High-reliability MT ferrule optical fiber connection monitoring system and fault early warning method

Abstract

The invention belongs to the technical field of fault monitoring, and discloses a high-reliability MT ferrule optical fiber connection monitoring system and a fault early warning method. Comprising the steps of collecting contact force distribution data of an MT insertion core and performing data cleaning to obtain an original contact force distribution data set; performing memory effect identification, and constructing an effect residual label to mark the original contact force distribution data set to obtain a marked contact force distribution data set; performing local spectrum decomposition, and outputting a spectrum characteristic frequency cluster; offset correction is carried out, and a corrected feature cluster is output; resonance mode distinguishing is carried out, mode marking is carried out on the corrected feature cluster based on a distinguishing result, and a mode distinguishing feature cluster is obtained; a return loss change curve is collected, and a degradation characteristic space is constructed; carrying out feature sub-region division to obtain degradation mode sub-regions; performing early warning evaluation, and outputting an early warning signal; and the accuracy of monitoring the optical fiber connection state and the reliability of early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of fault monitoring technology, and more specifically, to a high-reliability MT ferrule fiber optic connection monitoring system and fault early warning method. Background Technology

[0002] With the booming development of cloud computing and big data, data transmission demands are advancing towards higher bandwidth and higher density at an unprecedented pace. MT ferrules, with their unique high-density fiber optic connection solutions, are precisely addressing this trend, becoming a key choice for data center expansion and fiber optic technology innovation. Their compact design not only optimizes space utilization but also contributes to energy conservation and emission reduction, meeting the evolving needs of high-speed fiber optic standards. Furthermore, they demonstrate high reliability and precise alignment capabilities in critical sectors such as finance, healthcare, and defense. In addition, the high-density design of MT ferrules reduces the cost per connection and improves overall operational efficiency, making them a preferred choice for cost-effectiveness. Therefore, fault monitoring and early warning for fiber optic connections using MT ferrules have become an important research direction in this field.

[0003] During repeated insertion and removal, the tiny elastic elements on the ferrule end face or pin often exhibit extremely minor plastic deformation. Existing traditional monitoring methods lack the precision to identify this mechanical memory effect, and the loss at the connection interface cannot be assessed, leading to delayed fault detection. For example, irreversible structural degradation may occur after hundreds of insertions and removals, while traditional monitoring methods only observe short-term fluctuations in insertion and removal losses, ignoring the degradation trend. Furthermore, spectral characteristic frequencies often drift under temperature changes or contact angle shifts; however, traditional monitoring methods lack dynamic identification and compensation for environmentally induced shifts, easily generating environmentally sensitive false alarms and causing a decrease in early warning accuracy over long-term operation. Simultaneously, adhesive impurities such as dust or iron filings, as well as particulate matter, can damage the connection interface. Cracks or vibrations can excite specific frequency clusters, which may overlap with the structural resonant frequencies of the ferrule material itself. Traditional monitoring methods lack the ability to distinguish the source of resonance, which can easily lead to confusion between the intrinsic resonant response of the material and external interference signals. On the other hand, traditional monitoring methods lack a mechanism for handling overlapping fault boundaries of multiple categories. For example, fiber core end-face contamination and elastic fatigue often have overlapping characteristics in the frequency and power dimensions, but they have failed to make accurate distinctions, which limits the accuracy of fault mode identification. Since Mahalanobis distance calculation generally uses a fixed covariance matrix, and traditional monitoring methods fail to dynamically update the covariance estimate with the insertion and removal cycle, the accumulation of state drift misjudgments becomes increasingly serious, affecting the long-term monitoring accuracy of fiber optic connections and indirectly causing early warning failures.

[0004] In view of this, the present invention proposes a high-reliability MT ferrule fiber connection monitoring system and fault early warning method to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a high-reliability MT ferrule fiber connection fault early warning method, comprising:

[0006] S1. Collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the original contact force distribution dataset;

[0007] S2. Based on the original contact force distribution dataset, memory effect identification is performed, and residual effect labels are constructed to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset;

[0008] S3. Perform local spectral decomposition on the marked contact force distribution dataset and output spectral feature frequency clusters; perform offset correction on the spectral feature frequency clusters and output corrected feature clusters;

[0009] S4. Perform resonance mode differentiation on the modified feature clusters, and perform mode labeling on the modified feature clusters based on the differentiation results to obtain mode-discriminative feature clusters;

[0010] S5. Acquire the return loss variation curve, and construct the degradation feature space based on the return loss variation curve and mode differentiation feature cluster; divide the degradation feature space into feature sub-regions to obtain the degradation mode sub-regions.

[0011] S6. Perform early warning assessment on the deterioration mode sub-region and output early warning signal; send the early warning signal to the preset fault monitoring terminal for identification.

[0012] Furthermore, the method for identifying memory effects includes:

[0013] Extract the data belonging to each fiber core end face position from the original contact force distribution dataset and group them based on the known insertion and extraction cycles to obtain the unit fiber core cycle contact force group; calculate the extreme value difference of the normal force in the unit fiber core cycle contact force group to obtain the periodic normal force fluctuation amplitude;

[0014] Extract the shear force angle of the fiber core end face relative to the central axis of the fiber core array during each insertion and extraction in the unit fiber core cycle contact force group, and count the number of turning points where the change in shear force angle is higher than a preset angle threshold as the number of turning events.

[0015] The periodic normal force fluctuation amplitude of each fiber core end face is sorted according to the corresponding insertion and extraction cycle length. The difference in periodic normal force fluctuation amplitude between adjacent insertion and extraction cycles is calculated to construct a normal force fluctuation residual sequence. The rate of change of the first derivative of this normal force fluctuation residual sequence is calculated as the periodic deformation fluctuation rate.

[0016] The cumulative residual deformation is obtained by calculating the difference in the periodic normal force fluctuation amplitude between the first and last insertion / extraction cycles in a time-sequential sequence for the same fiber core end face.

[0017] The number of turning events on each fiber core end face is sorted according to the length of the corresponding insertion and extraction cycle, and the periodic growth rate of the number of turning events in adjacent insertion and extraction cycles is calculated to obtain the angle mutation growth rate.

[0018] The cumulative residual deformation, periodic deformation fluctuation rate, and angle mutation rate of the same fiber core end face are jointly normalized and constructed into a vector form to obtain the residual effect label.

[0019] Furthermore, the method for performing local spectral decomposition includes:

[0020] Construct a sliding window to traverse the marker contact force distribution dataset, perform wavelet transform on a portion of the marker contact force distribution data in each sliding window, and output the frequency domain energy band.

[0021] Calculate the energy density of each frequency band, select frequency bands with energy densities higher than a preset density threshold as high-density bands, and sort them according to their energy density to obtain an energy density sequence;

[0022] Calculate the density growth slope of the energy density sequence, and calculate the correlation coefficient between the density growth slope and each component in the residual effect label; select high-density frequency band combinations with correlation coefficients higher than a preset correlation threshold as spectral feature frequency clusters.

[0023] Furthermore, the method for performing offset correction includes:

[0024] Based on the residual effect labels, all fiber core end faces are grouped by wear degree to obtain worn fiber core subsets; the frequency values ​​of each frequency point belonging to the same worn fiber core subset and the same fiber core end face in the spectral feature frequency cluster are extracted in different insertion and removal cycles, and the average frequency difference between adjacent insertion and removal cycles is calculated to obtain the frequency change.

