Broadcasting and television network fault prediction and self-healing method based on AI

Abstract

The invention relates to the technical field of learning methods, in particular to an AI-based broadcast television network fault prediction and self-healing method, which comprises the following steps of: extracting a channel quality difference value through edge detection, recombining to form a change trend set, performing three-point windowing to execute differential analysis, identifying a confidence interval compression trend, and extracting a prediction and detection symbol relationship. And identifying an error reversal direction, and finally generating a repair trigger signal label set. According to the method, a dynamic trend structure of a prediction difference value is constructed, a label index is extracted and triggered in combination with symbol consistency and direction reversal features, aggregation and label binding of time sequence features are completed, a training sample set capable of being used for learning is generated, the association identification capability between the features is enhanced, the sensing precision of weak change signals is improved, and the early warning range is expanded. Abnormal misjudgment is reduced, a judgment basis based on trend evolution is formed, and reliable data support is provided for fault prediction and self-healing response.

Description

Technical Field

[0001] This invention relates to the field of learning methods technology, and in particular to an AI-based method for predicting and self-healing faults in broadcast networks. Background Technology

[0002] The field of learning methods technology mainly involves the research and implementation of algorithms, methods, and strategies for training artificial intelligence models, especially neural network models. Core aspects include model parameter adjustment, training data processing, loss function optimization, learning rate control, and overfitting prevention. This field encompasses various machine learning paradigms such as supervised learning, unsupervised learning, and reinforcement learning. It focuses on how to enable computational models to self-adjust using historical data or simulated environments, thereby improving their adaptability and predictive ability to new data. It is widely applied in multiple technical directions such as speech recognition, image recognition, natural language processing, intelligent control, and fault diagnosis, and is closely integrated with computer science, control engineering, statistics, and other fields, forming an interdisciplinary technical system. Traditional methods for fault prediction and self-healing in broadcast television networks refer to technical solutions that analyze potential fault phenomena during the operation of broadcast television network systems and take countermeasures to restore normal system operation after a fault occurs. These methods typically rely on static rule bases and expert experience for fault judgment, monitor network operation status through manually set threshold detection methods, and, after a fault occurs, rely on manual inspection or pre-set procedures to perform partial restarts, path switching, or equipment replacement to complete fault location and system recovery.

[0003] Existing fault handling methods in broadcast networks rely on fixed thresholds for status judgment, which cannot perceive the subtle fluctuations in channel quality. Fault identification is highly dependent on specific rule matching, making it difficult to dynamically reflect the state evolution under different operating scenarios. The processing flow is mainly based on manual inspection and static methods, lacking an early response mechanism for abnormal evolution trends. The predictive ability is limited to a simple mapping of known patterns, making it difficult to capture key signals in the early stages of fault evolution. It has failed to form a systematic assessment basis for abnormal states, resulting in a lag in response to changing network environments. Fault location and response measures lack precise data-driven guidance, and the system as a whole lacks a new feedback path for sustainable updates and adaptations, making it difficult to support the needs of intelligent and proactive network maintenance. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based method for predicting and self-healing faults in broadcast networks, comprising the following steps: S1: Collect channel quality detection data at continuous time points through edge detection in the broadcast television network, extract the difference between the predicted data and the upper and lower limits of the corresponding confidence interval, reorganize the difference in time order, and construct a set of predicted change trends; S2: Based on the predicted trend set, select the difference at every three time points to construct an analysis window, perform front-to-back difference on the difference within the window, determine whether the sign is continuous and consistent, and generate confidence interval compression trend features; S3: Based on the predicted trend set, extract the channel quality detection content corresponding to the time point of the predicted data, compare the two symbols, identify whether the symbols are reversed at adjacent time points, and construct the error reversal direction feature; S4: Based on the confidence interval compression trend characteristics and the error reversal direction characteristics, the feature states are synchronously compared according to the time index, and the index positions that simultaneously meet the contraction and reversal conditions are selected to generate a set of repair trigger signal labels. S5: Select the predicted data in the predicted change trend set and bind the labels in the repair trigger signal label set to construct input and output field pairs, divide the training and validation data, and construct the broadcast network fault prediction and self-healing learning sample structure.

[0005] As a further embodiment of the present invention, the predicted change trend set includes the difference change amplitude, the difference change direction, and the difference time series structure; the confidence interval compression trend feature includes the sign continuity identifier, the trend direction consistency identifier, and the difference stability feature; the error reversal direction feature includes the sign deviation feature, the direction switching point, and the reversal trend identifier; the repair trigger signal label set includes the synchronization index position, the trend intersection mark, and the trigger status label; and the broadcast network fault prediction and self-healing learning sample structure includes the predicted data input feature, the repair trigger output label, and the time series partitioning configuration.

