A method of optical communication link light source monitoring

Abstract

This invention discloses a method for monitoring light sources in optical communication links, relating to the field of optical communication monitoring technology. The method includes: acquiring the emission driving current sequence, monitoring photocurrent sequence, forward-sampled optical power sequence, and source-side back-feedback sensing sequence of the light source; obtaining the emission response residual sequence and back-feedback perturbation residual sequence after removing the common-mode response; constructing a path difference spectrum and extracting local peak sets; calculating the path difference discreteness, dominant oscillation frequency, and coherent back-feedback cluster intensity index; applying a constrained micro-probe current waveform through reinforcement learning to calculate the detection gain; and generating a pseudo-stability instability risk value by combining apparent stability metrics and outputting the monitoring results. This invention can identify the latent instability risk of light sources under surface stability conditions, improving monitoring sensitivity and timely early warning.

Description

Technical Field

[0001] This invention relates to the field of optical communication monitoring technology, and in particular to a method for monitoring the light source of an optical communication link. Background Technology

[0002] In high-speed optical communication systems, silicon photonics transmission links, pluggable optical modules, and transmission links containing multiple connectors and chip light-emitting end faces, the forward light emitted by the light source is easily affected by end face contamination, minor misalignment, reflection, and device aging when passing through connector end faces, edge interfaces, light-collecting structures, and chip light-emitting positions. This causes some light energy to return to the source side and act on the light source again, resulting in abnormal fluctuations in the light emission state of the light source. For such links, a difficult-to-identify problem often occurs in actual operation: the forward sampling optical power and monitoring photocurrent remain relatively stable on the surface, while a back disturbance caused by multiple weak reflections has formed on the source side. This causes the light source to gradually enter a latent unstable state. If it cannot be identified in time, it can easily develop into output instability, increased link bit error rate, or even a decline in communication quality.

[0003] When addressing the aforementioned issues, most existing technologies still rely on average optical power, monitored photocurrent, return loss results, or bit error rate as the primary criteria for judgment. These technologies typically only reflect existing power attenuation or significant instability, failing to reveal latent anomalies caused by the combined effects of multiple weak reflections on time delay distribution and oscillation rhythm. Furthermore, they lack an active detection process for the current operating state, making it difficult to amplify the hidden return disturbances without significantly disrupting normal service transmission. Consequently, they are prone to misjudging seemingly stable but internally risky states as normal, resulting in delayed early warnings, insufficient monitoring sensitivity, and inadequate identification of early instability processes. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies in identifying the latent instability risk of light sources under stable surface conditions, and to propose a light source monitoring method for optical communication links.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: A method for monitoring the light source of an optical communication link, comprising: S1. Collect the emission drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence and source-side back sensing sequence of the light source, remove the common-mode response, and obtain the emission response residual sequence and the back disturbance residual sequence. S2. Construct the path difference spectrum based on the luminescence response residual sequence and the return perturbation residual sequence and extract the local peak set to calculate the path difference discreteness. S3. Extract the dominant oscillation frequency from the luminescence response residual sequence, and calculate the coherent return cluster intensity index by combining the local peak set and the path difference discrepancy. S4. Based on the forward sampling optical power sequence, monitoring photocurrent sequence, coherent back cluster intensity index and path difference discreteness, a constrained micro-probe current waveform is applied at the current operating point through reinforcement learning, and the detection gain is calculated based on the coherent back cluster intensity index before and after detection. S5. Calculate the apparent stability measure based on the forward sampling optical power sequence and the monitoring photocurrent sequence, combine the coherent return cluster intensity index and the detection gain to calculate the pseudo-stability and instability risk value, and output the monitoring results.

[0006] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention, by simultaneously acquiring the emission drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence, and source-side back-feedback sensing sequence, first eliminates the common-mode response directly introduced by the drive change. Then, based on the emission response residual sequence and the back-feedback disturbance residual sequence, a path difference spectrum is constructed to extract the local peak set, path difference discreteness, dominant oscillation frequency, and coherent back-feedback cluster intensity index. This separates the back-feedback effect caused by multiple weak reflections from the surface-stable power change, and further reveals the concentration of multiple main back-feedback paths in the time delay distribution and the degree of synergy in the dominant oscillation rhythm. It can more accurately identify the latent back-feedback disturbances and instability trends inside the light source, improve the ability to sense latent anomalies under surface-stable conditions, and enhance the identification effect of early risks in complex optical communication links.

[0007] 2. This invention further utilizes forward-sampled optical power sequences, monitored photocurrent sequences, coherent back-return cluster intensity indices, and path difference discrepancies to apply a constrained micro-probe current waveform at the current operating point through reinforcement learning. It then determines the probe gain by combining the coherent back-return cluster intensity indices before and after probe, and calculates the pseudo-stability risk value by combining apparent stability metrics. The invention outputs normal state, latent back-return cluster state, and explicit instability state. By introducing an active probe mechanism without significantly disrupting normal service transmission, the impact of hidden back-return clusters can be amplified. This allows risk assessment to not only reflect the already manifested instability process but also to reveal instability evolution in the latent stage, thereby improving the sensitivity, accuracy, and timeliness of monitoring results. Attached Figure Description

[0008] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for monitoring the light source of an optical communication link according to an embodiment of the present invention. Figure 2 This is a functional block diagram of an optical communication link light source monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0010] Example: This example provides a method for monitoring the light source of an optical communication link. See [link to example]. Figure 1 Specifically, including: S1. Collect the emission drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence and source-side back sensing sequence of the light source, remove the common-mode response, and obtain the emission response residual sequence and the back disturbance residual sequence. S2. Construct the path difference spectrum based on the luminescence response residual sequence and the return perturbation residual sequence and extract the local peak set to calculate the path difference discreteness. S3. Extract the dominant oscillation frequency from the luminescence response residual sequence, and calculate the coherent return cluster intensity index by combining the local peak set and the path difference discrepancy. S4. Based on the forward sampling optical power sequence, monitoring photocurrent sequence, coherent back cluster intensity index and path difference discreteness, a constrained micro-probe current waveform is applied at the current operating point through reinforcement learning, and the detection gain is calculated based on the coherent back cluster intensity index before and after detection. S5. Calculate the apparent stability measure based on the forward sampling optical power sequence and the monitoring photocurrent sequence, combine the coherent return cluster intensity index and the detection gain to calculate the pseudo-stability and instability risk value, and output the monitoring results.

[0011] In embodiments of the present invention, the emission driving current sequence, monitoring photocurrent sequence, forward-sampled optical power sequence, and source-side back-sensing sequence of the light source are acquired, and the common-mode response is removed to obtain the emission response residual sequence and the back-perturbation residual sequence, including: In a 100G silicon photonics emission link that includes two LC-type fiber optic connectors and a set of chip edge coupling interfaces, this embodiment uses a distributed feedback laser light source as the emission light source, and is equipped with an emission drive circuit, a backlight monitoring photodiode, a forward sampling light extraction structure, a source-side return sensing circulator branch, and a four-channel synchronous high-speed analog-to-digital converter. When the link is in a normal service transmission working state, the laser light source operates within the rated drive current and rated output optical power range. The duration of a single monitoring window is set to 10 microseconds, the sampling rate of the four-channel synchronous high-speed analog-to-digital converter is set to 10 GSa / s, the single-channel analog bandwidth is not less than 20 GHz, the linear dynamic range is not less than 60 dB, the number of sampling points for each sequence in a single monitoring window is 100,000, all acquisition channels share the same 100 MHz high-precision temperature-controlled reference clock and use a synchronous edge-triggered mechanism. Before acquisition, the inherent time delay deviation between channels is eliminated through coaxial cable time delay matching and digital calibration, ensuring that the four acquisition channels are in the same clock domain and the synchronization deviation between channels does not exceed 50 picoseconds, and the sampling moments are completely aligned, eliminating the influence of timing deviation on subsequent cross-correlation calculations. The isolation degree from the emitter end to the return receiver end of the circulator used in the source-side return sensing branch is not less than 45 dB, the response bandwidth of the supporting high-speed photodetector is not less than 20 GHz, the forward sampling light extraction structure uses a fused biconical taper optical splitter with a splitting ratio of 1:99, and the insertion loss of the optical splitter does not exceed 0.5 dB. The bandwidth of the transimpedance amplifier supporting the backlight monitoring photodiode is not less than 20 GHz, and the linear output range covers the backlight current range corresponding to the rated working state of the light source. Within a single monitoring window, the emission drive current sequence is acquired through a high-precision milliohm-level sampling resistor connected in series between the emission drive circuit and the anode of the laser light source, in cooperation with the four-channel synchronous high-speed analog-to-digital converter. The monitored photocurrent sequence is synchronously acquired through the backlight monitoring photodiode integrated at the end of the laser light source, in cooperation with the transimpedance amplifier and the four-channel synchronous high-speed analog-to-digital converter. The forward sampling optical power sequence is synchronously acquired through the forward sampling light extraction structure in the optical path, in cooperation with the photodetection module and the four-channel synchronous high-speed analog-to-digital converter. The source-side return sensing sequence is synchronously acquired through the isolation branch of the source-side return sensing circulator built into the transmitter end, in cooperation with the high-speed photodetector and the four-channel synchronous high-speed analog-to-digital converter.