[0025] The normal force and shear force at the fiber core end face corresponding to the frequency change are extracted, and a contact force vector is generated based on the normal force and shear force. At the same time, the interface temperature of the same insertion and extraction cycle is obtained, and the ratio of the frequency change of adjacent insertion and extraction cycles to the contact force vector change and the interface temperature change is calculated. The results are then integrated to obtain the environmental induced offset at the corresponding frequency point.

[0026] Frequency points with environmental induced offset greater than a preset environmental sensitivity threshold are selected as sensitive frequency points, and the remaining frequency points are stable frequency points. Based on the environmental induced offset, the frequency value of the sensitive frequency point in each insertion and removal cycle is linearly backed up to obtain the corrected frequency value of the frequency point, and the frequency value of the stable frequency point is retained.

[0027] The corrected frequency values ​​of all frequency points are indexed and reorganized based on the corresponding fiber core end face positions, and all adjusted frequency points are integrated to obtain the corrected feature cluster.

[0028] Furthermore, the method for distinguishing resonance modes includes:

[0029] Based on the timestamp range corresponding to the insertion and removal cycle, the insertion and removal cycles are sorted by time to identify and correct the initial insertion and removal cycle of each frequency point in the feature cluster.

[0030] Within a preset time window before the initial insertion / removal cycle at each frequency point, extract the shear force time change sequence of the corresponding fiber core end face; calculate the ratio of the shear force change rate to the reference rate, and if the ratio is greater than the preset disturbance threshold, mark the corresponding frequency point as a disturbance trigger candidate point;

[0031] The frequency value corresponding to each disturbance triggering candidate point is counted in the subsequent continuous insertion and removal cycles. If the number of occurrences is lower than the preset continuity threshold, the disturbance triggering candidate point is preliminarily determined to be a non-structural resonance.

[0032] For disturbance triggering candidate points that are initially determined to be non-structural resonances, identify whether they exist simultaneously on multiple non-adjacent fiber core end faces in the corresponding MT ferrule fiber core array structure. If they do not appear on multiple non-adjacent fiber core end faces, they are determined to be non-structural resonances; otherwise, they are changed to structural resonances.

[0033] For frequency points that are not identified as disturbance triggering candidates, it is determined whether the frequency value change amplitude of the frequency point in the continuous insertion and removal cycle is less than the preset stability threshold. If it is less than the threshold, the point is identified as structural resonance. All the differentiated frequency points are integrated to obtain the mode distinguishing feature cluster.

[0034] Furthermore, the method of constructing the degraded feature space includes:

[0035] The statistical model distinguishes the energy density of each frequency point in the feature cluster, and constructs time-aligned data pairs based on the index points of the corresponding insertion and extraction cycles in the curve of energy density and return loss variation.

[0036] Calculate the cumulative energy value of the local region of the timestamp of the time-aligned data pair corresponding to the frequency point of non-structural resonance on each fiber core end face; calculate the energy mutation ratio based on the cumulative energy value and match it with the return loss value of the local region of the timestamp to construct the energy mutation loss data pair and integrate it into the mutation loss mapping dataset; perform fitting calculation on the mutation loss mapping dataset to obtain the energy loss mapping relationship.

[0037] The frequency drift is obtained by calculating the difference between the center frequency of the structural resonance point and the average frequency of each insertion and extraction cycle in a continuous insertion and extraction cycle. Based on the frequency drift and the return loss value of the corresponding insertion and extraction cycle, clustering is performed, and the cluster space is output to represent the drift loss mapping relationship.

[0038] Based on the energy loss mapping relationship, drift loss mapping relationship, effect residual label, number of resonance types and return loss value, construct the corresponding degradation feature vector of the fiber core end face, and construct the degradation feature space based on all degradation feature vectors.

[0039] Furthermore, the method for performing feature sub-region division includes:

[0040] Contamination behavior labels are constructed based on known fiber optic connection contamination behaviors, and contamination prototype vectors are constructed based on the contamination behavior labels for each type of fiber optic connection contamination behavior.

[0041] Normalize each type of pollution prototype vector and all deterioration feature vectors respectively, and calculate the cosine similarity between the normalized pollution prototype vector and the normalized deterioration feature vector. If it is higher than the preset similarity threshold, match the corresponding normalized deterioration feature vector with the corresponding pollution behavior label, and perform preliminary division of the deterioration feature space based on the pollution behavior label to obtain a preliminary set of feature sub-regions.

[0042] If a preliminary segmentation feature sub-region has a cosine similarity higher than a preset similarity threshold with multiple pollution prototype vectors, then principal component analysis is performed on the corresponding preliminary segmentation feature sub-region to output the relevant principal component vectors.

[0043] Extract the projection coordinates of each pollution prototype vector onto the relevant principal component vectors and integrate them into an intra-class projection point set; calculate the average vector of the current preliminary segmentation of feature sub-regions, and calculate the projection value of the average vector onto the relevant principal component vectors; calculate the Euclidean distance between the projection value and each point in the intra-class projection point set, and normalize each Euclidean distance to obtain the class weight, and use the pollution behavior label corresponding to the largest class weight as the label of the corresponding preliminary segmentation of feature sub-regions;

[0044] Integrating all the adjusted initial feature sub-region sets yields the deterioration pattern sub-region.

[0045] Furthermore, the methods for conducting early warning assessments include:

[0046] The degradation feature vector of the fiber core end face of each MT ferrule is acquired in real time. The Mahalanobis distance between the degradation feature vector and the statistical center of each degradation mode sub-region is calculated. Drift compensation is performed on the Mahalanobis distance, and the compensated Mahalanobis distance is output. If the compensated Mahalanobis distance is less than the preset safety distance, the corresponding fiber core end face is judged as a dangerous situation, and the corresponding pollution behavior label is output as a warning signal.

[0047] Furthermore, the method for performing drift compensation includes:

[0048] Extract the degradation feature vector sample set corresponding to the current Mahalanobis distance in adjacent consecutive insertion and extraction cycles; calculate the mean vector of the corresponding degradation feature vector sample set, and construct the time series covariance matrix based on the mean vector;

[0049] Obtain the historical covariance matrix of the degraded feature vector corresponding to the current Mahalanobis distance, and linearly fuse the historical covariance matrix with the temporal covariance matrix to obtain the dynamic compensation covariance matrix.

[0050] The Mahalanobis distance is recalculated based on the dynamically compensated covariance matrix to obtain the compensated Mahalanobis distance.

[0051] A high-reliability MT ferrule fiber optic connection monitoring system, used to implement a high-reliability MT ferrule fiber optic connection fault early warning method, characterized in that it includes:

[0052] The data acquisition module is used to collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the raw contact force distribution dataset;

[0053] The effect identification module is used to identify memory effects based on the original contact force distribution dataset, and to construct effect residual labels to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset.

[0054] The spectrum correction module is used to perform local spectrum decomposition on the marked contact force distribution dataset and output a cluster of spectral feature frequencies; it also performs offset correction on the spectral feature frequency clusters and outputs a corrected feature cluster.