[0006] As a further aspect of the present invention, the simultaneous satisfaction of the contraction and reversal conditions means that at the time index, the signs of the difference between the upper and lower limits of the confidence interval are continuously consistent, forming a contraction trend, while the sign of the prediction error at that moment is reversed compared to the previous moment.

[0007] As a further aspect of the present invention, the identification of whether the symbols at adjacent time points are reversed refers to determining the symbol by comparing the difference between the predicted value and the actual channel quality at two consecutive time points. If the symbol at the current time point is opposite to that at the previous time point, it is determined to be a symbol reversal.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect channel quality detection data frames at multiple consecutive time points in the edge detection of the broadcast television network, perform difference calculations between the predicted value and the corresponding upper and lower limits of the confidence interval for each time point, and organize the obtained differences in chronological order to generate a set of confidence interval difference sequences. S102: Call the time point differences in the confidence interval difference sequence set, and perform sequence reconstruction processing in chronological order by sliding window merging and time index alignment to obtain a continuous sequence matrix and generate the difference sequence reconstruction matrix. S103: Based on the numerical change trend of the time points in the reconstruction matrix according to the difference sequence, calculate the rate of change and pre-set the rate of change threshold to divide and mark the intervals. After aggregating the trend intervals, obtain the predicted change trend set.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the predicted trend set, select the difference content corresponding to every three consecutive time points to construct an analysis window, combine them into a difference triplet sequence in chronological order, index and organize the sequence, and generate a difference sliding window sequence group. S202: Call the difference sequence within the window of the difference sliding window sequence group, perform the front and back difference calculation between adjacent values ​​in sequence, construct the adjacent difference change sequence based on the difference results, and classify and aggregate them according to the time index to generate a continuous difference change sequence set; S203: Based on the set of continuous difference change sequences, retrieve the sign information of adjacent terms in the difference sequence, and determine whether there are continuous segments with consistent signs in the same sequence. If the preset continuous consistency condition is met, mark it as a stable trend segment. After aggregating all marked segments, obtain the confidence interval compression trend feature.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the predicted change trend set, extract the predicted data and channel quality detection data frame corresponding to each time point, and compare the numerical symbol attributes of the two types of data after pairing according to the time index. Integrate the comparison results in chronological order to generate a predicted detection symbol comparison sequence. S302: Call the predicted detection symbol comparison sequence to identify the difference between the symbols of the predicted data and the detected data in the time point comparison results. If the symbols are inconsistent, they are marked as direction difference items, and all differences are recorded in chronological order to generate a direction difference mark sequence. S303: Based on the directional difference marker sequence, compare the difference markers of adjacent time points one by one to identify whether the sign change has a reversal feature. If a reversal relationship is identified, extract the change segment and perform sequence aggregation processing to obtain the error reversal direction feature.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the confidence interval compression trend feature and the error reversal direction feature, extract the time index of the two types of features, construct a joint index lookup table according to the index order, organize the feature state values ​​at the corresponding positions, and generate an index feature mapping matrix; S402: Call the index feature mapping matrix, perform a logical comparison operation on the compression trend state value and the reversal direction state value of each index position, mark the index position when the combination of the two states meets the bidirectional consistency condition, aggregate all positions that meet the condition, and generate a consistent matching index set; S403: Based on the consistency matching index set, mark and encode all index positions in the set, assign a corresponding tag value to each index, construct a tag data frame structure, and obtain a repair trigger signal tag set.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the predicted change trend set, extract all predicted data as feature sources, retrieve the index position in the repair trigger signal label set, bind the predicted data and label value corresponding to the position to the field, establish the corresponding combination relationship, and generate a predicted label pairing structure set. S502: Call the predicted label pairing structure set, sort it according to the time index of the predicted data, divide the training time period and the verification time period according to the sequence order, and store the predicted values ​​and label pairs in the interval into independent structures to generate an interval partitioned data pair set. S503: Based on the structure of the training and validation segments in the interval-divided data set, construct the mapping relationship between the input feature value matrix and the label vector, and jointly encapsulate them into a unified format data unit to obtain the sample structure for broadcast network fault prediction and self-healing learning.