[0012] After the synchronous acquisition of the four sequences is completed, effective data preprocessing is first performed on each acquired sequence to remove saturated sampled values ​​that exceed the linear range of the analog-to-digital converter, retaining effective sampled points within the linear operating range. Then, robust normalization is performed on the preprocessed transmit drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence, and source-side back-feedback sensing sequence. For a single sequence to be processed, the median of all effective sampled values ​​in the sequence is first calculated, and then the median of the absolute values ​​of the differences between each effective sampled value and the median of the sequence is calculated to obtain the absolute median difference of the sequence. For each effective sampled point in the sequence, the median of the sequence is subtracted from the value of that sampled point. The number is then divided by the sum of the absolute median difference and the minimum value of the sequence. The minimum value to prevent the denominator from being zero is set to 0.000001 to complete the normalization calculation for the sampling point. After traversing all valid sampling points of the sequence, the normalized sequence corresponding to the sequence is obtained. After completing the robust normalization processing of the four sequences in sequence, the normalized sequences corresponding to the emission drive current sequence, the monitoring photocurrent sequence, the forward sampling optical power sequence, and the source-side back-feedback sensing sequence are obtained respectively. In this embodiment, the median and the absolute median difference are used to complete the robust normalization processing, which can effectively suppress the interference of outliers in the sampling process on the normalization results and improve the anti-interference ability and robustness of the results in the subsequent data processing process.

[0013] It should be noted that the emission drive current sequence refers to the ordered data set formed by the change of the drive current applied to the active region of the light source over time during a continuous monitoring period. It reflects the direct electrical input state of the light source when it is excited to emit light. The changes in the current magnitude in this sequence affect the carrier injection level, the stimulated emission buildup process, and the light output intensity. Therefore, this sequence not only characterizes the workload of the emitter but also the dynamic changes in the internal luminescence conditions of the light source, serving as a fundamental reference quantity for distinguishing between normal drive fluctuations and abnormal back-feedback disturbances. The monitoring photocurrent sequence refers to the arrangement of the current signal generated by the monitoring photoelectric device after receiving the light emitted by the light source over continuous time. Essentially, it reflects the change in the actual light emission state of the light source over time. Since the stronger the light received by the photoelectric device, the larger the photocurrent generated is usually, this sequence can characterize the light source output energy, short-term fluctuations, and signs of local instability. Furthermore, this sequence is simultaneously affected by drive changes and optical path disturbances, making it suitable as an observation to describe the light emission response of the light source.

[0014] It should be noted that the forward sampling optical power sequence refers to the ordered data set formed by the change of optical power over time after extracting a portion of the light energy from the forward emission main optical path of the light source. It reflects the intensity and stability of the light energy output along the normal transmission direction. This sequence corresponds to the change of effective emission energy entering the subsequent communication optical path, and can characterize whether the output on the link surface remains stable. Therefore, it can be used to determine whether the current light source is in a state of near-stable average power but with latent anomalies inside. The source-side back-feedback sensing sequence refers to the time sequence obtained after the sensing device arranged near the emission side of the light source detects the disturbance signal returned to the vicinity of the emission end. It reflects the intensity and change process of the light disturbance returned to the source side from the connector end face, coupling interface, chip light-emitting end face, or other reflection positions. This sequence essentially describes the dynamic degree of the re-injection effect of the external optical path on the light source, and can characterize whether multiple weak reflections are superimposed in time, thus providing a direct basis for identifying the state of latent back-feedback clusters.

[0015] Within the same monitoring window of the aforementioned 100G silicon photonics transmission link, based on the obtained normalized transmit drive current sequence, normalized monitoring photocurrent sequence, and normalized source-side feedback sensing sequence, the first scaling factor and the second scaling factor are solved sequentially, and the residual sequence is calculated. The first scaling factor is used to characterize the linear common-mode contribution of the normalized transmit drive current sequence to the normalized monitoring photocurrent sequence. The goal is to minimize the absolute error between the normalized monitoring photocurrent sequence and the scaled normalized transmit drive current sequence. The corresponding calculation formula is as follows: , This is the first proportional coefficient corresponding to the current monitoring window. Let be the proportionality coefficient variable to be solved, and t be the sampling point number within the monitoring window. This represents the normalized monitoring photocurrent sequence value corresponding to the t-th sampling point within the current monitoring window. Let be the normalized transmit drive current sequence value corresponding to the t-th sampling point within the current monitoring window. The summation operation covers all valid sampling points from the first to the last within the monitoring window. This step uses absolute error minimization instead of conventional least squares fitting because absolute error minimization has a stronger anti-interference capability against outliers and impulse interference during the sampling process. This is consistent with the anti-interference design logic of the aforementioned robust normalization processing, and can avoid the fitting result deviation caused by extreme outliers. It can more accurately isolate the common-mode change of the monitoring photocurrent directly caused by the change in drive current. The solution process is completed using the iterative weighted least squares numerical solution algorithm commonly used in the field of optical communication digital signal processing. The initial value of the scaling factor is set to zero. The iteration process updates in the direction of decreasing absolute error summation. The iteration convergence threshold is set to... The iteration stops when the change in the scaling factor is less than the convergence threshold, and the first scaling factor after convergence is output. Then, the second scaling factor is calculated using the same solution logic. The second scaling factor characterizes the linear common-mode contribution of the normalized emitter-driven current sequence to the normalized source-side back-feedback sensing sequence. The objective is to minimize the absolute error between the normalized source-side back-feedback sensing sequence and the scaled normalized emitter-driven current sequence. The corresponding calculation formula is as follows: , This is the second proportional coefficient corresponding to the current monitoring window. Let the proportionality coefficient variable be the one to be solved. This is the normalized source-side feedback sensing sequence value corresponding to the t-th sampling point within the current monitoring window. The meanings of the other parameters are consistent with the aforementioned formula. The solution process is also completed using the iterative weighted least squares method, and the initial value, convergence threshold, and first proportional coefficient are solved in the same way.

