Optical module damage detection method and system based on machine learning

Abstract

The invention discloses an optical module damage detection method and system based on machine learning, and relates to the technical field of optical communication, and the method comprises the steps: synchronously capturing a transmitting and receiving light power dynamic fluctuation track, thermal response gradient data and a power ripple instantaneous frequency spectrum, and generating an original signal flow after the alignment of a timestamp engine; based on the original signal flow, identifying a causal feature group through conditional independence test, and generating a causal feature vector matrix in combination with a semiconductor thermodynamic equation and a photoelectric conversion principle; converting the causal feature vector matrix into a dynamic tense graph structure, calculating an edge weight through a time convolution network, aggregating spatiotemporal features through a graph attention mechanism, and then outputting an impairment mode embedded vector; according to the method, the false correlation between the temperature gradient and the voltage ripple is screened out through the triplex condition independence test, so that the misjudgment rate of feature selection is reduced.

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to a method and system for detecting optical module damage based on machine learning. Background Technology

[0002] Current damage detection technologies for optical communication equipment mainly rely on machine learning algorithms to analyze operating parameters. Existing solutions focus on two types of implementation paths: one is deep learning methods based on time series modeling (such as LSTM, TCN, etc.), which predict damage risk by analyzing single-dimensional parameters such as optical power and temperature; the other is lightweight models driven by transfer learning, which adapt to new deployment environments through pre-trained models.

[0003] However, existing technologies have significant limitations: First, there are defects in the fusion of multi-source heterogeneous data: there is an asynchronous sampling problem between the dynamic fluctuation of optical power, thermal response gradient and power supply ripple spectrum, and the timing error introduced by traditional interpolation alignment methods is high, which leads to distortion of cross-domain feature correlation and an increase in false alarm rate; Second, there is a lack of physical mechanism modeling: existing feature engineering mostly relies on statistical correlation, but the damage of optical modules is essentially governed by semiconductor thermodynamic equations and photoelectric conversion principles, and pure data-driven models cannot distinguish between the causal link of temperature gradient-photocurrent and pseudo-correlation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based optical module damage detection method to solve the problems of feature distortion caused by asynchronous multi-source heterogeneous data and the lack of physical interpretability of damage features.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a machine learning-based optical module damage detection method, which includes simultaneously capturing the dynamic fluctuation trajectory of received and emitted power, thermal response gradient data and instantaneous spectrum of power ripple, and generating the original signal stream after alignment by a timestamp engine; Based on the original signal stream, causal feature groups are identified through conditional independence tests, and a causal feature vector matrix is ​​generated by combining semiconductor thermodynamic equations and photoelectric conversion principles. The causal feature vector matrix is ​​converted into a dynamic temporal graph structure. The edge weights are calculated through a temporal convolutional network. After the spatiotemporal features are aggregated through a graph attention mechanism, the damage pattern embedding vector is output. Based on the damage pattern embedding vector, difficult cases are screened using Monte Carlo perturbation samples, and an optimized diagnostic model is generated by integrating a knowledge distillation mechanism. When the damage probability output by the optimized diagnostic model exceeds the warning threshold, a causal intervention decision is triggered, generating a level-three alarm command and updating the device health record database.

[0007] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the generation of the original signal stream includes the following steps: The system captures the dynamic fluctuation trajectory of optical power, thermal response gradient data, and instantaneous spectrum data of power supply ripple. After the hardware coprocessor superimposes microsecond-level time stamps, it outputs time-stamped data points of dynamic fluctuation trajectory of optical power, thermal response gradient data points, and instantaneous spectrum data points of power supply ripple. The clock domain crossover control logic is used to process time-stamped optical power dynamic fluctuation trajectory data points, thermal response gradient data points, and power supply ripple instantaneous spectrum data points to generate the original signal stream.

[0008] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the generation of the causal feature vector matrix includes the following steps. The PC algorithm is used to perform a triple conditional independence test on the optical power dynamic fluctuation trajectory data points, thermal response gradient data points and power supply ripple instantaneous spectrum data points in the original data stream, and outputs the photocurrent fluctuation-temperature gradient causal feature set and the photocurrent fluctuation-voltage ripple causal feature set. For each thermal response gradient data point, thermal carrier mobility is calculated to generate a temperature-power coupling characteristic value sequence. For each optical power dynamic fluctuation trajectory data point, carrier recombination rate is calculated to generate a carrier distortion index sequence. The causal feature sets of photocurrent fluctuation-temperature gradient and photocurrent fluctuation-voltage ripple from the conditional independence test outputs are combined and temperature-power coupling feature values ​​and carrier distortion index sequences are added to form a causal feature vector matrix.