[0013] As a further aspect of the present invention, the joint encapsulation into a unified format data unit refers to combining the predicted feature matrix divided by time with the corresponding label vector in a one-to-one correspondence and organizing it into a unified input and output format that the model can recognize.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing a dynamic trend structure of the predicted difference, and combining the signature consistency and direction reversal feature extraction to trigger the label index, the aggregation of time-series features and label binding are completed, generating a training sample set that can be used for learning, enhancing the correlation recognition ability between features, improving the perception accuracy of weak change signals, expanding the warning range, reducing abnormal misjudgment, forming a judgment basis based on trend evolution, and providing reliable data support for fault prediction and self-healing response. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in 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.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] Please see Figure 1 This invention provides an AI-based method for predicting and self-healing faults in broadcast networks, comprising the following steps: S1: Collect channel quality detection content at multiple consecutive time points through edge detection in the broadcast television network, extract the difference between the prediction data at each time point and the upper and lower limits of the corresponding confidence interval in chronological order, perform time-series reconstruction processing on the difference content, and construct a set of predicted change trends; S2: Based on the predicted trend set, select the difference content of every three consecutive time points to form an analysis window, perform front and back difference processing on the difference in the window, identify whether there is a continuous difference sequence with the same sign, determine whether the continuous change direction is consistent, and generate confidence interval compression trend features. S3: Based on the predicted trend set, extract the channel quality detection content corresponding to the time point of the predicted data, compare the predicted data with the detection content to form a direction difference sequence, identify the symbol changes of adjacent time points in the sequence one by one, and construct the error reversal direction feature. S4: Based on the confidence interval compression trend feature and error reversal direction feature, the two types of feature content are synchronously matched according to the time index. The feature state combination under each index point is compared by bidirectional consistency logic. The index position that simultaneously satisfies the contraction trend and reversal trend is extracted, and a set of repair trigger signal labels is generated. S5: Select the predicted data in the predicted change trend set as the source of input features, bind the predicted data at the corresponding index position of the repair trigger signal label set with the associated label to establish a multi-dimensional pairing structure between input values ​​and output labels, and divide the training interval and validation interval according to the time order of data features to construct the sample structure for broadcast network fault prediction and self-healing.

[0020] The predicted trend set includes the difference change magnitude, difference change direction, and difference time series structure. The confidence interval compression trend features include sign continuity identifier, trend direction consistency identifier, and difference stability feature. The error reversal direction features include sign deviation feature, direction switching point, and reversal trend identifier. The repair trigger signal label set includes synchronization index position, trend intersection marker, and trigger status label. The cable TV network fault prediction and self-healing learning sample structure includes predicted data input features, repair trigger output labels, and time series partitioning configuration.

[0021] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect channel quality detection data frames at multiple consecutive time points in the edge detection of the broadcast television network, perform difference calculations between the predicted value and the corresponding upper and lower limits of the confidence interval for each time point, and organize the obtained differences in chronological order to generate a set of confidence interval difference sequences. Using edge monitoring probes in a fiber-coaxial hybrid network, physical layer parameters of the transmission channel are read in real time at a sampling frequency of 50 times per second. Specifically, the signal-to-noise ratio (SNR) and bit error rate (BER) values ​​are obtained as channel quality detection data frames. Simultaneously, channel quality data of the same type from the previous 72 hours are retrieved from the network's historical operation database. By calculating the mean and standard deviation of the historical data, a prediction interval with a 95% confidence level is constructed, thereby obtaining the predicted value, upper limit, and lower limit of the confidence interval for each sampling time point. Subsequently, numerical difference calculations are performed to calculate the difference between the predicted value and the upper limit of the confidence interval, and the difference between the predicted value and the lower limit of the confidence interval. These two sets of difference data are linearly arranged according to nanosecond-level timestamps to form upper limit difference subsequences and lower limit difference subsequences, respectively. These are then merged and stored to obtain a set of confidence interval difference sequences. In this process, taking the channel signal-to-noise ratio (SNR) data of a certain edge node as an example, assuming that at the time point of timestamp 1700000001, the predicted value collected is 35 dB, the upper limit of the confidence interval calculated based on historical data is 38 dB, and the lower limit of the confidence interval is 32 dB. Subtraction is performed: first, the difference between the predicted value of 35 dB and the upper limit of the confidence interval of 38 dB is calculated, i.e., 35 minus 38, resulting in an upper limit difference of -3 dB; second, the difference between the predicted value of 35 dB and the lower limit of the confidence interval of 32 dB is calculated, i.e., 35 minus 32, resulting in a lower limit difference of 3 dB. The advantage of this calculation logic is that by separating the deviation of the upper and lower limits, the offset vector of the current signal state relative to the safe range is accurately quantified. The calculated -3 dB and 3 dB are stored as feature values ​​at this time point in the sequence. Table 1 shows some of the collected and calculated data: Table 1. Calculation of Channel Quality Confidence Interval Difference

[0022] As shown in Table 1, through the above calculation process, the system can accurately capture the signal deviation within each time step. When the upper limit difference is positive or the lower limit difference is negative, it indicates that the signal has broken through the predicted safe envelope.