[0016] After solving for the two scaling factors, the luminescence response residual sequence and the return perturbation residual sequence are calculated. The luminescence response residual sequence is calculated using the following formula: , Let be the residual sequence value of the emission response corresponding to the t-th sampling point within the current monitoring window. This calculation process removes the normalized emission drive current sequence scaled by the first proportional coefficient point by point from the normalized monitoring photocurrent sequence, completely eliminating the linear common-mode response of the monitoring photocurrent directly caused by the change in emission drive current. The resulting residual sequence of emission response retains only the non-common-mode signal related to the change in the internal emission state of the light source and the return light re-injection disturbance, eliminating the interference of synchronous changes in drive current on subsequent disturbance analysis; the return disturbance residual sequence is calculated using the following formula: , The return disturbance residual sequence value is the sampling point t in the current monitoring window. This calculation process removes the normalized transmit drive current sequence scaled by the second scaling factor from the normalized source-side return sensing sequence point by point, thus removing the common-mode change of the return signal caused by the output power fluctuation due to the change of the transmit drive current. The resulting return disturbance residual sequence retains only the signal related to the specific return disturbance caused by the multi-interface reflection of the link, providing two pure input sequences related only to the return disturbance for the subsequent construction of the path difference spectrum. In this embodiment, all calculation processes are completed with 32-bit floating-point precision to avoid the numerical truncation error in the calculation process from affecting the accuracy of the residual sequence.

[0017] It should be noted that the first scaling factor is used to characterize the contribution of the normalized emission drive current sequence to the common behavior in the normalized monitoring photocurrent sequence. It corresponds to the matching strength that minimizes the overall absolute error between the two within the current monitoring window. This factor reflects the proportion of the monitoring photocurrent directly caused by normal drive changes. Therefore, scaling the normalized emission drive current sequence using this factor allows the synchronization component directly caused by the drive to be subtracted from the monitoring photocurrent, making the remaining part more representative of abnormal luminescence changes not explained by conventional drive. The second scaling factor is used to characterize the contribution of the normalized emission drive current sequence to the common changes in the normalized source-side back-sensing sequence. It corresponds to the matching strength that minimizes the overall absolute error between the two within the current monitoring window. This factor reflects how much of the change in the source-side back-sensing signal can be directly explained by normal drive fluctuations. Therefore, scaling the normalized emission drive current sequence using this factor and removing it from the source-side back-sensing sequence retains the remaining changes that better represent back propagation, interface reflection, and re-injection perturbations.

[0018] It should be noted that the luminescence response residual sequence refers to the time series obtained by removing the normalized emission drive current sequence scaled by the first proportional factor from the normalized monitoring photocurrent sequence. It represents the remaining response in the monitoring photocurrent that cannot be directly explained by normal drive changes. This sequence reflects the degree of deviation of the actual luminescence behavior of the light source from the conventional drive relationship. Therefore, when the light source is affected by back-feedback disturbances, mode fluctuations, or local instability, these changes not directly explained by the drive will appear in this sequence, making it an important observation for describing abnormal luminescence responses. The back-feedback disturbance residual sequence refers to the time series obtained by removing the normalized emission drive current sequence scaled by the second proportional factor from the normalized source-side back-feedback sensing sequence. It represents the remaining disturbance portion in the source-side back-feedback sensing signal that cannot be directly explained by normal drive changes. This sequence mainly reflects the re-injection effect and dynamic changes of the back-feedback from the connector end face, coupling interface, chip emission end face, or other reflection positions back to the source side. Therefore, it can more effectively characterize the strength and timing characteristics of the back-feedback disturbances formed by the external optical path on the light source.

[0019] In an embodiment of the present invention, the dominant oscillation frequency is extracted from the luminescence response residual sequence, and the coherent return cluster intensity index is calculated by combining the local peak set and the path difference discrepancy. Hanning windows are applied to the luminescence response residual sequence and the back-disturbance residual sequence to suppress spectral leakage. Then, a Discrete Fast Fourier Transform (DFT) is performed to obtain the corresponding first and second frequency domain results. The DFT algorithm uses a power-of-two (HOF) algorithm with a number of points, padding the sequence length to 100,000 points, matching the number of sampling points in the monitoring window. Zero-padding is used at the tail. The system sampling rate is set to 10 GSa per second. The frequency resolution and time delay resolution are uniquely determined by the sampling rate and sequence length. This transform converts the time-domain residual signal to the frequency domain, facilitating the extraction of phase and time delay features. The corresponding frequency domain transformation formula is: , , This is the first frequency domain result corresponding to the luminescence response residual sequence. This is the second frequency domain result corresponding to the returned perturbation residual sequence. For Discrete Fast Fourier Transform operators, For frequency domain variables, It is the residual sequence of luminescence response. The result is the return perturbation residual sequence, where t is the sampling point number and n is the current monitoring window number. A phase transform generalized cross-correlation function is constructed by normalizing the conjugate product of the first and second frequency domain results. This function retains only the phase information of the two frequency domain signals and eliminates amplitude interference, accurately extracting the time delay characteristics between the two signals and avoiding the impact of amplitude fluctuations on path difference calculation. The corresponding calculation formula is: In the formula, For phase transform generalized cross-correlation function, Let be the conjugate complex number of the second frequency domain result. To prevent extremely small quantities with a denominator of zero, the meanings of the remaining parameters remain consistent with the aforementioned formula. After constructing the generalized cross-correlation function of the phase transformation, a discrete fast Fourier inverse transform is performed to obtain the path difference spectrum. The inverse transform can convert the frequency domain phase correlation characteristics into the time domain path difference distribution, facilitating the direct extraction of path difference information from multiple reflection paths. The corresponding calculation formula is as follows: , For path difference spectrum, For the discrete inverse fast Fourier transform operator, The round-trip time delay variable corresponding to the path difference is defined, and the meanings of the other parameters remain consistent with the aforementioned formula. Local peak values ​​and their corresponding path differences are extracted within the positive time delay interval of zero to one thousand picoseconds in the path difference spectrum to form a local peak set. This interval covers the time delay range of all effective reflection paths in this link. The local peak extraction adopts the five-sampling-point neighborhood maximum determination method, traversing all path difference spectrum values ​​within the positive time delay region. Spectral values ​​that are greater than the values ​​of the five adjacent sampling points and whose amplitudes are greater than three times the global mean of the path difference spectrum are determined as effective local peaks. The amplitude of each effective local peak and its corresponding path difference are recorded. The amplitudes of all effective local peaks and their path differences together constitute the local peak set of the current monitoring window.

[0020] It should be noted that the first frequency domain result refers to the complex distribution result obtained after frequency domain transformation of the luminescence response residual sequence. This reflects the strength, phase order, and energy distribution along the frequency axis of each frequency component in the luminescence response residual. This result no longer directly represents the temporal fluctuations but rather separates the different oscillation components originally mixed in the time domain to different frequency positions. This facilitates the identification of which frequency components are related to abnormal luminescence changes of the light source and provides a foundation for subsequent analysis of the correspondence between the luminescence response and the back-return perturbation. The second frequency domain result refers to the complex distribution result obtained after frequency domain transformation of the back-return perturbation residual sequence. This reflects the amplitude and phase state of different frequency components in the back-return perturbation residual. This result characterizes the decomposition of the source-side back-return perturbation within the frequency range, allowing the back-return fluctuations that were originally intertwined over time to be separated into multiple distinguishable frequency components. This helps in subsequent analysis of the correspondence and propagation timing relationships between these perturbation components and the luminescence response components.

[0021] It should be noted that the conjugate product refers to the complex product formed by multiplying the conjugate results of the first and second frequency domain results frequency by frequency. It reflects the consistency and sequential relationship between the luminous response residual and the return perturbation residual at various frequency positions. The phase transform generalized cross-correlation function is a frequency domain correlation function constructed from the normalized conjugate product. Its function is to uniformly represent the correspondence between the luminous response residual and the return perturbation residual at various frequencies. This function prominently preserves the phase difference information between the two signals, making it easier to reveal the time offset caused by the propagation path and the return path when subsequently transformed to the time delay domain. Therefore, it is suitable for extracting the time delay position corresponding to the main path difference from complex return scenarios.

