obtain the original data patch and the evolution feature patch; For the directional degradation amount, change acceleration, continuous unidirectional degradation duration and cross-channel coordination amount within the same evolutionary feature patch, perform patch aggregation processing to obtain the aggregation vector; The aggregated vector is projected onto the state space to obtain the patch evolution state vector.

[0016] As a preferred embodiment of the power communication network fault prediction system of the present invention, it is characterized by including a feature vector generation module, a synchronization block module, a modulation module, an attention calculation module, a fusion module, and an output module. The feature vector generation module is used to generate a basic time series matrix based on the acquired target power communication network multi-monitoring channel data, and to generate a fault evolution feature vector according to the characteristics of equipment deterioration direction. The synchronous block module is used to perform synchronous block processing on the basic time series matrix and the fault evolution feature vector to obtain the original data patch and the patch evolution state vector. The modulation module is used to generate a basic patch code based on the original data patch, and modulate the basic patch code using the patch evolution state vector to obtain the modulated basic patch code. The attention calculation module is used to establish a temporal prediction model, dynamically adjust the historical visibility range of local attention according to the patch evolution state vector, perform mask attention calculation on the modulated basic patch code, and obtain local attention coding features. The fusion module is used to determine the fusion weight of each monitoring channel based on the local attention coding features, and to perform a weighted summation on the local attention coding features to obtain patch-level fusion features; The output module is used to output the fault category probability and fault advance based on the patch-level fusion features to obtain the fault prediction result of the power communication network.

[0017] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a fault prediction method for a power communication network.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a fault prediction method for a power communication network.

[0019] The beneficial effects of this invention are as follows: This invention introduces the characteristics of equipment degradation direction into the generation process of fault evolution feature vector, so that the patch represents the degradation direction of different monitoring quantities rather than just reflecting the numerical form; the patch evolution state vector is passed between the stages of block division, encoding, attention calculation and channel fusion, driving the model structure to adaptively adjust with the degree of degradation, reducing the dependence on limited fault samples; the combined output of the classification head and the lead prediction head, and the application of monotonicity constraints to the fault risk score of the continuous observation window, make the prediction results show a stable and interpretable trend. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The above is a flowchart of a power communication network fault prediction method provided in one embodiment of the present invention.

[0022] Figure 2 The flowchart illustrates the generation of fault evolution feature vectors in a fault prediction method for power communication networks, as provided in one embodiment of the present invention.

[0023] Figure 3 The flowchart illustrates the output local attention coding feature of a fault prediction method for power communication networks, as provided in one embodiment of the present invention. Detailed Implementation

[0024] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0025] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting faults in a power communication network is provided, comprising: S1: Generate a basic time series matrix based on the acquired target power communication network multi-monitoring channel data, and generate a fault evolution feature vector according to the equipment deterioration direction characteristics.

[0026] S2: Perform synchronous block processing on the basic time series matrix and fault evolution feature vector to obtain the original data patch and the patch evolution state vector.

[0027] S3: Generate a basic patch code based on the original data patch, and modulate the basic patch code using the patch evolution state vector to obtain the modulated basic patch code.

[0028] S4: Establish a time-series prediction model, dynamically adjust the historical visibility range of local attention based on the patch evolution state vector, perform mask attention calculation on the modulated basic patch encoding, and obtain local attention encoding features.

[0029] S5: Determine the fusion weight of each monitoring channel based on the local attention coding features, and perform weighted summation on the local attention coding features to obtain patch-level fusion features.

[0030] S6: Output the fault category probability and fault advance based on the patch-level fusion features to obtain the fault prediction results of the power communication network.

[0031] It should be noted that this invention dynamically adjusts the historical visibility range of local attention by patching the evolution state vector, converging the neighborhood in stable sections and expanding the neighborhood in progressively faulty sections, balancing computational overhead and long-range degradation capture capability; it determines the fusion weight of each monitoring channel by the evolution state, suppressing the interference of instantaneous disturbances of a single channel on the channel weight; and it combines the monotonicity constraints of adjacent observation windows to jointly output the fault category and lead, thereby improving the matching degree between the fault prediction results and the actual progressive degradation process of the equipment.