[0023] S102: Call the time point differences in the confidence interval difference sequence set, and perform sequence reconstruction processing in chronological order by sliding window merging and time index alignment to obtain a continuous sequence matrix and generate the difference sequence reconstruction matrix. The system reads all upper and lower bound difference data from the stored confidence interval difference sequence set. A sliding window with a length of 10 time steps is set, and data is truncated on the time axis with a step size of 1 time step at a time. For the data within each sliding window, the continuity of timestamps is checked. If a breakpoint with a timestamp interval exceeding 20 milliseconds is found, cubic spline interpolation is used, employing three valid data points before and after the breakpoint for fitting calculations to fill in the missing time point difference data, thus completing time index alignment. Subsequently, the aligned upper bound difference sequence is used as the first column vector, and the lower bound difference sequence as the second column vector, stacked row by row in ascending time order to construct a two-dimensional numerical matrix. The number of rows in this matrix corresponds to the total number of sampling points, and the number of columns is fixed at 2, thus obtaining a continuous difference sequence reconstruction matrix. In this process, it is assumed that there is a missing point within the time period currently covered by the sliding window, with adjacent upper bound differences of -1.7 dB and -6.3 dB, respectively. Interpolation is performed, using the slope of the change in data before and after to derive the value of the missing points. The calculated imputation value is assumed to be -4.0 dB. Inserting this imputation value into the corresponding position in the sequence ensures the integrity of the time axis. Next, the upper limit difference of -4.0 dB and the lower limit difference of 2.0 dB at the same time point are combined and filled into the corresponding row of the matrix. For example, for a data segment containing 1000 sampling points, the final generated matrix is ​​a 1000-row by 2-column floating-point matrix. This process transforms the discrete time series into a matrix form that can be used for linear algebra operations, providing a standardized data foundation for subsequent batch feature extraction.

[0024] S103: Based on the difference sequence, reconstruct the numerical change trend of time points in the matrix, calculate the rate of change, and pre-set the rate of change threshold to divide and mark the intervals. After aggregating the trend intervals, obtain the predicted trend set. The process iterates through each row of data in the difference sequence reconstruction matrix, extracting the upper limit difference data between the current time point and the previous time point, and performing a rate of change calculation operation. Specifically, it first calculates the numerical difference between the upper limit difference at the current time point and the upper limit difference at the previous time point, then divides this numerical difference by the absolute value of the upper limit difference at the previous time point to obtain the rate of change of the upper limit difference; similarly, it calculates the rate of change of the lower limit difference. Subsequently, a pre-set rate of change threshold of 0.15 (i.e., 15%) is introduced. This threshold is an empirical value derived by analyzing the average amplitude of signal fluctuations before a fault occurred by statistically analyzing network fault logs over the past year. The calculated real-time rate of change is compared with this threshold. If the absolute value of the rate of change is greater than 0.15, the current time interval is marked as "high volatility"; otherwise, it is marked as "stable". Finally, all continuous time intervals marked as "high volatility" are spliced ​​and aggregated, and isolated fluctuation points with a length of less than 3 time steps are removed to obtain the predicted trend set. In this process, it is assumed that the upper limit difference at the previous time point is -3.0 dB, and the upper limit difference at the current time point is -1.7 dB. The rate of change is calculated as follows: First, -1.7 is subtracted from -3.0, resulting in a difference of 1.3. Then, 1.3 is divided by the absolute value of -3.0, yielding approximately 0.433. This result of 0.433 is compared with a preset threshold of 0.15. Since 0.433 is greater than 0.15, the signal fluctuation at this time point is considered significant, falling into the category of drastic change. If the rate of change at the next time point is 0.05, it is considered stable. The advantage of this logic is that by calculating the rate of change using normalization, the influence of the signal's fundamental strength is eliminated, focusing solely on the acceleration characteristics of signal quality deterioration. This allows for the precise screening of potentially risky trend segments from massive amounts of data.

[0025] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the predicted trend set, select the difference content corresponding to every three consecutive time points to construct an analysis window, combine them into a difference triplet sequence in chronological order, index and organize the sequence, and generate a difference sliding window sequence group. Based on the high-fluctuation time periods marked in the predicted trend set, the corresponding time index range is locked, and the corresponding difference data is extracted from the difference sequence reconstruction matrix. An analysis window is constructed using every three consecutive time points as a basic analysis unit. For example, the upper and lower limit differences corresponding to time points T1, T2, and T3 are selected, and their values ​​are arranged sequentially to form a data group containing six numerical elements, defined as a difference triplet sequence. Subsequently, all generated difference triplet sequences are chained together in forward order along the time axis, and an integer index starting from 0 is established to ensure that each triplet can be accessed through a unique index value, ultimately generating a difference sliding window sequence group. In this process, it is assumed that the difference values ​​of the three consecutive time points selected are: upper limit difference of -3.0 at time T1, upper limit difference of -1.7 at time T2, and upper limit difference of -6.3 at time T3. These three values ​​are encapsulated into a unit, namely (-3.0, -1.7, -6.3). This operation expands instantaneous data into short time-series segments, enabling subsequent analysis to capture the local micro-oscillation patterns of the signal, rather than focusing solely on single-point values.