[0022] It should be noted that the path difference spectrum refers to the distribution result formed on the time delay axis after the inverse transformation of the generalized cross-correlation function of the phase transform. It reflects the correlation strength between the residual of the luminous response and the residual of the return disturbance at different equivalent path difference positions. The stronger the response at a certain time delay position in the spectrum, the higher the consistency of the two signals near the corresponding propagation time difference, which also means that there is a more obvious propagation path difference or return path difference. Therefore, the path difference spectrum can be used to characterize the time difference structure corresponding to multiple reflection positions in the link. The positive time delay region refers to the region on one side of the path difference spectrum corresponding to the positive time offset. It represents the situation where the return disturbance arrives with a time lag relative to the luminous response. Since the return disturbance needs to go through the process of being emitted from the source side, propagating along the link, being reflected at the interface, and returning to the vicinity of the source side, it usually shows a lag change compared to the luminous response. Therefore, when extracting the main return path difference, the positive time delay region can more effectively reflect the return delay information with actual propagation significance.

[0023] It should be noted that a local peak refers to a point in the path difference spectrum that has a relatively larger response amplitude at its nearest time delay position. It reflects that the correlation corresponding to a specific path difference is most significant within a local range. The higher the local peak, the stronger the consistency between the two signals near the time offset corresponding to that path difference. This also means that the propagation path or return path corresponding to that time delay in the link contributes more significantly to the current observation results. Therefore, local peaks can serve as an important basis for identifying the main path difference components. The local peak set refers to an ordered data set composed of multiple local peaks extracted within the positive time delay region of the path difference spectrum and their corresponding path differences. It reflects the overall distribution of all main return path difference components within the current monitoring window. This set not only retains the positional information of each path difference but also the corresponding response strength information. Therefore, it can serve as a direct input for subsequent calculations of path difference discreteness, path difference clustering strength, and phase consistency coefficient, comprehensively characterizing whether multiple return paths exhibit clustering and locking phenomena.

[0024] The weighted path difference center is calculated based on the amplitude of each local peak in the local peak set and its corresponding path difference. Using the amplitude of the local peak as the weighting factor can highlight the dominant reflection path with stronger return signal, making the calculated center value more consistent with the core path difference distribution that actually disturbs the light source. The corresponding calculation formula is as follows: In the formula, The weighted path difference center corresponding to the current monitoring window. Let k be the total number of valid local peaks in the set of local peaks, and k be the index of the valid local peak. Let be the amplitude of the k-th effective local peak in the set of local peaks. The path difference corresponding to the k-th valid local peak in the local peak set is summed to cover all valid local peaks within the set. The absolute deviation of each path difference relative to the weighted path difference center is calculated, and then the amplitude corresponding to each local peak is used as a weight to weight and sum the absolute deviations of all path differences, ultimately obtaining the path difference dispersion. This calculation method can weaken the influence of low-amplitude noise peaks and accurately reflect the degree of path difference clustering of the dominant reflection path. The smaller the path difference dispersion, the more concentrated the path differences of multiple reflection paths. The corresponding calculation formula is: middle, This represents the path difference dispersion corresponding to the current monitoring window. To prevent extremely small quantities where the denominator is zero, the meanings of the remaining parameters remain consistent with the aforementioned formula for calculating the weighted path difference center.

[0025] It should be noted that the weighted path difference center is a central location determined by the path difference positions and amplitudes corresponding to each local peak in the local peak set. It reflects the overall concentration of multiple major return path difference components within the current monitoring window. Since a larger local peak amplitude indicates a more significant contribution of the corresponding path difference to the current related structure, using amplitude as the weight to determine the center position yields a result that better represents the actual clustering core of the major return path difference components, rather than simply treating all path differences equally. This quantity characterizes the dominant center of the current return path structure in terms of time delay distribution and is an important reference for subsequently determining whether path differences cluster around the same area.

[0026] It should be noted that the degree of deviation refers to the distance of each path difference relative to the center of the weighted path difference, reflecting the degree of dispersion of a certain return path difference component relative to the overall dominant clustering area. If the distance between a certain path difference and the center of the weighted path difference is small, it means that the path difference is closer to the main return path group and belongs to the core component of the clustering structure; if the distance is large, it means that the path difference is more deviated from the current dominant clustering area, possibly corresponding to a weaker or more dispersed return component. By calculating the degree of deviation of each path difference, the path difference set can be transformed from a simple location distribution into a quantifiable description of the clustering density.

[0027] It should be noted that the path difference dispersion refers to the quantity obtained by comprehensively summarizing the deviations of all path differences from the weighted path difference center, using the amplitude of each local peak as a weight. It reflects the degree of dispersion or clustering of the main return path difference components within the current monitoring window. If most of the path differences are distributed around the same central area, the path difference dispersion is small, indicating that the return path structure is more concentrated and it is easier to form a path difference locking phenomenon. If the path differences are scattered and far apart, the path difference dispersion is large, indicating that there is no obvious clustering relationship between different return paths. This quantity is an important basic indicator for characterizing the clustering intensity of path differences and identifying coherent return clusters.

[0028] In an embodiment of the present invention, the dominant oscillation frequency is extracted from the luminescence response residual sequence, and the coherent return cluster intensity index is calculated by combining the local peak set and the path difference discrepancy. The power spectrum of a single-sided positive frequency range is constructed based on the frequency domain results of the luminescence response residual sequence. This power spectrum reflects the energy distribution of the luminescence response residual signal at different frequencies. The single-sided positive frequency range is set to 100MHz to 5GHz, covering the effective range of the intrinsic oscillation of the distributed feedback laser source in this link. Simultaneously, DC components, negative frequency components, and high-frequency aliasing noise are removed. The corresponding calculation formula is as follows: , This is the one-sided power spectrum corresponding to the residual sequence of the emission response in the current monitoring window. The frequency domain result corresponding to the luminescence response residual sequence is given, where f is the frequency domain variable and n is the current monitoring window number. The dominant oscillation frequency is extracted from the single-sided power spectrum. The extraction process is completed by maximizing the ratio of the power spectrum value to the global median in the frequency dimension. The median calculation range is consistent with the effective frequency range of the single-sided power spectrum. This method can suppress frequency domain noise and sidelobe interference, and accurately locate the core intrinsic oscillation frequency of the light source. The corresponding calculation formula is: In the formula, This represents the dominant oscillation frequency within the current monitoring window. This represents the median of the one-sided power spectrum along the effective frequency dimension. To prevent extremely small values ​​with a denominator of zero, the meanings of the remaining parameters remain consistent with the aforementioned formula. The phase consistency coefficient is calculated based on the dominant oscillation frequency and the path differences within the local peak set. During the calculation, the unit of the path difference is first converted from picoseconds to seconds to match the Hertz unit of the dominant oscillation frequency. Phase calculation uses radians, and the imaginary unit j satisfies the engineering definition that the square of j equals negative one. This coefficient measures the degree of phase synchronization of each reflection path at the dominant oscillation frequency. Higher phase consistency results in a more significant coherent superposition effect and a stronger disturbance to the light source. The corresponding calculation formula is: In the formula, The phase consistency coefficient for the current monitoring window. Let k be the total number of valid local peaks in the set of local peaks, and k be the index of the valid local peak. Let the amplitude of the kth effective local peak be . The summation operation covers all valid local peaks within the set of local peaks after unit conversion. The meanings of the remaining parameters are consistent with the aforementioned formula.

[0029] The path difference clustering intensity is calculated based on the amplitude of each local peak in the local peak set and the path difference discrepancy. The smaller the path difference discrepancy, the higher the degree of path difference clustering, and the greater the corresponding clustering intensity. The corresponding calculation formula is as follows: , The path difference aggregation intensity of the current monitoring window. This represents the path difference discreteness corresponding to the current monitoring window. The meanings of the other parameters remain consistent with the aforementioned formula. Finally, the path difference clustering strength is multiplied by the phase consistency coefficient to obtain the coherent return cluster strength index. This index integrates the path difference clustering characteristics and phase coherence characteristics of multiple reflection paths, and can directly quantify the strength of the coherent return cluster. The larger the value, the closer the link is to the coherent return cluster state locked by the reflection path difference. The corresponding calculation formula is as follows: , This represents the coherent return cluster intensity index for the current monitoring window; the meanings of the other parameters remain consistent with the aforementioned formula.