[0032] Example 2, refer to Figure 1 - Figure 2 This is a second embodiment of the present invention, which provides a method for predicting faults in a power communication network, including: In step S1, generating the basic time series matrix based on the acquired target power communication network multi-monitoring channel data includes the following steps: Data from multiple monitoring channels in the target power communication network originates from the integrated network management system, equipment performance management platform, and alarm logs, and is associated according to equipment identifiers, link identifiers, and unified timestamps. When the original sampling periods of different monitoring quantities are inconsistent, each monitoring quantity is resampled to a unified sampling interval. And complete the alignment according to the same timestamp. Fault Category The time when the actual failure occurred Obtained from confirmed alarm records, and saved separately as monitoring information.

[0033] After completing association, resampling, and time alignment, with the inclusion The original multivariate time series matrix is ​​constructed using observation windows at each sampling time point as units. The expression is: in, Indicates the first At the sampling time, the first The original detection values ​​of each detection channel The number of sampling points within the observation window. In this embodiment, to monitor the number of channels, , These correspond to received optical power, link bit error rate, transmission delay, equipment operating temperature, and packet loss rate, respectively. This represents a real matrix with L rows and C columns.

[0034] Furthermore, the original multivariate time series matrix is ​​preprocessed. For isolated missing values ​​within a single channel, linear interpolation is performed using two adjacent valid sampling points. If the missing time... Located at the valid time and Between, and The interpolation calculation expression is: in, For the first Each channel is missing at the moment The completion value, The most recent valid sampling time before the missing time. It is the most recent valid sampling time after the missing time. For the first Each channel at time Valid monitoring values, For the first Each channel at time Valid monitoring values.

[0035] After completing the missing value completion, the first... Each channel calculates its mean based on historical training data. and standard deviation When a certain monitoring value meets When an outlier is detected, it is identified as an isolated outlier and replaced using the linear interpolation method described above.

[0036] After outlier replacement, the Min-Max method is used for normalization to obtain the basic time series matrix. The expression is: in, For the first The minimum value of each channel in the training history data. These are the normalized monitoring values. For the first The maximum value of each channel in the training history data. To prevent extremely small positive numbers with a denominator of zero, the normalization parameter is calculated only from training historical data and remains fixed during the validation and inference phases.

[0037] Furthermore, in step S1, generating the fault evolution feature vector based on the equipment deterioration direction characteristics includes steps S11~S13: S11: Determine the directional coefficient based on the direction of equipment deterioration in each monitoring channel, and determine the health baseline value for each monitoring channel based on the health history interval.

[0038] To address the inconsistency in the deterioration direction of different indicators, a directional coefficient matrix is ​​introduced, expressed as: in, This indicates a deterioration in optical power. This indicates a deterioration in the bit error rate. This indicates a deterioration in latency. This indicates a deterioration in temperature. This indicates a worsening packet loss rate. .

[0039] No. The steady-state health baseline value for each channel is calculated from the same or similar equipment within a healthy historical period characterized by no faults, no maintenance, and no valid alarms. The expression is as follows: in, Indicates the first The healthy steady-state baseline value of each channel For the first A collection of health history sampling points from each channel. The number of sampling points in the current set is used. In this embodiment, the data that meets the health conditions within the last 3 natural days is used as the baseline statistical interval, and the baseline value is updated every 24 hours. If there is no complete interval that meets the health conditions within the last 3 natural days, the baseline value of the previous update cycle is used.

[0040] S12: The directional degradation amount is obtained by mapping the basic time series matrix using the directional coefficient and the health baseline value.

[0041] The expression for directional degradation is: in, For the first The monitoring channel is at the first The directional degradation at each sampling time. For the first The direction coefficients of each channel, after the current mapping, The larger the value, the more significant the degradation relative to a healthy baseline.

[0042] S13: Determine the changing acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount based on the directional degradation amount, and concatenate the directional degradation amount, changing acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount to obtain the fault evolution feature vector.