[0026] S202: Call the difference sequence within the window of the difference sliding window sequence group, perform the front and back difference calculations between adjacent values ​​in sequence, construct the adjacent difference change sequence based on the difference results, and classify and aggregate them according to the time index to generate a continuous difference change sequence set; Read each triplet sequence in the difference sliding window sequence group. For the three difference values ​​within the sequence, perform a first-order backward difference operation sequentially. Specifically, subtract the difference at the first time point from the difference at the second time point to obtain the first change; subtract the difference at the second time point from the difference at the third time point to obtain the second change. Combine the calculated first and second changes in their original time order to form a new binary sequence, i.e., the adjacent difference change sequence. Then, traverse the entire sequence group and classify and store all generated adjacent difference change sequences according to their original time indices to form a continuous difference change sequence set. In this process, continue using the aforementioned data: T1 is -3.0, T2 is -1.7, and T3 is -6.3. Perform the difference calculation: First, calculate -1.7 minus -3.0, i.e., -1.7 plus 3.0, resulting in 1.3; second, calculate -6.3 minus -1.7, i.e., -6.3 plus 1.7, resulting in -4.6. The final generated sequence of adjacent difference changes is (1.3, -4.6). This calculation reveals the "peak" characteristic of the signal, which rises sharply and then falls drastically within a short period. This operational logic directly quantifies the direction and amplitude of signal jitter by eliminating the absolute reference of the data, and can sensitively capture transient pulse interference in the network.

[0027] S203: Based on the set of continuous difference change sequences, retrieve the sign information of adjacent terms in the difference sequence, and determine whether there are continuous segments with consistent signs in the same sequence. If the preset continuous consistency condition is met, mark it as a stable trend segment. After aggregating all marked segments, obtain the confidence interval compression trend feature. The algorithm iterates through each binary sequence in the set of continuous difference change sequences, checking the mathematical signs (positive or negative) of the two change values ​​in the sequence. A logical judgment is then performed: if the first and second change values ​​in a sequence have the same sign (both positive or both negative), the trend within that small time window is considered monotonic. Further scanning of adjacent sequences is performed, counting the number of sequences with consistently consistent signs. If five or more consecutive sequences maintain the same sign characteristic (e.g., all positive or all negative), the signal is in a continuous unidirectional compression or expansion state, satisfying the preset continuous consistency condition. These consecutive time periods that meet the condition are marked as stable trend segments, and the start and end times of these segments are recorded, aggregating them to generate a confidence interval compression trend feature. In this process, it is assumed that the detected changes in five consecutive sequences are (0.2, 0.3), (0.4, 0.1), (0.3, 0.5), (0.2, 0.2), and (0.1, 0.4), and all values ​​are positive. This indicates that within the corresponding 15 time steps, the confidence interval difference is continuously increasing in one direction, and the signal quality is steadily approaching a certain boundary. The system marks this period as a "compression trend feature." Extracting this feature can effectively identify those "gradual" potential problems that, although not yet exceeding the boundary, are irreversibly heading towards failure.

[0028] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the set of predicted change trends, extract the predicted data and channel quality detection data frames corresponding to each time point, and compare the numerical symbol attributes of the two types of data after pairing according to the time index. Integrate the comparison results in chronological order to generate a predicted detection symbol comparison sequence. Based on the key time points determined in the predicted trend set, channel quality detection data frames (actual detection values) for each time point are extracted from the original acquisition database, and corresponding predicted data is extracted from the prediction model output library. For each pair of data, a symbol attribute extraction operation is performed to determine whether the actual value is greater than or less than zero (usually positive in the decibel system, referring to the deviation sign relative to the reference level). The symbols of the predicted data and the detected data are aligned by time and listed in the same row, forming a predicted-detection symbol comparison sequence with a three-column structure of "time-predicted symbol-detection symbol". In this process, assuming that at time point 1001, the predicted value is 35 dB (positive deviation relative to the reference), and the detected value is 31 dB (also positive deviation relative to the reference), then the record for this point is (1001, positive, positive). If at time point 1003, the predicted value is 31.5 dB, but the detected value suddenly changes to 28 dB (below the reference, recorded as a negative deviation), then the record is (1003, positive, negative). This step simplifies the complex numerical comparison into a directional comparison, which intuitively reflects the consistency between the predictive model and the actual physical link.