[0055] The mode differentiation module is used to differentiate the resonant modes of the modified feature clusters, and to perform mode labeling on the modified feature clusters based on the differentiation results, thereby obtaining the mode-differentiated feature clusters;

[0056] The feature delimitation module is used to collect the return loss variation curve, construct the degradation feature space based on the return loss variation curve and mode differentiation feature clusters, and divide the degradation feature space into feature sub-regions to obtain the degradation mode sub-regions.

[0057] The early warning compensation module is used to perform early warning assessment on the deterioration mode sub-area and output early warning signals; the early warning signals are sent to the preset fault monitoring terminal for identification; and the modules are connected to each other via wired and / or wireless means.

[0058] The technical effects and advantages of the high-reliability MT ferrule fiber connection monitoring system and fault early warning method of this invention are as follows:

[0059] By acquiring and cleaning the contact force distribution data of the MT ferrule fiber core end face in real time, a raw contact force distribution dataset was obtained. Based on this dataset, memory effect identification, local spectrum decomposition, offset correction, resonance mode differentiation, degradation feature space construction, feature sub-region division, and drift compensation early warning assessment were carried out, realizing a fault early warning method for MT ferrule fiber connections. Compared with existing experience, the degradation phenomenon of the fiber core end face was quantified by identifying the mechanical memory effect after micro-plastic deformation of the connection interface and constructing effect residual labels. By extracting spectral feature frequency clusters and performing linear back-back correction based on environmentally induced offsets, false alarms of feature frequency drift caused by temperature or contact disturbances were reduced. By distinguishing between structural resonance and non-structural resonance and performing mode labeling, the difference between the intrinsic resonance characteristics of the material and external disturbances was clearly identified. By constructing a contamination prototype vector and using principal component projection weighted discrimination, accurate sub-region division was achieved under the condition of overlapping fault boundaries of multiple categories. By performing drift compensation on the calculated Mahalanobis distance, the misjudgment of state drift caused by covariance matrix error was avoided, improving the accuracy of fiber connection status monitoring and the reliability of early warning. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of a high-reliability MT ferrule fiber connection fault early warning method according to the present invention;

[0061] Figure 2 This is a schematic diagram of a high-reliability MT ferrule fiber connection monitoring system according to the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1

[0064] Please see Figure 1 As shown in this embodiment, a high-reliability MT ferrule fiber connection fault early warning method includes:

[0065] S1. Collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the original contact force distribution dataset;

[0066] S2. Based on the original contact force distribution dataset, memory effect identification is performed, and residual effect labels are constructed to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset;

[0067] S3. Perform local spectral decomposition on the marked contact force distribution dataset and output spectral feature frequency clusters; perform offset correction on the spectral feature frequency clusters and output corrected feature clusters;

[0068] S4. Perform resonance mode differentiation on the modified feature clusters, and perform mode labeling on the modified feature clusters based on the differentiation results to obtain mode-discriminative feature clusters;

[0069] S5. Acquire the return loss variation curve, and construct the degradation feature space based on the return loss variation curve and mode differentiation feature cluster; divide the degradation feature space into feature sub-regions to obtain the degradation mode sub-regions.

[0070] S6. Conduct early warning assessment on the deterioration mode sub-region and output early warning signal; send the early warning signal to the preset fault monitoring terminal for identification;

[0071] The MT ferrule is a ferrule with a high-density fiber optic connection configuration. The contact force distribution data of this ferrule includes the values ​​of normal pressure and shear force of each fiber core end face in the MT ferrule during the connection process. The normal pressure refers to the degree of compression between the fiber core end face and the contact surface, and the shear force reflects the degree of disturbance or misalignment of the fiber core end face during the connection insertion process. In this embodiment, the contact force distribution data is filled with missing values ​​and time aligned to obtain a higher quality original contact force distribution dataset.

[0072] Methods for identifying memory effects include:

[0073] Data belonging to each fiber core end face position in the original contact force distribution dataset are extracted and grouped based on known insertion and extraction cycles to obtain unit fiber core cycle contact force groups. The known insertion and extraction cycles refer to different insertion and extraction times. Each unit fiber core cycle contact force group belongs to the same fiber core end face and corresponds to a partial original contact force distribution dataset under a specific number of insertion and extraction times.

[0074] The extreme difference of normal force in the unit fiber core periodic contact force group is calculated to obtain the periodic normal force fluctuation amplitude. The periodic normal force fluctuation amplitude refers to the difference between the maximum and minimum normal force in the unit fiber core periodic contact force group, which reflects the activity level of the crimping state of the connection interface.

[0075] Extract the shear force angle of the fiber core end face relative to the central axis of the fiber core array during each insertion and extraction in the unit fiber core cycle contact force group, and count the number of turning points where the change in shear force angle exceeds a preset angle threshold as the number of turning events.

[0076] The fiber core array central axis refers to the geometric symmetry axis of the overall fiber array formed by the structural arrangement of all fiber core end faces in the MT ferrule. The shear force angle is obtained by extracting the angle relative to this axis during each insertion and removal. At the same time, the shear force angles are sorted based on the timestamps of the corresponding insertion and removal cycles, and the change in shear force angle between adjacent timestamps is calculated. An angle threshold is set based on the historical memory effect identification experience. If the change in shear force angle is higher than the preset angle threshold, the situation is judged as a turning point. The number of all turning points is counted to obtain the number of turning events. This value is used to reflect whether the force on the fiber core end face has deviated and the degree of deviation.

[0077] The periodic normal force fluctuation amplitude of each fiber core end face is sorted according to the corresponding insertion and extraction cycle length. The difference in periodic normal force fluctuation amplitude between adjacent insertion and extraction cycles is calculated to construct a normal force fluctuation residual sequence. The insertion and extraction cycle length refers to the number of insertions and extractions. The insertion and extraction cycles are sorted based on the number of insertions and extractions. By calculating the difference in periodic normal force fluctuation amplitude between adjacent insertion and extraction cycles, the crimping stability fluctuation of each fiber core end face under different insertion and extraction cycles is reflected. The difference in periodic normal force fluctuation amplitude between adjacent insertion and extraction cycles is combined to form the normal force fluctuation residual sequence.

[0078] The rate of change of the first derivative of the normal force fluctuation residual sequence is calculated as the periodic deformation fluctuation rate. The rate of change of the first derivative value is used as the periodic deformation fluctuation rate, which is the rate of increase or decrease, reflecting the changing characteristics of the deformation rate.

[0079] The cumulative deformation residual is obtained by calculating the difference in the periodic normal force fluctuation amplitude between the first and last insertion / extraction cycles of the same fiber core end face in a series of insertion / extraction cycles ordered by time. Specifically, by calculating the difference in the periodic normal force fluctuation amplitude between the initial insertion / extraction cycle and the most recently acquired insertion / extraction cycle of a certain fiber core end face, the cumulative deformation effect caused by the crimping state borne by that end face in the entire series of collectable insertion / extraction cycles is quantified.

[0080] The number of turning events on each fiber core end face is sorted according to the length of the corresponding insertion and extraction cycle, and the periodic growth rate of the number of turning events in adjacent insertion and extraction cycles is calculated to obtain the angle mutation growth rate. Specifically, by sorting the number of turning events on each fiber core end face under different insertion and extraction cycles, the growth rate of this value between adjacent cycles is calculated to identify whether the turning events show a periodic increase.