[0030] It should be noted that the dominant oscillation frequency refers to the frequency position in the power spectrum corresponding to the luminous response residual sequence that best represents the main periodic fluctuation within the current monitoring window. It reflects the most significant oscillation rhythm of the current luminous emission change of the light source. The oscillation component corresponding to this frequency usually contributes the most to the output fluctuation of the light source. Therefore, it can be used as a reference benchmark to measure whether the return disturbance has a consistent effect with the main oscillation process of the light source. Using this frequency as the analysis base point, it is possible to determine whether the return components corresponding to different return path differences show a synergistic superposition trend under the dominant oscillation rhythm.

[0031] It should be noted that the phase consistency coefficient refers to the degree of consistency of the phase relationship of each path difference in a local peak set at the dominant oscillation frequency. It reflects whether different return path difference components act together in similar directions on the oscillation rhythm. If the phase relationship corresponding to each path difference is relatively consistent, it means that these return components are more likely to form a unidirectional superposition at the dominant oscillation frequency, thereby enhancing the common influence on the light emission state of the light source. If the phases corresponding to each path difference are dispersed, it means that different return components cancel each other out or lack synergy. This coefficient is used to characterize the strength of synergy of multiple return paths on the frequency rhythm and is an important quantity for identifying coherent return phenomena.

[0032] It should be noted that path difference clustering intensity refers to the degree of concentration determined by the magnitude of each local peak in the local peak set and the path difference dispersion. It reflects the degree of clustering and effective contribution of the main return path difference components in the time delay distribution within the current monitoring window. If the local peaks are strong overall and the corresponding path difference distribution is relatively concentrated, the path difference clustering intensity is large, indicating that multiple main return components tend to form a dominant structure around similar path difference regions. If the local peaks are weak or the path difference distribution is relatively dispersed, the path difference clustering intensity is small. This quantity is used to characterize the concentration trend of multiple main return paths in the path difference dimension and is an important basis for judging whether there is a path difference locking phenomenon.

[0033] It should be noted that the coherent return cluster intensity index is a comprehensive quantity determined by the path difference aggregation intensity and the phase consistency coefficient. It reflects whether multiple major return path difference components within the current monitoring window simultaneously possess both path difference concentration and oscillation phase coordination. The larger the index, the more likely multiple return paths are not only close in path difference location but also more likely to form in-direction superposition at the dominant oscillation frequency, thus making them more likely to act together on the light source and amplify the effects of latent instability. When the index is small, it indicates that the return paths are either scattered or have inconsistent phase relationships, making it difficult to form a significant coherent return cluster. Therefore, this index can be used as a core monitoring quantity to characterize the strength of the return cluster state.

[0034] The current monitoring state is defined as a four-dimensional set of state parameters. The four parameters are the coherent back-feed cluster intensity index, path difference dispersion, absolute median difference of the forward sampled optical power sequence, and absolute median difference of the monitored photocurrent sequence for the current monitoring window. This state set simultaneously covers both the core link instability risk characteristics and the apparent operational stability characteristics, providing comprehensive state input for micro-probe action selection. The corresponding state vector expression is: , This is the state vector of the current monitoring window. To detect the coherent return cluster intensity index of the current monitoring window, This represents the path difference dispersion for the current monitoring window. This represents the absolute median difference of the forward-sampled optical power sequence within the current monitoring window. The absolute median difference of the photocurrent sequence monitored in the current monitoring window is denoted by n, where n is the current monitoring window number.

[0035] Before inputting the state vector into the policy network, the maximum and minimum values ​​of each parameter from 100 historical monitoring windows are linearly normalized to map all parameters to the range of 0 to 1, ensuring the consistency and stability of the policy network input. Then, based on a pre-trained two-layer fully connected policy network, a constrained micro-probe current waveform is generated and applied as a micro-probe action at the current operating point. The input layer dimension of the policy network is 4, matching the state vector dimension; the hidden layer dimension is 16 and uses the ReLU activation function; the output layer dimension matches the number of sampling points within the monitoring window and uses the tanh activation function. The output result is scaled to a preset perturbation amplitude range, with the upper limit of the perturbation amplitude set to one percent of the rated operating current of the light source. The fundamental frequency of the micro-probe current waveform is set to less than one-tenth of the link service modulation frequency, ensuring that the micro-probe action does not cause frequency aliasing with the service signal and does not significantly disturb the normal operation of the link. For transmission, the allowable operating range of the distributed feedback laser source in this link is set to 10mA to 80mA. The micro-probe current waveform output by the policy network is synchronously superimposed on the original transmit drive current sequence at each sampling point, with the timing completely aligned with the monitoring window duration. After superposition, the total current value is verified point by point, and the current values ​​of sampling points exceeding the allowable operating range are clamped to the range boundary to ensure that the total transmit drive current after applying the micro-probe current waveform is always within the allowable operating range. The policy network is pre-trained through reinforcement learning. The training process uses the optical communication link simulation environment corresponding to this link. The training data covers the full-scene link data in normal state, latent back-cluster state, and explicit instability state. The reward function used in the training takes into account both the back-cluster amplification effect and the transmission disturbance suppression. The number of training iterations is set to 100,000. When the change in the rolling mean of the reward function after 1,000 consecutive iterations is less than 0.0001, it is determined to be converged.

[0036] While applying the micro-probe current waveform, a four-channel synchronous acquisition process identical to that before detection is employed. This process simultaneously acquires the transmit drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence, and source-side back-feedback sensing sequence. Strictly adhering to all calculation parameters and rules from before detection, the entire process of residual sequence construction, path difference spectrum calculation, local peak extraction, and coherent back-feedback cluster intensity index calculation is repeated to obtain the coherent back-feedback cluster intensity index after the micro-probe action. Finally, the detection gain is determined based on the difference between the coherent back-feedback cluster intensity index after detection and the coherent back-feedback cluster intensity index before detection. The detection gain directly reflects the amplification effect of the micro-probe action on the hidden coherent back-feedback cluster phenomenon; a larger gain value indicates a higher risk of link latent instability. The corresponding calculation formula is as follows: , The detection gain for the current monitoring window. The coherent return cluster intensity index is obtained after applying the microprobe action. The coherent return cluster intensity index is the current monitoring window before detection; the meanings of the other parameters are consistent with the aforementioned state vector formula.

[0037] It should be noted that the state parameter set refers to a combination of parameters used to characterize the operating state and return disturbance state of the light source within the current monitoring window. It is composed of the coherent return cluster intensity index, path difference dispersion, absolute median difference of the forward sampling optical power sequence, and absolute median difference of the monitoring photocurrent sequence. This parameter set reflects, on the one hand, whether there is a concentrated and coordinated return effect of multiple main return paths, and on the other hand, the surface fluctuation level of the light source's forward output and monitoring response. Therefore, it can describe more completely whether the light source within the current window is in a state of surface stability and internal sensitivity, and provide a basis for subsequent selection of detection actions.

[0038] It should be noted that the policy network refers to a mapping structure that generates micro-probing actions based on the current set of state parameters. Its function is to establish a correspondence between the current monitoring state and the appropriate micro-perturbation methods. The network receives input information reflecting the strength of the return cluster, the dispersion of path difference, and the apparent fluctuation, and outputs a detection method that is more sensitive to the current light source and has controlled perturbation. This allows the system to selectively stimulate hidden return cluster responses without significantly disrupting normal transmission, thereby improving the observability of latent instability phenomena.