[0043] The overall process of generating fault evolution feature vectors is as follows: Figure 2 As shown, it specifically includes: Based on the directional degradation, a change acceleration, continuous unidirectional degradation length, and cross-channel coordination are constructed for each time step and each channel, and together with the directional degradation, they form a multi-dimensional evolutionary feature matrix.

[0044] The expression for the changing acceleration is: in, For the first Each channel at time Discrete-changing acceleration, This represents the directional degradation of the current channel at the previous sampling time. This represents the directional degradation of the current channel during the first two sampling times.

[0045] The expression for the length of continuous degradation in the same direction is: in, Indicates the first Channel cutoff time The number of sampling points that continuously show degradation and enhancement. Indicates the first Channel cutoff time The number of sampling points that continuously show degradation and enhancement. Indicates dissatisfaction The situation.

[0046] To describe the synchronization degradation relationship between different monitoring channels, in a length of Calculate the Pearson correlation coefficient within a sliding window for each channel. With another channel ,definition: in, Indicates channel With channel In recent Pearson correlation coefficient at each sampling point To remove Other channel indexes, Indicates channel The mean, Indicates channel The mean, Indicates channel standard deviation Indicates channel standard deviation For the index of sampling points within the window, For channel c in time period The amount of targeted degradation, For channel During the period The amount of targeted degradation.

[0047] Based on correlation, for the first Each channel defines a cross-channel coordination quantity, expressed as follows: in, Indicates the first The first channel and the remaining channels have the highest degree of synchronous degradation among the cross-channel coordination quantities. For the first The monitoring channel is at the first The directional degradation at each sampling time. For the first The monitoring channel is at the first The amount of directional degradation at each sampling time.

[0048] By concatenating the four types of evolutionary variables according to their channels, we obtain the evolutionary feature vector, expressed as: in, For a moment The evolutionary feature vectors, and Concat(·) represents vector concatenation.

[0049] In step S2, synchronous block processing is performed on the basic time series matrix and fault evolution feature vector to obtain the original data patch and patch evolution state vector, including steps S21~S23: S21: Perform block processing on the basic time series matrix and fault evolution feature vector using the same patch length and patch step size to obtain the original data patch and evolution feature patch.

[0050] The basic time series matrix obtained in step S1 and multidimensional evolutionary feature matrix Use the same patch length and step length Perform synchronous block division.

[0051] Let the first The starting sampling point for each patch is Then the original data patch and evolutionary feature patches They are defined as follows: in, For patch index, For the first The patch is located at the starting sampling point in the original observation window. For patch length, The step size between adjacent patches. The total number of patches, due to and Use the same , and The two correspond strictly in time. This represents a real matrix with P rows and C columns.

[0052] Patch overlap rate is defined as This embodiment adopts ,therefore To disclose the method for determining patch length and step size, a candidate set is constructed during the training phase. and for each Pick Calculate the Macro-F1 and mean absolute error of lead on the validation set. The expression is: in, The number of validation samples with real fault time labels. For the first The true failure lead time for each sample For the first The lead time for prediction of a sample For sample index.

[0053] When selecting parameters, first eliminate combinations whose average inference latency is greater than the uniform sampling interval. Then, select the combination with the highest Macro-F1 among the remaining combinations. If the Macro-F1 values ​​are the same, select the combination with the highest Macro-F1 value. Smaller combinations.

[0054] S22: Perform intra-patch aggregation processing on the directional degradation amount, change acceleration, continuous unidirectional degradation duration and cross-channel coordination amount within the same evolutionary feature patch to obtain the aggregation vector.

[0055] For the i-th evolutionary feature patch Intra-channel aggregation is performed, and the mean directional degradation, maximum acceleration, mean length of continuous unidirectional degradation, and mean cross-channel coordination are calculated separately within each patch. These are then concatenated to form an aggregation vector. The expression is: in, Let i be the directional degradation amount of the i-th patch. The acceleration of the change for the i-th patch, Let be the consecutive unidirectional degradation length of the i-th patch. For the cross-channel coordination quantity of the i-th patch, This indicates that the mean is calculated along the patch time dimension. This indicates finding the maximum value along the patch time dimension.