[0081] The cumulative residual deformation, periodic deformation fluctuation rate, and angle mutation rate of the same fiber core end face are jointly normalized and constructed into a vector form to obtain the effect residual label. The joint normalization process makes the data of each dimension have the same dimension, thus forming a multi-dimensional vector corresponding to a certain fiber core end face. This multi-dimensional vector is used as the effect residual label. This label is used to reflect the overall trajectory of the mechanical state evolution of the corresponding fiber core end face in the past multiple insertion and extraction cycles.

[0082] Methods for performing local spectral decomposition include:

[0083] A sliding window is constructed to traverse the marked contact force distribution dataset. Wavelet transform is performed on a portion of the marked contact force distribution data in each sliding window to output the frequency domain energy band. The size of the sliding window is set based on historical spectrum decomposition experience. Each sliding window corresponds to a portion of the marked contact force distribution data of a certain fiber core end face within a local insertion / extraction cycle. In this embodiment, wavelet transform is performed on this portion of the marked contact force distribution data to convert the original time domain data into frequency domain data, thereby identifying the frequency domain response. The obtained frequency domain energy band refers to the local energy distribution of the contact force signal in the corresponding time period under different frequency bandwidth conditions.

[0084] The energy density of each frequency band is calculated, and frequency bands with energy densities higher than a preset density threshold are selected as high-density bands. The energy density is then sorted based on the energy density to obtain an energy density sequence. Energy density refers to the energy concentration per unit bandwidth within a frequency band. A preset density threshold is set based on historical spectrum decomposition experience, and frequency bands with energy densities higher than this threshold are selected as high-density bands. The energy density sequence is obtained by sorting the frequency bands based on their energy density. This sequence is used to characterize the dominant activity of the frequency response in the high-density bands.

[0085] The density growth slope of the energy density sequence is calculated, and the correlation coefficient between the density growth slope and each component in the residual effect label is calculated. The density growth slope of the sequence is obtained by calculating the ratio of the difference between adjacent energy densities to the time interval. In this embodiment, the corresponding correlation coefficient is obtained by calculating the density growth slope and the Pearson coefficient of each dimension component in the residual effect label. It should be noted that all data have been normalized before calculating the correlation coefficient to eliminate dimensional differences.

[0086] High-density frequency band combinations with correlation coefficients higher than a preset correlation threshold are selected as spectral characteristic frequency clusters. Based on historical correlation judgment experience, correlation thresholds are set for the density growth slope and each component in the residual effect label. Since the components are different, the corresponding correlation thresholds are different, which are determined by historical experience or by consulting relevant theoretical knowledge. The above correlation coefficients are calculated to quantify the degree of correlation between frequency domain energy changes and mechanical state changes. This helps to determine whether the spectral behavior during MT ferrule connection has a trend correlation with mechanical states such as structural deformation. For example, local wear or loosening of the connection interface will cause corresponding frequency responses.

[0087] Methods for offset correction include:

[0088] Based on the residual effect labels, all fiber core end faces are grouped by wear degree to obtain worn fiber core subsets. In this embodiment, the threshold range of each component in the residual effect label is set by querying theoretical data related to fiber core wear. The wear level corresponding to the corresponding residual effect label is determined by matching the value of each component in each residual effect label with the corresponding threshold range. This wear level is used to characterize the wear degree of the fiber core end face. Based on the wear level, all fiber core end faces are grouped to obtain worn fiber core subsets, where all fiber core end faces in each worn fiber core subset belong to the same wear degree. The purpose of this operation is to classify fiber core end faces that show consistent or similar degradation trends.

[0089] The frequency values ​​of each frequency point belonging to the same worn fiber core subset and the same fiber core end face in the spectral feature frequency cluster are extracted in different insertion and removal cycles. The average frequency difference between adjacent insertion and removal cycles is calculated to obtain the frequency change. Specifically, the frequency point corresponding to a certain fiber core end face with a certain degree of wear in the spectral feature frequency cluster is identified. The frequency value refers to the average frequency of the frequency corresponding to that frequency point in the insertion and removal cycle. The difference of the average frequency in adjacent insertion and removal cycles is calculated to obtain the frequency change, which is used to reflect the frequency change trend of the same frequency point during the insertion and removal cycle.

[0090] The normal force and shear force at the fiber core end face corresponding to the frequency change are extracted, and a contact force vector is generated based on the normal force and shear force. The normal force and shear force at the fiber core end face corresponding to the frequency change are the average normal force and average shear force of the same period as the insertion and extraction period corresponding to the frequency value extracted above. These two components form a two-dimensional vector, which is the contact force vector.

[0091] Simultaneously, the interface temperature during the same insertion / removal cycle is acquired, and the ratios of the frequency change to the contact force vector change and the interface temperature change during adjacent insertion / removal cycles are calculated. These ratios are then integrated to obtain the environmental induced offset at the corresponding frequency point.

[0092] The formula for calculating the change in contact force vector is: ;in This represents the change in the contact force vector; This represents the average normal force value during a specific insertion / extraction cycle. This represents the average shear force value during a specific insertion / extraction cycle. This represents the average normal force value between adjacent insertion / removal cycles described above. The average shear force value of adjacent insertion and extraction cycles is represented by the above insertion and extraction cycle; the response ratio of frequency change to mechanical and thermal changes is formed by calculating the ratio of frequency change to contact force vector change and interface temperature change respectively, reflecting whether the frequency is easily affected by environmental factors and fluctuates; the environmental induced offset is obtained by weighted summation of the two ratios, where the weight is set based on historical offset correction experience.

[0093] Frequency points with environmentally induced offsets greater than a preset environmental sensitivity threshold are selected as sensitive frequency points, while the remaining frequency points are considered stable frequency points. The preset environmental sensitivity threshold is set based on historical offset correction experience. If the environmentally induced offset of a certain frequency point is large, the point is determined to be a sensitive frequency point, indicating that the frequency point is greatly affected by temperature or contact disturbances. Conversely, the frequency point is determined to be a stable frequency point if the environmentally induced offset is small.

[0094] Based on the environmentally induced offset, the frequency value of the sensitive frequency point in each insertion / removal cycle is linearly backed up to obtain the corrected frequency value of that frequency point. The frequency value of the stable frequency point is retained. In this embodiment, the environmentally induced offset is used as the adjustment amount, and the frequency value of the sensitive frequency point in each insertion / removal cycle is adjusted in reverse to achieve linear backing up, returning to its value before the disturbance, which is the corrected frequency value. The stable frequency point is not adjusted to ensure that its actual connection degradation signal is not affected.

[0095] The corrected frequency values ​​of all frequency points are indexed and reorganized based on the corresponding fiber core end face positions. All adjusted frequency points are integrated to obtain a corrected feature cluster. All frequency points with adjusted frequency values ​​are rebound to their original corresponding fiber core end face positions to prevent confusion of frequency points in subsequent operations. The frequency points in all spectral feature frequency clusters that need to be offset are adjusted accordingly to obtain a corrected feature cluster.

[0096] Methods for distinguishing resonance modes include:

[0097] The insertion and removal cycles are sorted by time based on the timestamp range corresponding to the insertion and removal cycles. The insertion and removal cycles that first appear at each frequency point in the corrected feature cluster are identified. All insertion and removal cycles are sorted based on the time period of the ferrule connection. Each frequency point in the corrected feature cluster is matched with the sorted sequence of insertion and removal cycles to confirm the corresponding insertion and removal cycle that was first detected at the frequency point in the entire insertion and removal history, which serves as the data basis for subsequent operations.