[0039] It should be noted that the constrained micro-probe current waveform refers to a small-amplitude, time-varying current change applied to the transmit drive current near the current operating point. Its amplitude, range of variation, and mode of action are constrained by the operational safety boundary. The purpose of this waveform is not to change the long-term operating mode of the light source, but rather to apply a controllable small perturbation to the light source without disrupting normal transmission conditions, in order to observe whether the return path cluster will amplify the response to this small perturbation. Therefore, this waveform must have sufficient excitation capability without causing the light source to deviate from its permissible operating state. A micro-probe action refers to a controlled, small perturbation operation selected and applied by the policy network based on the current set of state parameters. Its direct carrier can be the constrained micro-probe current waveform. Essentially, this action actively introduces a small, controllable change near the normal operating state of the light source to test whether the latent coherent return path clusters in the current link will exhibit a more significant enhancement trend. Therefore, this action does not serve the transmission service itself, but is an active detection method set up to expose hidden anomalies and improve anomaly detectability.

[0040] It should be noted that the detection gain refers to the increase in the coherent back-cluster intensity index after detection relative to the coherent back-cluster intensity index before detection. It reflects the amplification capability of micro-detection actions on hidden back-cluster phenomena. If the detection gain is large, it indicates that there is already a latent coherent back-cluster structure in the current link, and small detection disturbances can make it more obvious. If the detection gain is small or does not increase, it indicates that the back-cluster sensitivity in the current window is weak or there is no significant latent anomaly. Therefore, this quantity can be used as an important criterion to distinguish between ordinary stable state and latent unstable state.

[0041] The reward function calculation and policy network iteration are completed simultaneously. This embodiment adopts a trinomial reward function design. The core objective is to maximize the amplification effect of micro-probing actions on coherent backhaul clusters while minimizing the disturbance of micro-probing actions to normal service transmission on the link, ensuring that micro-probing actions are always within a safe and low-disturbance range. The corresponding reward function calculation formula is as follows: In the formula, This is the reward value corresponding to the current monitoring window. The coherent return cluster intensity index is obtained after applying the microprobe action. To detect the coherent return cluster intensity index of the current monitoring window, This is the absolute median difference of the coherent return cluster intensity index within the rolling historical 100 monitoring windows preceding the current monitoring window. If historical data from the previous 100 monitoring windows is insufficient, the calibration value under normal link operation is used as a substitute. To prevent extremely small values ​​with a denominator of zero, the first term of the reward function is a gain enhancement term, which characterizes the effect of micro-probe actions on the detection gain. It is normalized using the rolling historical absolute median difference, which eliminates numerical magnitude differences under different link operating states, ensuring the robustness of the reward function across the entire operating range and effectively suppressing the interference of outliers on reward calculation. The forward optical power perturbation weighting coefficient is set to 0.5. The driving current disturbance weighting coefficient is set to 0.3. The two weighting coefficients are obtained based on the service transmission sensitivity and light source driving safety boundary calibration of this 100G silicon photonics transmission link. They are used to balance the penalty intensity of different disturbance terms to ensure that the micro-probe action does not exceed the link safety working boundary. For other types of link scenarios, the weighting coefficients can be adjusted according to the same calibration logic. This is the forward-sampled optical power sequence obtained after applying the micro-probe action. This is the forward-sampled optical power sequence of the current monitoring window before detection. To detect the absolute median difference in the sequence of forward sampling optical power differences, The first term is the arithmetic mean of the absolute median difference of the forward sampling optical power sequence in the 100 historical monitoring windows preceding the current monitoring window. If the historical data of the first 100 monitoring windows is insufficient, the calibration value under normal link operation is used instead. The second term of the reward function is the forward optical power perturbation penalty term, which is used to characterize the impact of microprobe actions on the fluctuation of forward sampling optical power. The larger the value, the stronger the perturbation of the microprobe action on the normal output optical power of the link, and the greater the corresponding penalty. This is the total emitter drive current sequence after the micro-probe action is applied. To detect the emitter drive current sequence in the current monitoring window before detection, To detect the absolute median difference of the sequence of transmit drive current differences before and after, The reward function is the arithmetic mean of the absolute median difference of the transmit drive current sequence in each of the 100 historical monitoring windows preceding the current monitoring window. If the historical data of the first 100 monitoring windows is insufficient, the calibration value under normal link operation is used instead. The third term of the reward function is the drive current disturbance penalty term, which is used to characterize the impact of micro-probe actions on the transmit drive current fluctuation. The larger the value, the stronger the disturbance of the micro-probe action on the light source drive state, and the greater the corresponding penalty.

[0042] After calculating the reward function, a baseline-based single-step policy gradient algorithm is used to iteratively update the weight parameters of the policy network. The policy network employs a stochastic policy framework. For the continuous action space of the micro-probe current waveform, the output layer outputs the mean and variance of a Gaussian distribution. The mean is mapped to the perturbation amplitude range using a tanh activation function, and the variance is ensured to be non-negative using a softplus activation function. This determines the probability distribution of the actions, and the corresponding update formula is as follows: , These are the weight parameters of the policy network. Set the learning rate to [value]. , For the policy network in the current state Down Output Action policy gradient, The action vector corresponding to the micro-probe current waveform applied to the current monitoring window. This is the state vector of the current monitoring window, which is completely consistent with the set of state parameters defined in the preceding steps. The advantage function is used to reduce the variance during policy gradient updates and improve training stability. The formula for calculating the advantage function is: , As the baseline term, it is calculated using the arithmetic mean of the reward values ​​of the previous 100 monitoring windows in the rolling history. If the historical data of the previous 100 monitoring windows is insufficient, the calibration reward value under normal link operation is used as a substitute. The policy network iteratively updates in the direction of maximizing the reward function. This update logic can automatically increase the gain of the first term of the reward function while suppressing the negative impact of link disturbances corresponding to the second and third terms, perfectly matching the low-disturbance and high-sensitivity design goal of micro-probing actions. The policy network performs a single-step update after completing the probing process in each monitoring window. When the change in the rolling reward mean of 1000 consecutive monitoring windows is less than When the policy network converges, iterative updates are stopped to ensure that the micro-probing actions output by the policy network always adapt to the current working state of the link.

[0043] In embodiments of the present invention, an apparent stability metric is calculated based on the forward-sampled optical power sequence and the monitored photocurrent sequence. A pseudo-stability / instability risk value is calculated by combining the coherent return cluster intensity index and the detection gain, and the monitoring results are output, including: Apparent stability metrics are calculated based on the fluctuation levels of the forward-sampled optical power sequence and the monitored photocurrent sequence. The fluctuation level is quantified using the absolute median difference, consistent with the previous steps. The absolute median difference, as a robust statistic, effectively suppresses outlier interference during sampling and avoids misjudgments of apparent stability caused by occasional pulses. Before calculation, the absolute median differences of the two sequences are subjected to rolling history normalization to eliminate differences in dimensions and numerical magnitudes, ensuring a balanced contribution weight between the two sequences to the apparent stability metric. The apparent stability metric characterizes the apparent operational stability of the link; the smaller the sequence fluctuation, the larger the apparent stability metric value. The corresponding calculation formula is as follows: In the formula, As a measure of apparent stability for the current monitoring window, This represents the absolute median difference of the forward-sampled optical power sequence within the current monitoring window. This is the arithmetic mean of the absolute median differences of the forward-sampled optical power sequences over the previous 100 monitoring windows in the rolling history preceding the current monitoring window. The absolute median difference of the photocurrent sequence is monitored within the current monitoring window. This is the arithmetic mean of the absolute median difference of the monitored photocurrent sequences within the previous 100 monitoring windows in the rolling historical data. If the historical data for the previous 100 monitoring windows is insufficient, both rolling historical statistics are replaced with the calibration values ​​under normal link operation conditions. To prevent extremely small quantities with a denominator of zero, n is the current monitoring window number.