[0056] S23: Project the aggregated vector onto the state space to obtain the patch evolution state vector.

[0057] Will Linear projection onto a D-dimensional state space yields the patch evolution state vector. The expression is: in, It is a linear projection matrix. Let be the bias vector, and D be the dimension of the patch evolution state.

[0058] In step S3, the basic patch code is modulated using the patch evolution state vector to obtain the modulated basic patch code, including steps S31 to S33: S31: Perform linear mapping and positional encoding on the original data patch to obtain the basic patch encoding.

[0059] For the i-th original data patch Let the P-dimensional patch vector of the c-th physical channel be denoted as . The basic patch code is obtained using shared linear mapping and positional encoding. The expression is: in, Encode the base patch for the i-th patch and the c-th channel. To share linear mappings, To share the bias parameters corresponding to the linear mapping, Encode the location of the i-th patch.

[0060] S32: Generate a modulation vector based on the patch evolution state vector.

[0061] S33: Modulate the basic patch code according to the modulation vector to obtain the modulated basic patch code.

[0062] The modulation processing in this invention is residual modulation, expressed as: in, Encoding the modulated patch. Basic patch coding, For the trainable weight matrix used to generate the modulation vector, This is a trainable bias vector. For activation function, This indicates element-wise multiplication.

[0063] It should be noted that the current processing, while preserving the original time series form, injects the degradation background of the current patch into the coded representation, so that normal short-term fluctuations of the same amplitude and continuous degradation produce different model responses.

[0064] In step S4, the historical visibility range of local attention is dynamically adjusted according to the patch evolution state vector, and mask attention calculation is performed on the modulated basic patch code to obtain local attention coding features, including steps S41~S43: It should be noted that the temporal prediction model is an improved PatchTST model with evolutionary feature coupling, consisting of a patch embedding layer, an evolutionary feature modulation layer, a local attention coding layer, and a channel attention fusion layer. PatchTST serves as the main body of long temporal coding, and the model obtained in step S2 is... Used for modulation patch representation in step S3, control history visibility range in step S4, and correction channel weights in step S5.

[0065] The overall process of outputting local attention encoded features is as follows: Figure 3 As shown, it specifically includes: S41: Determine the overall evolutionary feature strength based on the patch evolutionary state vector, and set a fixed local neighborhood for the current patch.

[0066] Furthermore, the overall evolutionary characteristic strength is calculated. The expression is: in, The trainable weight matrix used to calculate the overall evolutionary feature strength. is the trainable bias vector corresponding to the trainable weight matrix.

[0067] While fixed local attention can reduce computation during the stable phase, it may truncate the slow evolutionary process across multiple patches during the gradual degradation phase. Therefore, it depends on the overall evolutionary feature strength. The system dynamically determines whether to extend the historical visibility of the current patch.

[0068] Define the fixed local neighborhood of the i-th patch. The expression is: in, For historical patch index, To fix the number of patches contained in a local neighborhood, this embodiment takes... Based on experience, the settings are configured so that during the stable phase, only the current patch and the three most recent historical patches are accessed.

[0069] S42: Perform an intensity judgment on the overall evolutionary feature intensity, and generate an evolutionary neighborhood based on the intensity judgment result.

[0070] The specific operations for strength determination in this invention include, specifically, setting an evolutionary feature strength threshold to determine whether the current patch is in a state of continuous degradation. and requires continuity One patch satisfies In this embodiment, This is set based on experience. If the conditions are met, it only applies to a fixed local neighborhood of the current patch. Internal computation of masked attention; if the condition is not met, then constructing an evolutionary neighborhood from earlier historical patches. .

[0071] For any candidate historical patch First, define the patch-level degradation increment vector based on the directional degradation amount. The expression is: in, Indicates the first The time average of the directional degradation of each monitoring channel within each patch. Indicates the first The time average of the directional degradation of each monitoring channel within each patch.