[0098] Within a preset time window before the initial insertion / removal cycle at each frequency point, the shear force time change sequence of the corresponding fiber core end face is extracted. In this embodiment, a fixed-length time window is extended forward from the initial insertion / removal cycle time point at the frequency point. This length is set based on historical pattern differentiation experience. The fiber core end face recorded data related to the frequency point is extracted within this time window, and the shear force time change sequence within this time window is identified.

[0099] Calculate the ratio of the shear force change rate to the reference rate. If the ratio is greater than the preset disturbance threshold, mark the corresponding frequency point as a disturbance triggering candidate point. The reference rate refers to the standard shear force change rate set by querying historical samples. Calculate the ratio of the shear force change rate at this time to the reference rate. If the ratio is greater than the disturbance threshold set based on historical pattern differentiation experience, it indicates that the frequency point is accompanied by an abnormal mechanical state when it first appears. Therefore, mark the frequency point as a disturbance triggering candidate point.

[0100] The frequency value corresponding to each disturbance triggering candidate point is counted in the subsequent continuous insertion and removal cycles. If the number of occurrences is lower than the preset continuity threshold, the disturbance triggering candidate point is preliminarily determined to be a non-structural resonance.

[0101] This involves counting the number of times each disturbance trigger candidate point appears in a number of consecutive insertion / removal cycles after the first occurrence. The specific number of cycles can be adjusted based on the specific operating conditions, and this specific number of cycles must be less than a preset continuity threshold. It should be noted that the preset continuity threshold is set based on historical pattern differentiation experience. If the disturbance trigger frequency point appears in most subsequent consecutive cycles, it may be an inherent frequency of the structure. If it is less than the preset continuity threshold, it may be a frequency caused by short-term disturbances such as instantaneous disturbances. Therefore, this situation is initially judged as non-structural resonance.

[0102] For disturbance triggering candidate points that are initially determined to be non-structural resonances, it is determined whether they exist simultaneously on multiple non-adjacent fiber core end faces in the corresponding MT ferrule fiber core array structure. If they do not appear on multiple non-adjacent fiber core end faces, they are determined to be non-structural resonances; otherwise, they are changed to structural resonances.

[0103] Further spatial analysis is performed on the candidate points for disturbance triggering non-structural resonance to determine whether the point has appeared on multiple discontinuous fiber core end faces. In this embodiment, a certain judgment threshold is set based on historical experience. If it appears on multiple discontinuous fiber core end faces and the number is higher than the threshold, it indicates that the corresponding frequency point belongs to a certain general frequency disturbance (such as end face material modal characteristics, array resonance mechanism, etc.), and the resonance mode is then modified to structural resonance. Otherwise, it indicates that it is caused by random disturbance and has an occasional nature.

[0104] For frequency points that are not identified as candidate points for disturbance triggering, it is determined whether the frequency value change amplitude of the frequency point in the continuous insertion and removal cycle is less than the preset stability threshold. If it is less than the threshold, the point is determined to be a structural resonance.

[0105] For frequency points that were not initially identified as candidate points for disturbance triggering, the frequency value change amplitude of the point during continuous insertion and removal cycles is calculated. If the frequency value change amplitude is always less than the stability threshold set based on historical pattern differentiation experience, it indicates that the frequency of the point has high stability and may come from the inherent modal characteristics of the material or structure, and is not affected by temporary disturbances. Therefore, this situation is classified as structural resonance.

[0106] Integrating all the differentiated frequency points yields a pattern-distinguishing feature cluster.

[0107] Methods for constructing a degraded feature space include:

[0108] The statistical model distinguishes the energy density of each frequency point in the feature cluster, and constructs time-aligned data pairs based on the index points of the corresponding insertion and removal cycles in the return loss variation curve.

[0109] In this embodiment, the frequency energy of each frequency point in the mode distinguishing feature cluster is statistically processed in different insertion and removal cycles. The energy density of the point in any insertion and removal cycle is calculated, and the energy density is matched with the index data of the relevant points in the time segment of the corresponding insertion and removal cycle in the return loss change curve to form a time-aligned data pair. The sequence formed by sorting the return loss value at each time point in the corresponding insertion and removal cycle by time is plotted as a curve, which is the return loss change curve.

[0110] Calculate the energy accumulation value of the local region of the timestamp of the time-aligned data pair corresponding to the frequency point of the non-structural resonance on each fiber core end face.

[0111] The local time stamp region refers to a local time segment obtained by traversing a sliding window of a set length within the time segment formed by the insertion and removal cycles corresponding to the time alignment data pair. The time length of this local time segment is equal to the length of each sliding window, and the specific length is set based on historical experience. The total energy within each window is counted to obtain the cumulative energy value.

[0112] The energy mutation ratio is calculated based on the accumulated energy value and matched with the return loss value of the local area of ​​the timestamp. Energy mutation loss data pairs are constructed and integrated into a mutation loss mapping dataset. The energy mutation ratio refers to the ratio of the accumulated energy value of the local area of ​​the timestamp to the basic total energy set based on existing frequency domain related theoretical knowledge. At the same time, the return loss value of the corresponding time segment of the local area of ​​the timestamp is extracted. The mutation ratio and the return loss value are matched to construct energy mutation loss data pairs, which reflect the relationship between energy change and loss under mutation conditions.

[0113] The energy loss mapping relationship is obtained by fitting the dataset of abrupt loss. In this embodiment, the regression fitting model is used to process the dataset of abrupt loss to construct the mapping relationship between the degree of frequency energy burst and the degree of actual return loss degradation under the condition of non-structural resonance. This characterizes the relationship between non-structural resonance and optical return loss caused by conditions such as insertion and removal impact, particle intrusion or connector contamination.

[0114] The frequency drift is calculated by measuring the difference between the center frequency of the structural resonance point and the average frequency of each insertion / extraction cycle in consecutive insertion / extraction cycles. Based on this frequency drift and the return loss value of the corresponding insertion / extraction cycle, clustering is performed, and the cluster space is output to represent the drift loss mapping relationship.

[0115] In this embodiment, the center frequency refers to the inherent standard frequency of the structural resonance frequency point. Due to the inherent characteristics of the corresponding material or structure of the structural resonance, the center frequency is set by consulting relevant information on the fiber core end face material or design structure. The frequency drift is obtained by calculating the difference between the center frequency and the average frequency of each insertion and extraction cycle. In this embodiment, the frequency drift and the return loss value of the corresponding insertion and extraction cycle are used as input data for the clustering model to achieve matching of different frequency drift levels with corresponding loss states. Each cluster center in the formed cluster space corresponds to a matching region of frequency drift and return loss. There are several cluster centers in this cluster space, so this cluster space is used to characterize the drift loss mapping relationship.

[0116] Based on the energy loss mapping relationship, drift loss mapping relationship, effect residual label, number of resonance types, and return loss value, a degradation feature vector corresponding to the fiber core end face is constructed, and a degradation feature space is constructed based on all degradation feature vectors.