[0009] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the output damage mode embedding vector includes the following steps. Based on the causal feature vector matrix, a basic time series graph topology is constructed. Within the sliding window, a three-layer causal convolution kernel with increasing dilation factor is used to calculate the spatiotemporal correlation strength between nodes. A cross-time period dependency weight matrix is ​​generated by cross-layer weighted summation. The cross-time dependency weight matrix is ​​input into the graph attention mechanism to calculate the attention coefficients between nodes. The neighbor features are then weighted and aggregated to generate a spatiotemporal fusion vector. This vector is then processed alternately by a gated temporal recurrent unit and a graph convolutional network, and residual connections and pooling dimensionality reduction are performed to output the damage pattern embedding vector.

[0010] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the generation of the optimized diagnostic model includes the following steps. Based on the damage pattern embedding vector and the pre-stored physical verification damage sample set, a deep neural network architecture is initialized and iteratively optimized to convergence through backpropagation using cross-entropy loss, generating a benchmark diagnostic model. In the real-time raw signal stream of the optical module, random offset is superimposed on the optical power trajectory, Gaussian noise is injected into the temperature gradient, and harmonic distortion is added to the ripple spectrum to generate a disturbance sample set. After three-level transformation of synchronous capture, causal feature generation and temporal graph modeling, the disturbance damage mode embedding vector is output. Based on the perturbation loss pattern embedding vector, a difficult example is selected by generative adversarial network, and a pre-stored physical verification damage sample set is merged to form an enhancement set. Feature weighted training and knowledge distillation are used to output an optimized diagnostic model.

[0011] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the generation of the three-level alarm command includes the following steps. The damage probability value is calculated in real time based on the optimized diagnostic model. When the confidence threshold is exceeded, the causal intervention engine is activated to generate the theoretical value range by performing thermodynamic-photoelectric dual constraints and calculate the divergence values ​​of driving current and temperature. Dynamic thresholds are generated based on historical false alarm rates. Physical layer root causes are determined by exceeding JS divergence limits, and three-level alarm commands are generated by matching multi-level alarm rules.

[0012] As a preferred embodiment of the machine learning-based optical module damage detection method of the present invention, the updating of the device health record database refers to writing the damage probability value output by the optimized diagnostic model, the divergence value calculated by the causal intervention engine, and the alarm level generated by the three-level alarm command into the damage event record table of the device health record database.

[0013] Secondly, this invention provides a machine learning-based optical module damage detection system, comprising, The synchronous acquisition module synchronously captures and receives dynamic fluctuation trajectory of power output, thermal response gradient data, and instantaneous spectrum of power ripple, and generates the original signal stream after alignment by the timestamp engine; The causal feature module, based on the original signal stream, identifies causal feature groups through conditional independence tests and generates a causal feature vector matrix by combining semiconductor thermodynamic equations and photoelectric conversion principles. The graph modeling module converts the causal feature vector matrix into a dynamic temporal graph structure, calculates the edge weights through a temporal convolutional network, aggregates the spatiotemporal features through a graph attention mechanism, and outputs the damage pattern embedding vector. The incremental training module uses Monte Carlo perturbation samples to screen difficult examples based on the damage pattern embedding vector and integrates a knowledge distillation mechanism to generate an optimized diagnostic model. The decision center module triggers causal intervention decisions when the damage probability output by the diagnostic model exceeds the warning threshold, generates a three-level alarm command, and updates the device health record database.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based optical module damage detection method described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based optical module damage detection method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: sampling time difference is eliminated by cross-controlling the FPGA atomic clock timestamp engine and clock domain; the triple conditional independence test (PC algorithm) filters out the spurious correlation between temperature gradient and voltage ripple, thus reducing the feature selection misjudgment rate; thermal resistance attenuation characteristics are calculated based on the carrier mobility equation, and SRH composite lifetime quantification material defects are superimposed to generate temperature-power coupling characteristics and carrier distortion index; Monte Carlo perturbation samples are trained adversarially to generate a difficult example set, and the volume is compressed by combining knowledge distillation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a machine learning-based optical module damage detection method.