[0072] when At that time, determine the patch With historical patches The overall trend of degradation is consistent. For the first The patch-level degradation increment vector of each historical patch.

[0073] The degree of state similarity is measured by the cosine similarity of the patch's evolved state vectors, expressed as: in, Let i be the state similarity between the i-th and j-th patches. It is a 2-norm. for The transpose of , For the first The patch evolution state vector of each patch.

[0074] when The state similarity is considered to meet the requirements.

[0075] In satisfying , and In the historical patches, according to Select no more than K patches from high to low to form an evolutionary neighborhood, expressed as: in, Represents the evolutionary neighborhood. This indicates that the top K historical patches are selected based on state similarity, where K is the maximum number of patches in the evolutionary neighborhood.

[0076] S43: Merge the fixed local neighborhood with the evolved neighborhood to obtain a set of visible locations, and perform mask attention calculation on the modulated base patch code according to the set of visible locations to obtain the local attention coding features.

[0077] The final set of visible locations for the current patch is defined as follows: in, The set of patch indices that are allowed to be accessed when performing self-attention for the i-th patch.

[0078] when When the conditions for continuous degradation are not met, ,therefore When the condition for continuous degradation is met, Add a small number of earlier historical patches with similar states to the fixed local neighborhood.

[0079] The evolutionary feature strength threshold, state similarity threshold, and maximum number of evolutionary neighborhoods are determined using the training and validation sets. The specific steps are as follows: Construct based on patch distribution quantiles of training samples The candidate set is constructed based on the value range {0.6, 0.7, 0.8, 0.9}. The candidate set is constructed based on the value range {1,2,4,8}, and the candidate set of K is constructed while satisfying that the average inference delay does not exceed the uniform sampling interval. In the combination, the combination with the largest validation set Macro-F1 is selected first. If they are the same, the combination with the lower false positive rate is selected. If they are still the same, the combination with the smaller K is selected. In this embodiment, K is selected. .

[0080] For the i-th patch, it will belong to The location is set to visible, not belonging to The position is set to invisible.

[0081] Mask matrix The corresponding element is 0 in the visible position and negative infinity in the invisible position.

[0082] The expression for calculating self-attention is: in, This represents the mask self-attention computation function. The query vector is obtained by linearly mapping the current patch encoding. Set of visible locations The key matrix obtained by mapping the encoding of each patch within the matrix. Let be the transpose vector of the key matrix. Set of visible locations The value matrix obtained by mapping the encoding of each patch within the matrix. The dimension of the key vector. It is a mask matrix jointly controlled by local mask and evolutionary neighborhood.

[0083] In step S5, the fusion weights of each monitoring channel are determined based on the local attention coding features, and a weighted summation is performed on the local attention coding features to obtain the patch-level fusion features, including steps S51~S54: S51: Perform pooling and activation processing on the local attention encoding features to obtain the basic weights corresponding to each monitoring channel.

[0084] After local attention encoding is completed, the i-th patch and the c-th physical channel obtain encoded features. Different failure modes involve different key monitoring channels, and channel weights generated solely based on transient latent features are easily affected by single-channel spike noise. Therefore, the channel evolution prior obtained in step two is introduced for joint correction on the basis of data-driven channel attention.

[0085] The SENet-style channel attention mechanism extracts data to drive the basic weights for each channel's features, expressed as follows: in, The data-driven base weights for the i-th patch and the c-th channel are... This indicates that pooling is performed on the channel-encoded features. This is the first layer of trainable weight matrix. This is the trainable weight matrix for the second layer. It is a linear rectification activation function.

[0086] S52: Determine the channel evolution prior for each monitoring channel based on the patch evolution state vector.

[0087] The channel evolution prior vector is calculated based on the patch evolution state vector, and the expression is: in, Indicates the first The channel evolution prior vector of each patch This is a trainable weight matrix used to compute the channel evolution prior from the patch evolution state vector. To and The corresponding trainable bias vector S53: Perform correction processing on the basic weights using channel evolution priors to obtain the fusion weights of each monitoring channel.