[0117] The process involves transforming the functional relationship constructed by the regression fitting model, which pertains to the energy loss mapping relationship, into a multidimensional numerical vector. The cluster center frequencies of all frequency points corresponding to a certain cluster center in the drift loss mapping relationship are calculated. The multidimensional numerical vectors representing the energy loss mapping relationships of the same fiber core endface, along with the cluster center frequencies, components in the residual effect labels, the number of frequency points for each resonance type, and the return loss values ​​marked on the return loss variation curve, are combined into a multidimensional degradation feature vector to characterize the degradation state of the corresponding fiber core endface. All degradation feature vectors are then aggregated to generate a degradation feature space.

[0118] Methods for dividing feature subregions include:

[0119] Contamination behavior labels are constructed based on known fiber optic connection contamination behaviors, and contamination prototype vectors are constructed based on the contamination behavior labels for each type of fiber optic connection contamination behavior.

[0120] The known fiber optic connection contamination behaviors are determined based on expert experience and relevant domain knowledge. Examples include external foreign matter such as dust or metal shavings adsorbing onto the connection interface, moisture contamination of the end face due to high humidity, residual cleaning chemical solvents on the end face, and scratches on the fiber core end face caused by repeated insertion and removal. Each type of contamination behavior is labeled accordingly. In this embodiment, for each label, corresponding historical samples are extracted, and the average feature performance is calculated based on these samples to obtain the contamination prototype vector for that type of contamination. This contamination prototype vector reflects the average degradation under multi-dimensional indicators such as frequency shift, energy mutation, or mechanical residue during various contamination processes. To facilitate subsequent calculations, it is necessary to filter the components within the contamination prototype vector based on historical experience so that the dimensions of these components are the same as the dimensions of the components in the degradation feature vector.

[0121] Normalization is performed on each pollution prototype vector and all degradation feature vectors. The cosine similarity between the normalized pollution prototype vector and the normalized degradation feature vector is calculated. If the similarity exceeds a preset threshold, the corresponding normalized degradation feature vector is matched with the corresponding pollution behavior label. Based on this pollution behavior label, the degradation feature space is initially divided, resulting in a preliminary set of feature sub-regions.

[0122] The process involves normalizing all pollution prototype vectors and degradation feature vectors to eliminate dimensional differences, and calculating the cosine similarity between each pollution prototype vector and each degradation feature vector. A similarity threshold is set based on historical sub-region division experience. If the obtained cosine similarity is higher than this threshold, it indicates that a certain degradation feature vector has a high degree of consistency with the pollution pattern of the corresponding type of pollution prototype vector, and the degradation feature vector is then assigned to the corresponding type of pollution behavior label. After all degradation feature vectors are assigned to their corresponding pollution categories, the degradation feature space is initially divided, resulting in a preliminary set of feature sub-regions.

[0123] If a preliminarily defined feature sub-region has a cosine similarity higher than a preset similarity threshold with multiple pollution prototype vectors, then principal component analysis is performed on the corresponding preliminarily defined feature sub-regions to output the relevant principal component vectors.

[0124] If a preliminary segmentation feature sub-region has a similarity score higher than the threshold with multiple pollution prototype vectors, it indicates that there is regional overlap forming a fuzzy sub-region. To ensure the accuracy of segmentation, the maximum similarity score is not used directly for differentiation. In this embodiment, the principal component analysis algorithm is used to extract the dominant change direction of the preliminary segmentation feature sub-region in the degraded feature space, forming a relevant principal component vector. This relevant principal component vector represents the principal axis of the convergence of diverse features in the current fuzzy sub-region, corresponding to the combined projection line of the parameters of each dimension.

[0125] The projected coordinates of each pollution prototype vector on the relevant principal component vector are extracted and integrated into an intra-class projection point set. Each pollution prototype vector is projected onto the direction of the relevant principal component vector to obtain the projected coordinates in that direction. The projected coordinates of all pollution prototype vectors are integrated to obtain the intra-class projection point set, which is used to characterize the performance of various pollution patterns in the dominant change direction of the fuzzy sub-region. It should be noted that this calculation is performed in the coordinate system corresponding to the degraded feature space, and the subsequent feature sub-region division operation is also performed in this coordinate system.

[0126] Calculate the average vector of the currently preliminarily segmented feature sub-regions, and calculate the projection value of this average vector onto the direction of the relevant principal component vectors. The average vector of the sub-region is obtained by averaging all vector samples in the currently preliminarily segmented feature sub-regions with fuzzy sub-regions. This average vector is used to describe the collective bias feature of the overall fuzzy behavior. The projection value of this average vector onto the direction of the relevant principal component vectors is calculated, which is the projection coordinate in that direction.

[0127] Calculate the Euclidean distance between the projected value and each point in the in-class projection point set, and normalize each Euclidean distance to obtain the class weight. The pollution behavior label corresponding to the largest class weight is used as the label for the corresponding preliminary feature sub-region. Specifically, calculate the Euclidean distance between each point in the in-class projection point set and the above projection value to form a set of distances from the representative point of the sub-region to all pollution prototype labels. Then, normalize these distances to convert them into relative similarity weights for each pollution category. The label with the largest weight is determined as the final pollution label of the sub-region, representing the closest degradation mode of the fuzzy sub-region.

[0128] Integrating all the adjusted initial feature sub-region sets yields the deterioration pattern sub-region.

[0129] Methods for conducting early warning assessments include:

[0130] The degradation feature vector of the fiber core end face of each MT ferrule is acquired in real time. The Mahalanobis distance between the degradation feature vector and the statistical center of each degradation mode sub-region is calculated. The statistical center refers to the mean vector of each feature dimension in the corresponding degradation mode sub-region. The degradation feature vector of the fiber core end face of each MT ferrule is acquired in real time, and the Mahalanobis distance between the vector and the statistical center is calculated. This distance reflects the degree of difference and risk level of the corresponding state point relative to the mode center of the sub-region in the degradation feature space.

[0131] Drift compensation is performed on the Mahalanobis distance, and the compensated Mahalanobis distance is output. If the compensated Mahalanobis distance is less than the preset safety distance, the corresponding fiber core end face is judged as a dangerous situation, and the corresponding pollution behavior label is output as a warning signal. The safety distance is set based on historical assessment experience. If the compensated Mahalanobis distance is less than the safety distance, it means that the corresponding fiber core end face has approached or crossed the predefined degradation boundary in the degradation feature space. This state is considered a risk state, and the identification result is output according to the pollution behavior label bound to the current degradation mode sub-region. The identification result is encoded into a warning signal that the computer can recognize.

[0132] The methods for performing drift compensation include:

[0133] Extract the degradation feature vector sample set corresponding to the current Mahalanobis distance in the historical adjacent consecutive insertion and extraction cycles. Taking the current fiber core end face as the index, extract the degradation feature vector sample set in the historical adjacent consecutive insertion and extraction cycles. The sample set composed of these degradation feature vectors reflects the feature evolution trajectory of the fiber core end face in the recent consecutive insertion and extraction cycles.

[0134] Calculate the mean vector of the corresponding degradation feature vector sample set, and construct a time-series covariance matrix based on the mean vector. The time-series covariance matrix corresponding to the mean vector of the degradation feature vector sample set is calculated based on existing mathematical knowledge. This matrix reflects the joint relationship of recent degradation features across different variable dimensions.

[0135] Obtain the historical covariance matrix of the degraded eigenvector corresponding to the current Mahalanobis distance. Linearly fuse this historical covariance matrix with the temporal covariance matrix to obtain the dynamic compensation covariance matrix. This is achieved by using the historical covariance matrix corresponding to the degraded eigenvector in the earlier and more accurate historical records as the data basis. This matrix is ​​then fused and updated with the temporal covariance matrix using a linear weighting method to construct a new dynamic compensation covariance matrix. The weights are set based on historical drift compensation experience.