[0044] The pseudo-stability risk value for the current window is calculated based on the coherent return cluster strength index, probe gain, and apparent stability metric. This risk value integrates the link core instability risk, the amplification effect of latent risk, and the apparent stability level. It can accurately identify pseudo-stability states where the average power is nearly stable but internal instability disturbances have already occurred. The corresponding calculation formula is as follows: In the formula, This represents the pseudo-stability and instability risk value for the current monitoring window. This represents the coherent return cluster intensity index before detection within the current monitoring window. The detection gain for the current monitoring window. This method retains only the positively amplified detection gain, and only includes latent risk when the micro-detection action amplifies the coherent back-cluster phenomenon, thus avoiding interference from negative gain in risk calculation. This is the absolute median difference of the detection gain within the 100 historical monitoring windows preceding the current monitoring window. When historical data from the first 100 monitoring windows is insufficient, the calibration value under normal link operation is used as a substitute to robustly normalize the detection gain and eliminate the difference in numerical magnitude under different operating conditions. The meanings of the remaining parameters are consistent with the aforementioned apparent stability metric calculation formula.

[0045] After calculating the pseudo-stability and instability risk value, the pseudo-stability and instability risk value of the current window is adaptively segmented based on the rolling historical risk sequence to obtain the adaptive segmentation result. The rolling historical risk sequence uses the pseudo-stability and instability risk values ​​of the 1000 effective monitoring windows before the current monitoring window. The effective monitoring window is defined as a monitoring window that completes the entire process monitoring, has no data acquisition anomalies, and has no numerical overflow. Abnormal risk values ​​in the historical samples are removed using the 3-times absolute median difference criterion. When there are fewer than 1000 historical data in the initial stage, the calibrated risk values ​​under normal link operation are used to supplement the 1000 samples. All calibration values ​​are obtained by calibrating the corresponding parameters of the link under normal operation without external reflection and in rated normal operation for 100 consecutive monitoring windows. The adaptive segmentation adopts an unsupervised segmentation method based on kernel density estimation, using Gaussian kernels. The function performs kernel density estimation on the rolling historical risk sequence. The kernel bandwidth is calculated using the Silverman optimal bandwidth criterion, which is common in this field. It identifies the effective probability density peaks in the kernel density estimation results. The effective peaks are determined using the five-sampling-point neighborhood maxima method, and the peak amplitude must be greater than 1.5 times the global mean of the probability density. If more than 3 effective peaks are identified, the 3 effective peaks with the largest amplitudes are selected. If fewer than 3 effective peaks are identified, the calibrated risk values ​​of three states—link normal operation, latent back-to-back clusters, and explicit instability—are used to supplement the 3 peaks. The probability density valley between adjacent peaks is used as the segmentation threshold to divide the risk value into three continuous numerical intervals. This method does not require a preset fixed threshold and can adaptively adjust the segmentation boundary based on the historical working state of the link to adapt to changes in different link scenarios and working states.

[0046] The monitoring status of the current window is determined and output based on the adaptive segmentation results. The correspondence between the three numerical intervals and the monitoring status is anchored to the normal operation calibration risk value of the link. The interval containing the normal operation calibration risk value is the normal state interval. The remaining two intervals correspond to the latent return cluster state and the manifest instability state respectively, according to the risk value from low to high. The normal state means that there is no obvious coherent return cluster disturbance in the link and the working state is stable. The latent return cluster state means that the link has experienced coherent return cluster phenomenon with path difference aggregation. Micro-probe actions can significantly amplify this effect, but the apparent power remains stable and is in the instability latent stage. The manifest instability state means that the link has experienced strong coherent return cluster disturbance and the light source is in a manifest instability state. Finally, the pseudo-stability risk value, coherent return cluster intensity index, detection gain and monitoring status of the current monitoring window are output synchronously.

[0047] It should be noted that the apparent stability measure is a quantity calculated based on the fluctuation of the forward sampled optical power sequence and the fluctuation of the monitored photocurrent sequence. It reflects the stability of the light source in the signal that can be directly observed from the outside within the current window. The larger the value, the smaller the overall fluctuation of the forward output optical power and the monitored photocurrent, and the closer the link and the light source are to a stable state at the surface observation level. The smaller the value, the more obvious the unstable changes have been shown in the forward output or emission response. The role of the apparent stability measure is to distinguish between a window that appears to be stable on the surface and a window that has directly shown strong fluctuations, thereby providing a reference for identifying pseudo-stable and unstable states.

[0048] It should be noted that the pseudo-stability risk value is a comprehensive risk quantity calculated based on the coherent return cluster strength index, detection gain, and apparent stability measure. It reflects the degree to which the current window appears stable on the surface but has a latent tendency to become unstable internally. This value also takes into account the strength of the return cluster itself, the amplification capability under the action of micro-detection, and the level of stability observed on the surface. Therefore, when the return cluster is strong, the detection gain is large, and the appearance is relatively stable, the risk value will increase significantly, indicating that the current window is more likely to be in a state of hidden anomaly rather than true stability.

[0049] It should be noted that the normal state refers to a state where the pseudo-stability risk value of the current window is at a low level, and the coherent back-cluster phenomenon is not significant or difficult to be amplified by micro-probing actions. In this state, the forward output optical power and monitoring photocurrent are usually stable, the back-path difference structure does not have obvious aggregation and cooperative characteristics, and micro-probing actions will not cause a significant increase in the intensity of the coherent back-cluster. Therefore, this state indicates that the current light source and link do not show any noteworthy latent back-cluster anomalies. The latent back-cluster state refers to a state where the current window still maintains high stability on the surface observation, but the coherent back-cluster intensity index has increased and can be significantly amplified under the action of micro-probing actions. In this state, the forward optical power and monitoring photocurrent may still not show significant fluctuations, so it is easy to be regarded as normal from the perspective of conventional monitoring. However, multiple main back-path differences have shown aggregation and a cooperative trend in the dominant oscillation rhythm, indicating that there is a hidden sensitive structure inside the link. This state indicates that the current window is in the early stage of development from normal operation to overt instability, and has early warning value. Overt instability refers to a state where the pseudo-stability risk value of the current window is at a high level, the coherent back-tracking cluster phenomenon has been significantly enhanced, and this phenomenon can be further amplified under the action of micro-probes. In this state, not only is the concentration and synergistic effect of the internal back-tracking path cluster obvious, but observable unstable changes also gradually appear in the forward output optical power or monitoring photocurrent, indicating that latent anomalies have developed into a relatively clear instability trend. This state indicates that the link and light source have entered a high-risk operating stage requiring timely intervention, alarms, or maintenance.

[0050] It should be noted that, without departing from the core concept of this invention, the following alternative methods also exist in this embodiment: The source-side back-feeding sequence can be obtained from an independent back-feeding branch or from sensing optoelectronic devices located near the transmitter. The forward-sampled optical power sequence can be obtained through beam splitter sampling, on-chip light extraction structures, or other equivalent forward light extraction methods. The micro-probe current waveform can be a sinusoidal perturbation, a piecewise constant perturbation, a pulse perturbation, or other constrained waveforms, as long as the applied transmit drive current remains within the allowable operating range. The policy network can be a fully connected neural network or other function approximators suitable for small-dimensional state inputs. The normalization quantity in the reward function can be the rolling history absolute median difference or a rolling history robust scaling quantity, as long as it is used to achieve scaling uniformity between different dimensions. Adaptive segmentation of the monitoring state can be implemented using quantile methods, clustering methods, or rolling model methods.

[0051] like Figure 2 The diagram shown is a functional block diagram of an optical communication link light source monitoring system provided in an embodiment of the present invention.