[0088] The correction process in this invention is a multiplicative correction, specifically including extracting the evolutionary feature components of each channel from the patch evolutionary feature state to obtain the final fusion weight, expressed as: in, Let be the final fusion weight for the i-th patch and the c-th channel. This represents the total number of monitoring channels. For the first The evolutionary prior components of each channel.

[0089] S54: Perform weighted summation on the local attention coding features of each monitoring channel according to the fusion weight to obtain patch-level fusion features.

[0090] The coded features of each channel are weighted according to the final fusion weight to obtain the patch-level fusion features. The expression is: When the optical power channel continues to decay and the bit error rate channel increases synchronously, the cross-channel collaborative quantity, after aggregation, will simultaneously improve the overall evolution feature intensity and the evolution feature components of the corresponding channel, open the historical evolution neighborhood and increase the fusion weight of the relevant faulty channel. If only a single channel short-term spike pulse occurs, the overall evolution feature intensity remains low and no enhancement is triggered.

[0091] In step S6, the fault category probability and fault advance amount are output based on the patch-level fusion features to obtain the fault prediction result of the power communication network, including steps S61~S64: S61: Perform global pooling on the patch-level fused features to obtain observation window-level features.

[0092] The N patch-level fusion features obtained in step S5 Perform time-dimensional global pooling to obtain observation window-level features. The expression is: in, This indicates that average pooling is performed along the patch time dimension. For the first The patch-level fusion feature is obtained by fusing the patches through channel attention.

[0093] S62: Establish a fault classification head and an advance prediction head. Use the fault classification head to perform classification mapping processing on the observation window-level features and output the fault category probability.

[0094] The fault classification head outputs a category probability vector based on observation window-level features. The expression is: in, For the trainable weight matrix of the fault classification head, This is the trainable bias vector for the fault classification head.

[0095] , Corresponding to normal state Corresponding link failure, Corresponding to equipment hardware failure, Corresponding signal interference fault, corresponding The probability of power supply failure.

[0096] S63: Use the lead prediction head to perform regression mapping processing on the observation window-level features and output the fault lead.

[0097] Lead prediction head based on the same window level features Direct output of non-negative continuous fault lead time The expression is: in, For the trainable weight matrix of the lead prediction head, This is the trainable bias vector for the lead prediction head. The meaning is the remaining time from the end of the current observation window to the predicted time of the fault occurrence.

[0098] For the A training sample with real fault times, the end time of the observation window is denoted as . The actual time of the fault occurrence is recorded as .

[0099] The expression for the actual failure advance is: in, For training sample index, This represents the actual failure advance for the m-th sample.

[0100] The classification task uses cross-entropy loss. The lead time task only uses mean squared error loss on M samples with actual failure times. The expression is: in, The number of samples in the training batch. The total number of categories, For the m-th sample to belong to the th sample The true one-hot label of the class, Here, M represents the probability of the corresponding category output by the classification header, and M is the number of samples with valid real fault times in the current batch. and These are the predicted lead time and the actual lead time for the m-th time-labeled sample, respectively.

[0101] S64: Determine the fault risk score of the current observation window based on the fault category probability. If the subsequent observation window is closer to the fault occurrence time and the overall evolution characteristic intensity is not lower than that of the previous observation window, then perform monotonicity constraint processing on the fault risk scores of adjacent observation windows.

[0102] For the same equipment, a continuous observation window sequence formed in chronological order The fault classification head outputs a category probability vector. The sum of failure probabilities excluding the normal category is defined as the failure risk score, expressed as: in, Indicates the fault risk score. For continuous observation window index, The number of consecutive windows participating in the consistency constraint. For the first The normal category probability of each window is such that if the subsequent window is closer to the actual fault and the overall evolutionary feature strength is not lower than that of the previous window, then the fault risk score must not decrease.