[0029] S302: Call the predicted detection symbol comparison sequence to identify the difference between the symbols of the predicted data and the detected data in the time point comparison results. If the symbols are inconsistent, they are marked as direction difference items, and all differences are recorded in chronological order to generate a direction difference label sequence. The system scans the predicted and detected symbols against the sequence line by line, performing an XOR comparison on the predicted and detected symbols in each line. If the predicted and detected symbols are different (e.g., one positive and one negative), a "directional deviation" is identified at that time point, and it is marked as a directional difference item. The absolute error value (the absolute difference between the predicted and detected values) is recorded. If the symbols are the same, the system skips the mark. All marked time points and their error magnitudes are linked chronologically to generate a directional difference mark sequence. In this process, for the data (positive and negative) at time point 1003, due to the inconsistent symbols, the system performs a mark operation and calculates the error magnitude: 31.5 minus the absolute value of 28, resulting in 3.5. This point is recorded as a directional difference item. If such differences occur frequently within a certain time period, it indicates a fundamental physical reversal of the channel characteristics (such as changes in reflection characteristics caused by loose fiber optic cable joints), rather than simple noise interference. This step, through qualitative screening, quickly identifies anomalies indicating model failure or sudden changes in network physical characteristics.

[0030] S303: Based on the direction difference label sequence, compare the difference sign labels of adjacent time points one by one to identify whether the sign change has a reversal feature. If a reversal relationship is identified, extract the change segment and perform sequence aggregation to obtain the error reversal direction feature. The system reads the direction difference marker sequence and compares the specific manifestation of the difference sign between two adjacent time points. The specific judgment logic is as follows: check whether the difference at the current time point is "predicted positive, actually negative," and whether the difference at the next adjacent time point is "predicted negative, actually positive," or vice versa. If a complete reversal of the difference sign is detected between adjacent time points, an "error reversal" phenomenon is determined. Data segments containing this reversal phenomenon at four consecutive time points are extracted and aggregated into an independent feature unit to obtain the error reversal direction feature. In this process, it is assumed that time point T1 is (predicted positive, detected negative) and time point T2 is (predicted negative, detected positive). This instantaneous polarity reversal usually indicates high-frequency phase jitter in the network or internal logic errors in the device. The system packages T1 and T2 and their neighboring points as a specific fault mode feature. This feature is highly sensitive for identifying faults such as synchronization loss or clock drift.

[0031] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the confidence interval compression trend feature and the error reversal direction feature, extract the time index of the two types of features, construct a joint index lookup table according to the index order, organize the feature state values ​​at the corresponding positions, and generate an index feature mapping matrix; Create a complete index table containing all covered time points. Set two columns for status: the first column corresponds to compression trend features, and the second column corresponds to error reversal features. Iterate through the index table; if a time point exists in the compression trend feature index list, set the first column to 1; otherwise, set it to 0. Similarly, if it exists in the error reversal feature index list, set the second column to 1; otherwise, set it to 0. This generates the index feature mapping matrix. During this process, construct the matrix structure shown in Table 2. Table 2 Example of Index Feature Mapping Matrix

[0032] As shown in Table 2, both status bits at time point 1002 are 1, indicating that at this moment there is both a unidirectional compression trend of the signal and an error reversal. This composite feature usually corresponds to a very high probability of the equipment being paralyzed.

[0033] S402: Call the index feature mapping matrix, perform a logical comparison operation between the compression trend state value and the reversal direction state value at each index position, mark the index position when the combination of the two states meets the bidirectional consistency condition, aggregate all positions that meet the condition, and generate a consistent matching index set; The system iterates through each row of the index feature mapping matrix, performing a logical AND operation on the compression trend state value and the reversal direction state value. The result is True only when both state values ​​are 1, and the corresponding time index is then designated as a "bidirectional consistency matching point." If either state value is 0, the result is False and the index is not marked. All time indices with True results are extracted and arranged in ascending order to generate a consistency matching index set. During this process, based on the data in Table 2, for time index 1001, the state is 1 and 0, and the logical AND result is 0; for time index 1002, the state is 1 and 1, and the logical AND result is 1; for time index 1003, the state is 0 and 1, and the logical AND result is 0. Therefore, the system only includes 1002 in the consistency matching index set. This logic, through rigorous double verification, filters out false alarms that might arise from a single feature (such as simple external interference or simple model bias), ensuring that the accuracy of the triggered repair signal reaches over 99%.

[0034] S403: Based on the consistency matching index set, mark and encode all index positions in the set, assign a corresponding tag value to each index, construct a tag data frame structure, and obtain the repair trigger signal tag set; Each index value in the consistency matching index set is read and converted into a label format using binary one-hot encoding. Specifically, for indices present in the set, a label value of "1" is assigned, representing "trigger repair"; for time indices not present in the set but within the original observation period, a label value of "0" is assigned, representing "normal maintenance". These label values ​​are then mapped one-to-one with their corresponding timestamps to construct a single-column label data frame, resulting in the repair trigger signal label set. In this process, time point 1002 is assigned a label value of 1. This means that in subsequent machine learning training, the model will learn that when the input data exhibits features characteristic of time point 1002, it should output a "trigger repair" instruction. This step completes the crucial transformation from unsupervised feature analysis to supervised label generation, providing a "standard answer" for building self-healing learning models.