[0052] In this embodiment, the functions of each module / unit are as follows: The residual extraction module is used to collect the emission driving current sequence, monitoring photocurrent sequence, forward sampling optical power sequence and source-side back sensing sequence of the light source, remove the common-mode response, and obtain the emission response residual sequence and the back disturbance residual sequence. The path analysis module is used to construct the path difference spectrum based on the luminescence response residual sequence and the return perturbation residual sequence, extract the local peak set, and calculate the path difference discrete quantity. The coherent evaluation module is used to extract the dominant oscillation frequency from the luminescence response residual sequence and calculate the coherent return cluster intensity index by combining the local peak set and the path difference discrepancy. The enhanced detection module is used to apply a constrained micro-detection current waveform at the current operating point based on the forward sampled optical power sequence, the monitored photocurrent sequence, the coherent back cluster intensity index, and the path difference discreteness, and to calculate the detection gain based on the coherent back cluster intensity index before and after detection. The risk assessment module is used to calculate the apparent stability measure based on the forward sampling optical power sequence and the monitoring photocurrent sequence, combine the coherent return cluster intensity index and the detection gain to calculate the pseudo-stability and instability risk value, and output the monitoring results.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the light source of an optical communication link, characterized in that, include: S1. Collect the emission drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence and source-side back sensing sequence of the light source, remove the common-mode response, and obtain the emission response residual sequence and the back disturbance residual sequence. S2. Construct the path difference spectrum based on the luminescence response residual sequence and the return perturbation residual sequence and extract the local peak set to calculate the path difference discreteness. S3. Extract the dominant oscillation frequency from the luminescence response residual sequence, and calculate the coherent return cluster intensity index by combining the local peak set and the path difference discrepancy. S4. Based on the forward sampling optical power sequence, monitoring photocurrent sequence, coherent back cluster intensity index and path difference discreteness, a constrained micro-probe current waveform is applied at the current operating point through reinforcement learning, and the detection gain is calculated based on the coherent back cluster intensity index before and after detection. S5. Calculate the apparent stability measure based on the forward sampling optical power sequence and the monitoring photocurrent sequence, combine the coherent return cluster intensity index and the detection gain to calculate the pseudo-stability and instability risk value, and output the monitoring results.

2. The method for monitoring the light source of an optical communication link according to claim 1, characterized in that, After removing the common-mode response, the luminescence response residual sequence and the return perturbation residual sequence are obtained, including: The acquired emission drive current sequence, monitoring photocurrent sequence, forward sampling optical power sequence, and source-side back-sensing sequence are robustly normalized to obtain the corresponding normalized sequences. The first proportionality coefficient is determined based on the result of minimizing the absolute error between the normalized emission drive current sequence and the normalized monitoring photocurrent sequence. The second proportional coefficient is determined based on the result of minimizing the absolute error between the normalized transmit drive current sequence and the normalized source-side back-feed sensing sequence. The emission drive current sequence scaled by the first proportional coefficient is removed from the normalized monitoring photocurrent sequence to obtain the emission response residual sequence. The normalized emitter drive current sequence, scaled by the second scaling factor, is removed from the normalized source-side return sensing sequence to obtain the return disturbance residual sequence.

3. The method for monitoring the light source of an optical communication link according to claim 1, characterized in that, A path difference spectrum is constructed based on the luminescence response residual sequence and the return perturbation residual sequence, and a set of local peaks is extracted, including: The luminescence response residual sequence and the back-current perturbation residual sequence are transformed in the frequency domain to obtain the corresponding first frequency domain result and second frequency domain result; Based on the conjugate product of the first frequency domain result and the second frequency domain result, a generalized cross-correlation function of phase transformation is constructed. The path difference spectrum is obtained by performing an inverse transform on the phase transform generalized cross-correlation function. Local peaks and their corresponding path differences are extracted within the positive time delay region of the path difference spectrum to form a set of local peaks.

4. The method for monitoring the light source of an optical communication link according to claim 3, characterized in that, The path difference discrete quantity is calculated, including: The weighted path difference center is calculated based on the amplitude of each local peak in the local peak set and the corresponding path difference. Calculate the degree of deviation of each path difference from the center of the weighted path difference; By using the corresponding local peak amplitude as weight, the deviation of each path difference is weighted and summed to obtain the path difference discrete quantity.

5. The method for monitoring the light source of an optical communication link according to claim 4, characterized in that, The dominant oscillation frequency is extracted from the luminescence response residual sequence. Combined with the local peak set and path difference discretization, the coherent return cluster intensity index is calculated, including: A power spectrum is constructed based on the frequency domain results of the luminescence response residual sequence, and the dominant oscillation frequency is extracted from the power spectrum; Based on the dominant oscillation frequency and the path differences in the local peak set, the phase consistency coefficient of the local peak set at the dominant oscillation frequency is calculated. The path difference clustering intensity is calculated based on the amplitude of each local peak in the local peak set and the path difference discrepancy. The coherent return cluster intensity index is obtained based on the path difference cluster intensity and the phase consistency coefficient.

6. The method for monitoring the light source of an optical communication link according to claim 1, characterized in that, Based on the forward-sampled optical power sequence, the monitored photocurrent sequence, the coherent return cluster intensity index, and the path difference discrepancy, a constrained micro-probe current waveform is applied at the current operating point through reinforcement learning. The detection gain is calculated based on the coherent return cluster intensity index before and after detection, including: The current monitoring state is defined as a set of state parameters, which consists of the coherent back cluster intensity index, path difference dispersion, absolute median difference of the forward sampling optical power sequence, and absolute median difference of the monitoring photocurrent sequence. The micro-probe action is described by applying a restricted micro-probe current waveform at the current operating point based on the policy network, wherein the transmit drive current after applying the micro-probe current waveform is within the allowable operating range. Based on the monitoring results after the micro-probing action is applied, the coherent return cluster intensity index after probing is obtained; The detection gain is determined based on the difference between the coherent return cluster intensity index after detection and the coherent return cluster intensity index before detection.

7. The method for monitoring the light source of an optical communication link according to claim 6, characterized in that, The method involves applying a constrained microprobe current waveform at the current operating point using reinforcement learning, and also includes updating the policy network based on a reward function. The reward function includes a first term, a second term, and a third term; The first item is used to characterize the effect of the micro-detection action on the detection gain; The second item is used to characterize the effect of the micro-probe action on the fluctuation of the forward sampling optical power; The third item is used to characterize the effect of the micro-probe action on the fluctuation of the transmission drive current; The policy network is updated in the direction of increasing the first term and suppressing the corresponding effects of the second and third terms.

8. The method for monitoring the light source of an optical communication link according to claim 1, characterized in that, The pseudo-stability and instability risk values ​​are calculated by combining the coherent return cluster intensity index and the detection gain, and the monitoring results are output, including: The apparent stability metric is calculated based on the fluctuation of the forward-sampled optical power sequence and the fluctuation of the monitored photocurrent sequence. The pseudo-stability and instability risk value for the current window is calculated based on the coherent return cluster intensity index, detection gain, and apparent stability metric. Based on the rolling historical risk sequence, the pseudo-stability and instability risk value of the current window is adaptively segmented to obtain the adaptive segmentation result; Based on the adaptive segmentation results, the monitoring status of the current window is determined and output, wherein the monitoring status includes normal status, latent return cluster status, and explicit instability status.

9. A light source monitoring system for an optical communication link, applied in the light source monitoring method for an optical communication link according to any one of claims 1-8, characterized in that, The system includes: The residual extraction module is used to collect the emission driving current sequence, monitoring photocurrent sequence, forward sampling optical power sequence and source-side back sensing sequence of the light source, remove the common-mode response, and obtain the emission response residual sequence and the back disturbance residual sequence. The path analysis module is used to construct the path difference spectrum based on the luminescence response residual sequence and the return perturbation residual sequence, extract the local peak set, and calculate the path difference discrete quantity. The coherent evaluation module is used to extract the dominant oscillation frequency from the luminescence response residual sequence and calculate the coherent return cluster intensity index by combining the local peak set and the path difference discrepancy. The enhanced detection module is used to apply a constrained micro-detection current waveform at the current operating point based on the forward sampled optical power sequence, the monitored photocurrent sequence, the coherent back cluster intensity index, and the path difference discreteness, and to calculate the detection gain based on the coherent back cluster intensity index before and after detection. The risk assessment module is used to calculate the apparent stability measure based on the forward sampling optical power sequence and the monitoring photocurrent sequence, combine the coherent return cluster intensity index and the detection gain to calculate the pseudo-stability and instability risk value, and output the monitoring results.

Images