[0103] Define an indicator variable with the following expression: in, For the indicator variable in the k-th adjacent window pair, For the first Fault lead time corresponding to each observation window For the first Fault lead time corresponding to each observation window For the first The overall evolutionary characteristic intensity of each observation window For the first The overall evolutionary characteristic intensity of each observation window.

[0104] The monotonic regularization loss is constructed as follows: in, This represents the monotonicity regularization loss. For window Fault risk score, For window Fault risk score, This represents the number of consecutive windows participating in the consistency constraint.

[0105] The overall training objective of the model is expressed as: in, This represents the overall training objective of the model. To avoid losing weight due to advance timing, These are the monotonicity constraint weights.

[0106] This embodiment determines the lead loss weight and monotonicity constraint weight through a validation set grid search, specifically including: Choose from the value range {0.5, 1.0, 2.0}. Choose from the range {0.01, 0.05, 0.10, 0.20}, prioritizing the Macro-F1 validation set. The second-best option is selected based on the principle of finding the best option. In this embodiment, we choose... .

[0107] It should be noted that the role of joint fault prediction is to simultaneously answer two questions: whether a fault will occur and how long until the fault occurs. The lead time label is directly calculated from historical fault times, avoiding the lag caused by retrospective conversion of time from classification probability thresholds; the monotonicity constraint suppresses risk inversion and violent oscillations between continuous windows, making the output warning level, fault type, and response time window more consistent and executable.

[0108] Example 3 is the third embodiment of the present invention. This embodiment provides a fault prediction system for power communication networks, including a feature vector generation module, a synchronization block module, a modulation module, an attention calculation module, a fusion module, and an output module. The feature vector generation module is used to generate a basic time series matrix based on the acquired target power communication network multi-monitoring channel data, and to generate a fault evolution feature vector according to the characteristics of equipment deterioration direction. The synchronous block module is used to perform synchronous block processing on the basic time series matrix and the fault evolution feature vector to obtain the original data patch and the patch evolution state vector. The modulation module is used to generate a basic patch code based on the original data patch, and modulate the basic patch code using the patch evolution state vector to obtain the modulated basic patch code. The attention calculation module is used to establish a temporal prediction model, dynamically adjust the historical visibility range of local attention according to the patch evolution state vector, perform mask attention calculation on the modulated basic patch code, and obtain local attention coding features. The fusion module is used to determine the fusion weight of each monitoring channel based on the local attention coding features, and to perform a weighted summation on the local attention coding features to obtain patch-level fusion features; The output module is used to output the fault category probability and fault advance based on the patch-level fusion features to obtain the fault prediction result of the power communication network.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0110] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0112] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0113] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method for predicting faults in a power communication network, characterized in that: include, A basic time series matrix is ​​generated based on the acquired multi-monitoring channel data of the target power communication network, and a fault evolution feature vector is generated according to the characteristics of equipment deterioration direction. Synchronous block processing is performed on the basic time series matrix and fault evolution feature vector to obtain the original data patch and the patch evolution state vector; A basic patch code is generated based on the original data patch, and the basic patch code is modulated using the patch evolution state vector to obtain the modulated basic patch code. A temporal prediction model is established, and the historical visibility range of local attention is dynamically adjusted according to the patch evolution state vector. Mask attention calculation is performed on the modulated basic patch encoding to obtain the local attention encoding features. The fusion weights of each monitoring channel are determined based on the local attention coding features, and a weighted summation is performed on the local attention coding features to obtain the patch-level fusion features; Based on the patch-level fusion features, the fault category probability and fault advance amount are output to obtain the fault prediction results of the power communication network.

2. The fault prediction method for power communication networks as described in claim 1, characterized in that: The obtained modulated base patch code includes Perform linear mapping and positional encoding on the original data patch to obtain the basic patch code; Generate a modulation vector based on the patch evolution state vector; The basic patch code is modulated according to the modulation vector to obtain the modulated basic patch code.