[0035] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the set of predicted change trends, extract all predicted data as feature sources, retrieve the index positions in the set of repair trigger signal labels, bind the predicted data and label values ​​corresponding to the positions to the fields, establish corresponding combination relationships, and generate a set of predicted label pairing structures. Extract all raw predicted data from time periods marked as "high volatility" as feature input (FeatureX); simultaneously, read the set of repair trigger signal labels generated in step S403 as target labels (LabelY); using the timestamp as the key, perform an inner join operation on the database to bind the predicted data and label values ​​at the same time to the same record; if there is only predicted data but no corresponding label at a certain time (e.g., label missing), then discard that data; finally, establish a combination relationship between each row containing "predicted feature vector" and "repair trigger label", generating a predicted label pairing structure set. In this process, each generated record is in the form of: {Time: 1002, Feature: [36.5, -1.7, 4.3...], Label: 1}. This process ensures that each piece of data input to the model has a clear causal correspondence, that is, "under what data performance (feature), what decision (label) should be made".

[0036] S502: Call the predicted label pairing structure set, sort it according to the time index of the predicted data, divide the training time period and the validation time period according to the sequence order, and store the predicted values ​​and label pairs in the interval into independent structures to generate the interval partitioned data pair set. All records in the predicted label pairing structure set are sorted in ascending order by timestamp. The first 80% of the data records are truncated chronologically and assigned to the training dataset for model parameter learning. The remaining 20% ​​are assigned to the validation dataset to evaluate the model's generalization ability. In both the training and validation sets, predicted features are extracted and stored in a feature array, and label values ​​are extracted and stored in a label array, ensuring strict index correspondence between the two sets, generating a set of interval-partitioned data pairs. In this process, it is assumed that there are a total of 10,000 paired records. The system assigns records 1 to 8,000 to the training set and records 8,001 to 10,000 to the validation set. This time-based partitioning method (rather than random shuffling) strictly adheres to the principles of time series forecasting, avoiding the data leakage problem of "predicting the past with future data," and ensuring the real validity of the evaluation results for future practical applications.

[0037] S503: Based on the structure of the training and validation segments in the interval-divided data set, construct the mapping relationship between the input feature value matrix and the label vector, and jointly encapsulate them into a unified format data unit to obtain the sample structure for cable network fault prediction and self-healing learning. Based on the interval-divided data set, a tensor structure for the input of the deep learning model is constructed. Specifically, an input layer structure for a Long Short-Term Memory (LSTM) network is constructed, which includes three dimensions: batch size (BatchSize), time steps (TimeSteps, set to 10), and feature dimension (InputDim, set to 3, including predicted value, upper bound difference, and lower bound difference). Simultaneously, an output layer label vector is constructed, with its dimension being the batch size multiplied by 1 (binary classification result). The training and validation data are respectively filled into the above tensor structure and standardized (Z-score normalization), i.e., each feature value is subtracted from its mean and then divided by its standard deviation. Finally, it is packaged into a standard TFRecord or NumPy format file to obtain the sample structure for broadcast network fault prediction and self-healing learning. In this process, for feature normalization, it is assumed that the mean of a certain feature dimension is 0.5 and the standard deviation is 0.2. If the current input value is 0.9, the transformed value is (0.9-0.5) / 0.2=2.0. This processing eliminates the scale differences between data of different dimensions, accelerating the convergence speed of the model's gradient descent. The final generated data structure is shown in Table 3: Table 3. Example of self-healing learning sample structure

[0038] As shown in Table 3, this structure fully defines the information of the input neural network, enabling the model to directly read and start iterative training, and ultimately achieve early detection and automatic repair of network faults.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the technical solution.

Claims

1. An AI-based method for fault prediction and self-healing in broadcast networks, characterized in that, Includes the following steps: S1: Collect channel quality detection data at continuous time points through edge detection in the broadcast television network, extract the difference between the predicted data and the upper and lower limits of the corresponding confidence interval, reorganize the difference in time order, and construct a set of predicted change trends; S2: Based on the predicted trend set, select the difference at every three time points to construct an analysis window, perform front-to-back difference on the difference within the window, determine whether the sign is continuous and consistent, and generate confidence interval compression trend features; S3: Based on the predicted trend set, extract the channel quality detection content corresponding to the time point of the predicted data, compare the two symbols, identify whether the symbols are reversed at adjacent time points, and construct the error reversal direction feature; S4: Based on the confidence interval compression trend characteristics and the error reversal direction characteristics, the feature states are synchronously compared according to the time index, and the index positions that simultaneously meet the contraction and reversal conditions are selected to generate a set of repair trigger signal labels. S5: Select the predicted data in the predicted change trend set and bind the labels in the repair trigger signal label set to construct input and output field pairs, divide the training and validation data, and construct the broadcast network fault prediction and self-healing learning sample structure.

2. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The predicted trend set includes the difference change magnitude, the difference change direction, and the difference time series structure. The confidence interval compression trend features include sign continuity identifier, trend direction consistency identifier, and difference stability feature. The error reversal direction features include sign deviation feature, direction switching point, and reversal trend identifier. The repair trigger signal label set includes synchronization index position, trend intersection marker, and trigger status label. The broadcast network fault prediction and self-healing learning sample structure includes predicted data input features, repair trigger output labels, and time series partitioning configuration.

3. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The simultaneous satisfaction of the contraction and reversal conditions means that at the time index, the signs of the difference between the upper and lower limits of the confidence interval are continuously consistent, forming a contraction trend, while the sign of the prediction error at that moment is reversed compared to the previous moment.

4. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The identification of whether the symbol is reversed at adjacent time points refers to judging the symbol by the difference between the predicted value and the actual channel quality at two consecutive time points. If the symbol at the current time point is opposite to that at the previous time point, it is determined to be a symbol reversal.

5. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect channel quality detection data frames at multiple consecutive time points in the edge detection of the broadcast television network, perform difference calculations between the predicted value and the corresponding upper and lower limits of the confidence interval for each time point, and organize the obtained differences in chronological order to generate a set of confidence interval difference sequences. S102: Call the time point differences in the confidence interval difference sequence set, and perform sequence reconstruction processing in chronological order by sliding window merging and time index alignment to obtain a continuous sequence matrix and generate the difference sequence reconstruction matrix. S103: Based on the numerical change trend of the time points in the reconstruction matrix according to the difference sequence, calculate the rate of change and pre-set the rate of change threshold to divide and mark the intervals. After aggregating the trend intervals, obtain the predicted change trend set.

6. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the predicted trend set, select the difference content corresponding to every three consecutive time points to construct an analysis window, combine them into a difference triplet sequence in chronological order, index and organize the sequence, and generate a difference sliding window sequence group. S202: Call the difference sequence within the window of the difference sliding window sequence group, perform the front and back difference calculation between adjacent values ​​in sequence, construct the adjacent difference change sequence based on the difference results, and classify and aggregate them according to the time index to generate a continuous difference change sequence set; S203: Based on the set of continuous difference change sequences, retrieve the sign information of adjacent terms in the difference sequence, and determine whether there are continuous segments with consistent signs in the same sequence. If the preset continuous consistency condition is met, mark it as a stable trend segment. After aggregating all marked segments, obtain the confidence interval compression trend feature.

7. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the predicted change trend set, extract the predicted data and channel quality detection data frame corresponding to each time point, and compare the numerical symbol attributes of the two types of data after pairing according to the time index. Integrate the comparison results in chronological order to generate a predicted detection symbol comparison sequence. S302: Call the predicted detection symbol comparison sequence to identify the difference between the symbols of the predicted data and the detected data in the time point comparison results. If the symbols are inconsistent, they are marked as direction difference items, and all differences are recorded in chronological order to generate a direction difference mark sequence. S303: Based on the directional difference marker sequence, compare the difference markers of adjacent time points one by one to identify whether the sign change has a reversal feature. If a reversal relationship is identified, extract the change segment and perform sequence aggregation processing to obtain the error reversal direction feature.

8. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the confidence interval compression trend feature and the error reversal direction feature, extract the time index of the two types of features, construct a joint index lookup table according to the index order, organize the feature state values ​​at the corresponding positions, and generate an index feature mapping matrix; S402: Call the index feature mapping matrix, perform a logical comparison operation on the compression trend state value and the reversal direction state value of each index position, mark the index position when the combination of the two states meets the bidirectional consistency condition, aggregate all positions that meet the condition, and generate a consistent matching index set; S403: Based on the consistency matching index set, mark and encode all index positions in the set, assign a corresponding tag value to each index, construct a tag data frame structure, and obtain a repair trigger signal tag set.

9. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the predicted change trend set, extract all predicted data as feature sources, retrieve the index position in the repair trigger signal label set, bind the predicted data and label value corresponding to the position to the field, establish the corresponding combination relationship, and generate a predicted label pairing structure set. S502: Call the predicted label pairing structure set, sort it according to the time index of the predicted data, divide the training time period and the verification time period according to the sequence order, and store the predicted values ​​and label pairs in the interval into independent structures to generate an interval partitioned data pair set. S503: Based on the structure of the training and validation segments in the interval-divided data set, construct the mapping relationship between the input feature value matrix and the label vector, and jointly encapsulate them into a unified format data unit to obtain the sample structure for broadcast network fault prediction and self-healing learning.

10. The AI-based method for fault prediction and self-healing in broadcast networks according to claim 9, characterized in that, The joint encapsulation into a unified format data unit refers to combining the predicted feature matrix divided by time with the corresponding label vector in a one-to-one correspondence and organizing it into a unified input and output format that the model can recognize.

Images