[0136] The Mahalanobis distance is recalculated based on the dynamically compensated covariance matrix to obtain the compensated Mahalanobis distance. The newly constructed dynamically compensated covariance matrix is ​​then substituted back into the Mahalanobis distance calculation formula to obtain the compensated Mahalanobis distance. This compensated Mahalanobis distance is closer to the true dynamic distribution of variables in the current cycle connection interface operation state, avoiding misjudgments caused by covariance estimation errors.

[0137] This embodiment obtains the original contact force distribution dataset by real-time acquisition and data cleaning of the contact force distribution data of the MT ferrule fiber core end face. Based on this dataset, memory effect identification, local spectrum decomposition, offset correction, resonance mode differentiation, degradation feature space construction, feature sub-region division, and drift compensation early warning assessment are performed to realize the MT ferrule fiber connection fault early warning method. Compared with existing experience, the degradation phenomenon of the fiber core end face is quantified by identifying the mechanical memory effect after micro-plastic deformation of the connection interface and constructing effect residual labels. By extracting spectral feature frequency clusters and performing linear back-back correction based on environmentally induced offset, false alarms of feature frequency drift caused by temperature or contact disturbance are reduced. By distinguishing between structural resonance and non-structural resonance and performing mode labeling, the difference between the intrinsic resonance characteristics of the material and external disturbances is clearly identified. By constructing a contamination prototype vector and using principal component projection weighted discrimination, accurate sub-region division is achieved under the condition of overlapping fault boundaries of multiple categories. By performing drift compensation on the calculated Mahalanobis distance, the state drift misjudgment caused by covariance matrix error is avoided, improving the accuracy and reliability of fiber connection status monitoring and early warning.

[0138] Example 2

[0139] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A high-reliability MT ferrule fiber connection monitoring system is provided, including:

[0140] The data acquisition module is used to collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the raw contact force distribution dataset;

[0141] The effect identification module is used to identify memory effects based on the original contact force distribution dataset, and to construct effect residual labels to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset.

[0142] The spectrum correction module is used to perform local spectrum decomposition on the marked contact force distribution dataset and output a cluster of spectral feature frequencies; it also performs offset correction on the spectral feature frequency clusters and outputs a corrected feature cluster.

[0143] The mode differentiation module is used to differentiate the resonant modes of the modified feature clusters, and to perform mode labeling on the modified feature clusters based on the differentiation results, thereby obtaining the mode-differentiated feature clusters;

[0144] The feature delimitation module is used to collect the return loss variation curve, construct the degradation feature space based on the return loss variation curve and mode differentiation feature clusters, and divide the degradation feature space into feature sub-regions to obtain the degradation mode sub-regions.

[0145] The early warning compensation module is used to perform early warning assessment on the deterioration mode sub-area and output early warning signals; the early warning signals are sent to the preset fault monitoring terminal for identification; and the modules are connected to each other via wired and / or wireless means.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0147] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0148] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A high-reliability MT ferrule fiber connection fault early warning method, characterized in that, include: S1. Collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the original contact force distribution dataset; S2. Based on the original contact force distribution dataset, memory effect identification is performed, and residual effect labels are constructed to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset; S3. Perform local spectral decomposition on the marked contact force distribution dataset and output the spectral feature frequency clusters; Shift and correct the frequency clusters of spectral features, and output the corrected feature clusters; S4. Perform resonance mode differentiation on the modified feature clusters, and perform mode labeling on the modified feature clusters based on the differentiation results to obtain mode-discriminative feature clusters; S5. Acquire the return loss variation curve, and construct the degradation feature space based on the return loss variation curve and mode differentiation feature cluster; divide the degradation feature space into feature sub-regions to obtain the degradation mode sub-regions. S6. Perform early warning assessment on the deterioration pattern sub-region and output early warning signals; The warning signal is sent to a preset fault monitoring terminal for identification.

2. The high-reliability MT ferrule fiber connection fault early warning method according to claim 1, characterized in that, The methods for identifying memory effects include: Extract the data belonging to each fiber core end face position from the original contact force distribution dataset and group them based on the known insertion and extraction cycles to obtain the unit fiber core cycle contact force group; calculate the extreme value difference of the normal force in the unit fiber core cycle contact force group to obtain the periodic normal force fluctuation amplitude; Extract the shear force angle of the fiber core end face relative to the central axis of the fiber core array during each insertion and extraction in the unit fiber core cycle contact force group, and count the number of turning points where the change in shear force angle is higher than a preset angle threshold as the number of turning events. The periodic normal force fluctuation amplitude of each fiber core end face is sorted according to the corresponding insertion and extraction cycle length. The difference in periodic normal force fluctuation amplitude between adjacent insertion and extraction cycles is calculated to construct a normal force fluctuation residual sequence. The rate of change of the first derivative of this normal force fluctuation residual sequence is calculated as the periodic deformation fluctuation rate. The cumulative residual deformation is obtained by calculating the difference in the periodic normal force fluctuation amplitude between the first and last insertion / extraction cycles in a time-sequential sequence for the same fiber core end face. The number of turning events on each fiber core end face is sorted according to the length of the corresponding insertion and extraction cycle, and the periodic growth rate of the number of turning events in adjacent insertion and extraction cycles is calculated to obtain the angle mutation growth rate. The cumulative residual deformation, periodic deformation fluctuation rate, and angle mutation rate of the same fiber core end face are jointly normalized and constructed into a vector form to obtain the residual effect label.

3. The high-reliability MT ferrule fiber connection fault early warning method according to claim 2, characterized in that, The methods for performing local spectral decomposition include: Construct a sliding window to traverse the marker contact force distribution dataset, perform wavelet transform on a portion of the marker contact force distribution data in each sliding window, and output the frequency domain energy band. Calculate the energy density of each frequency band, select frequency bands with energy densities higher than a preset density threshold as high-density bands, and sort them according to their energy density to obtain an energy density sequence; Calculate the density growth slope of the energy density sequence, and calculate the correlation coefficient between the density growth slope and each component in the residual effect label; select high-density frequency band combinations with correlation coefficients higher than a preset correlation threshold as spectral feature frequency clusters.

4. The high-reliability MT ferrule fiber connection fault early warning method according to claim 3, characterized in that, The methods for performing offset correction include: Based on the residual effect labels, all fiber core end faces are grouped by wear degree to obtain worn fiber core subsets; the frequency values ​​of each frequency point belonging to the same worn fiber core subset and the same fiber core end face in the spectral feature frequency cluster are extracted in different insertion and removal cycles, and the average frequency difference between adjacent insertion and removal cycles is calculated to obtain the frequency change. The normal force and shear force at the fiber core end face corresponding to the frequency change are extracted, and a contact force vector is generated based on the normal force and shear force. At the same time, the interface temperature of the same insertion and extraction cycle is obtained, and the ratio of the frequency change of adjacent insertion and extraction cycles to the contact force vector change and the interface temperature change is calculated. The results are then integrated to obtain the environmental induced offset at the corresponding frequency point. Frequency points with environmental induced offset greater than a preset environmental sensitivity threshold are selected as sensitive frequency points, and the remaining frequency points are stable frequency points. Based on the environmental induced offset, the frequency value of the sensitive frequency point in each insertion and removal cycle is linearly backed up to obtain the corrected frequency value of the frequency point, and the frequency value of the stable frequency point is retained. The corrected frequency values ​​of all frequency points are indexed and reorganized based on the corresponding fiber core end face positions, and all adjusted frequency points are integrated to obtain a corrected feature cluster.