3. The fault prediction method for power communication networks as described in claim 2, characterized in that: The obtained local attention encoding features include, The overall evolutionary feature intensity is determined based on the patch evolutionary state vector, and a fixed local neighborhood is set for the current patch; An intensity judgment is performed on the overall evolutionary feature intensity, and an evolutionary neighborhood is generated based on the intensity judgment result; The fixed local neighborhood is merged with the evolved neighborhood to obtain a set of visible locations. Mask attention calculation is then performed on the modulated base patch code according to the set of visible locations to obtain the local attention coding features.

4. The fault prediction method for power communication networks as described in claim 3, characterized in that: The obtained patch-level fusion features include, Pooling and activation processing are performed on the local attention-encoded features to obtain the basic weights corresponding to each monitoring channel; The channel evolution priors corresponding to each monitoring channel are determined based on the patch evolution state vector; The basic weights are corrected using channel evolution priors to obtain the fusion weights of each monitoring channel; The local attention encoding features of each monitoring channel are weighted and summed according to the fusion weights to obtain patch-level fusion features.

5. The fault prediction method for power communication networks as described in claim 4, characterized in that: The step of outputting fault category probabilities and fault lead times based on patch-level fusion features to obtain fault prediction results for the power communication network includes: Global pooling is performed on the patch-level fused features to obtain observation window-level features; Establish a fault classification head and an advance prediction head. Use the fault classification head to perform classification mapping processing on the observation window-level features and output the fault category probability. The lead prediction head is used to perform regression mapping on the observation window-level features to output the fault lead. The fault risk score of the current observation window is determined based on the fault category probability. If the subsequent observation window is closer to the fault occurrence time and the overall evolution characteristic intensity is not lower than that of the previous observation window, then the fault risk score of the adjacent observation window is subject to monotonicity constraint processing.

6. The fault prediction method for power communication networks as described in claim 5, characterized in that: The generated fault evolution feature vector includes, The directional coefficient is determined based on the direction of equipment deterioration in each monitoring channel, and the health baseline value for each monitoring channel is determined based on the health history interval. The directional degradation amount is obtained by mapping the basic time series matrix with the directional coefficient and the health baseline value. The variable acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount are determined based on the directional degradation amount. The directional degradation amount, variable acceleration, continuous unidirectional degradation duration, and cross-channel coordination amount are then concatenated to obtain the fault evolution feature vector.

7. The fault prediction method for power communication networks as described in claim 6, characterized in that: The obtained original data patch and patch evolution state vector include, The base time series matrix and fault evolution feature vector are processed in blocks using the same patch length and patch step size to obtain the original data patch and the evolution feature patch; For the directional degradation amount, change acceleration, continuous unidirectional degradation duration and cross-channel coordination amount within the same evolutionary feature patch, perform patch aggregation processing to obtain the aggregation vector; The aggregated vector is projected onto the state space to obtain the patch evolution state vector.

8. A power communication network fault prediction system, employing the power communication network fault prediction method as described in any one of claims 1 to 7, characterized in that, It includes a feature vector generation module, a synchronization block segmentation module, a modulation module, an attention calculation module, a fusion module, and an output module; The feature vector generation module is used to generate a basic time series matrix based on the acquired target power communication network multi-monitoring channel data, and to generate a fault evolution feature vector according to the characteristics of equipment deterioration direction. The synchronous block module is used to perform synchronous block processing on the basic time series matrix and the fault evolution feature vector to obtain the original data patch and the patch evolution state vector. The modulation module is used to generate a basic patch code based on the original data patch, and modulate the basic patch code using the patch evolution state vector to obtain the modulated basic patch code. The attention calculation module is used to establish a temporal prediction model, dynamically adjust the historical visibility range of local attention according to the patch evolution state vector, perform mask attention calculation on the modulated basic patch code, and obtain local attention coding features. The fusion module is used to determine the fusion weight of each monitoring channel based on the local attention coding features, and to perform a weighted summation on the local attention coding features to obtain patch-level fusion features; The output module is used to output the fault category probability and fault advance based on the patch-level fusion features to obtain the fault prediction result of the power communication network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power communication network fault prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fault prediction method for a power communication network as described in any one of claims 1 to 7.

Images