5. The high-reliability MT ferrule fiber connection fault early warning method according to claim 4, characterized in that, The methods for distinguishing resonance modes include: Based on the timestamp range corresponding to the insertion and removal cycle, the insertion and removal cycles are sorted by time to identify and correct the initial insertion and removal cycle of each frequency point in the feature cluster. Within a preset time window before the initial insertion / removal cycle at each frequency point, extract the shear force time change sequence of the corresponding fiber core end face; calculate the ratio of the shear force change rate to the reference rate, and if the ratio is greater than the preset disturbance threshold, mark the corresponding frequency point as a disturbance trigger candidate point; The frequency value corresponding to each disturbance triggering candidate point is counted in the subsequent continuous insertion and removal cycles. If the number of occurrences is lower than the preset continuity threshold, the disturbance triggering candidate point is preliminarily determined to be a non-structural resonance. For disturbance triggering candidate points that are initially determined to be non-structural resonances, identify whether they exist simultaneously on multiple non-adjacent fiber core end faces in the corresponding MT ferrule fiber core array structure. If they do not appear on multiple non-adjacent fiber core end faces, they are determined to be non-structural resonances; otherwise, they are changed to structural resonances. For frequency points that are not identified as disturbance triggering candidates, it is determined whether the frequency value change amplitude of the frequency point in the continuous insertion and removal cycle is less than the preset stability threshold. If it is less than the threshold, the point is identified as structural resonance. All the differentiated frequency points are integrated to obtain the mode distinguishing feature cluster.

6. The high-reliability MT ferrule fiber connection fault early warning method according to claim 5, characterized in that, The methods for constructing the degraded feature space include: The statistical model distinguishes the energy density of each frequency point in the feature cluster, and constructs time-aligned data pairs based on the index points of the corresponding insertion and extraction cycles in the curve of energy density and return loss variation. Calculate the cumulative energy value of the local region of the timestamp of the time-aligned data pair corresponding to the frequency point of non-structural resonance on each fiber core end face; calculate the energy mutation ratio based on the cumulative energy value and match it with the return loss value of the local region of the timestamp to construct the energy mutation loss data pair and integrate it into the mutation loss mapping dataset; perform fitting calculation on the mutation loss mapping dataset to obtain the energy loss mapping relationship. The frequency drift is obtained by calculating the difference between the center frequency of the structural resonance point and the average frequency of each insertion and extraction cycle in a continuous insertion and extraction cycle. Based on the frequency drift and the return loss value of the corresponding insertion and extraction cycle, clustering is performed, and the cluster space is output to represent the drift loss mapping relationship. Based on the energy loss mapping relationship, drift loss mapping relationship, effect residual label, number of resonance types and return loss value, construct the corresponding degradation feature vector of the fiber core end face, and construct the degradation feature space based on all degradation feature vectors.

7. A high-reliability MT ferrule fiber connection fault early warning method according to claim 6, characterized in that, The methods for performing feature sub-region division include: Contamination behavior labels are constructed based on known fiber optic connection contamination behaviors, and contamination prototype vectors are constructed based on the contamination behavior labels for each type of fiber optic connection contamination behavior. Normalize each type of pollution prototype vector and all deterioration feature vectors respectively, and calculate the cosine similarity between the normalized pollution prototype vector and the normalized deterioration feature vector. If it is higher than the preset similarity threshold, match the corresponding normalized deterioration feature vector with the corresponding pollution behavior label, and perform preliminary division of the deterioration feature space based on the pollution behavior label to obtain a preliminary set of feature sub-regions. If a preliminary segmentation feature sub-region has a cosine similarity higher than a preset similarity threshold with multiple pollution prototype vectors, then principal component analysis is performed on the corresponding preliminary segmentation feature sub-region to output the relevant principal component vectors. Extract the projection coordinates of each pollution prototype vector onto the relevant principal component vectors and integrate them into an intra-class projection point set; calculate the average vector of the current preliminary segmentation of feature sub-regions, and calculate the projection value of the average vector onto the relevant principal component vectors; calculate the Euclidean distance between the projection value and each point in the intra-class projection point set, and normalize each Euclidean distance to obtain the class weight, and use the pollution behavior label corresponding to the largest class weight as the label of the corresponding preliminary segmentation of feature sub-regions; Integrating all the adjusted initial feature sub-region sets yields the deterioration pattern sub-region.

8. A high-reliability MT ferrule fiber connection fault early warning method according to claim 7, characterized in that, The methods for conducting early warning assessments include: The degradation feature vector of the fiber core end face of each MT ferrule is acquired in real time. The Mahalanobis distance between the degradation feature vector and the statistical center of each degradation mode sub-region is calculated. Drift compensation is performed on the Mahalanobis distance, and the compensated Mahalanobis distance is output. If the compensated Mahalanobis distance is less than the preset safety distance, the corresponding fiber core end face is judged as a dangerous situation, and the corresponding pollution behavior label is output as a warning signal.

9. A high-reliability MT ferrule fiber connection fault early warning method according to claim 8, characterized in that, The methods for performing drift compensation include: Extract the degradation feature vector sample set corresponding to the current Mahalanobis distance in adjacent consecutive insertion and extraction cycles; calculate the mean vector of the corresponding degradation feature vector sample set, and construct the time series covariance matrix based on the mean vector; Obtain the historical covariance matrix of the degraded feature vector corresponding to the current Mahalanobis distance, and linearly fuse the historical covariance matrix with the temporal covariance matrix to obtain the dynamic compensation covariance matrix; The Mahalanobis distance is recalculated based on the dynamically compensated covariance matrix to obtain the compensated Mahalanobis distance.

10. A high-reliability MT ferrule fiber optic connection monitoring system, used to implement the high-reliability MT ferrule fiber optic connection fault early warning method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect the contact force distribution data of the MT ferrule and perform data cleaning to obtain the raw contact force distribution dataset; The effect identification module is used to identify memory effects based on the original contact force distribution dataset, and to construct effect residual labels to label the original contact force distribution dataset to obtain a labeled contact force distribution dataset. The spectrum correction module is used to perform local spectrum decomposition on the marked contact force distribution dataset and output spectrum feature frequency clusters; Shift and correct the frequency clusters of spectral features, and output the corrected feature clusters; The mode differentiation module is used to differentiate the resonant modes of the modified feature clusters, and to perform mode labeling on the modified feature clusters based on the differentiation results, thereby obtaining the mode-differentiated feature clusters; The feature delimitation module is used to collect the return loss variation curve and construct the degradation feature space based on the return loss variation curve and mode differentiation feature clusters. The degradation feature space is divided into feature sub-regions to obtain degradation pattern sub-regions; The early warning compensation module is used to perform early warning assessment on the deterioration mode sub-region and output early warning signals; The warning signal is sent to a preset fault monitoring terminal for identification; the modules are connected to each other via wired and / or wireless means.

Images