[0019] Figure 2 A schematic diagram of the generation of the original data stream.

[0020] Figure 3 A schematic diagram for generating the causal feature vector matrix.

[0021] Figure 4 A schematic diagram of damage mode embedding vector generation. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a machine learning-based optical module damage detection method, comprising the following steps: S1. Synchronously capture the dynamic fluctuation trajectory of the received power, thermal response gradient data, and instantaneous spectrum of the power ripple, and generate the original signal stream after alignment by the timestamp engine.

[0026] S1.1. Capture the dynamic fluctuation trajectory of optical power, thermal response gradient data, and instantaneous spectrum data of power supply ripple. After the hardware coprocessor superimposes microsecond-level time stamps, output the time-stamped dynamic fluctuation trajectory data points of optical power, thermal response gradient data points, and instantaneous spectrum data points of power supply ripple.

[0027] Specifically, a sampling rate configuration command is sent to the high-speed ADC chip to set the capture parameters and continuously acquire the dynamic fluctuation trajectory of optical power received by the optical module; a synchronization trigger signal is sent to the temperature sensor to output millisecond-level thermal response gradient data in real time; a start command is sent to the onboard high-speed current probe to synchronously capture the instantaneous spectrum data of power supply ripple; the dynamic fluctuation trajectory data of optical power, thermal response gradient data, and instantaneous spectrum data of power supply ripple are input in parallel to the synchronization buffer of the FPGA timestamp engine through the SPI interface, I2C bus, and LVDS channel, respectively; the hardware coprocessor of the FPGA timestamp engine receives the dynamic fluctuation trajectory data of optical power, thermal response gradient data, and instantaneous spectrum data of power supply ripple in the synchronization buffer, and superimposes absolute time stamps with microsecond-level precision on them, outputting time-stamped dynamic fluctuation trajectory data points of optical power, thermal response gradient data points, and instantaneous spectrum data points of power supply ripple.

[0028] S1.2. Clock domain cross-control logic is used to process time-stamped optical power dynamic fluctuation trajectory data points, thermal response gradient data points, and power supply ripple instantaneous spectrum data points to generate the original signal stream.

[0029] Specifically, the clock domain cross controller of the FPGA timestamp engine is activated to detect the phase offset between the timestamps of the optical power dynamic fluctuation trajectory data, the thermal response gradient data, and the power supply ripple instantaneous spectrum data; the three sets of data are dynamically compensated for by the phase-locked loop to compensate for the clock domain deviation of the timestamps of the optical power dynamic fluctuation trajectory data, the thermal response gradient data, and the power supply ripple instantaneous spectrum data; and the atomic clock reference source built into the FPGA timestamp engine is connected to generate a reference clock signal.

[0030] Using a reference clock signal, the three sets of data timestamps are remapped into a timestamp sequence of dynamic fluctuation trajectory of optical power with nanosecond precision, a timestamp sequence of thermal response gradient, and a timestamp sequence of instantaneous spectrum of power supply ripple. The data alignment unit is activated to create a unified time axis register starting at time zero. The timestamp sequences of the three sets of data are loaded into the unified time axis register simultaneously. The triple data stream interpolation operator is started to obtain the target time node of the unified time axis. Based on the time interval of the timestamp sequences of the three sets of data, the linear interpolation weight of the corresponding data point at the target time node is calculated. The linear interpolation reconstruction of the optical power dynamic fluctuation trajectory data, thermal response gradient data, and power supply ripple instantaneous spectrum data is performed synchronously. The time synchronization data block containing the aligned three sets of data points is output. Extract the optical power dynamic fluctuation trajectory data points, thermal response gradient data points, and power supply ripple instantaneous spectrum data points at the same time from the time synchronization data block, and package them into the original signal stream of the fluctuation trajectory-temperature gradient-ripple spectrum triplet sequence with an additional absolute timestamp of the unified time axis.

[0031] S2. Based on the original signal stream, identify causal feature groups through conditional independence test, and generate causal feature vector matrix by combining semiconductor thermodynamic equation and photoelectric conversion principle.

[0032] S2.1. The PC algorithm is used to perform a triple conditional independence test on the optical power dynamic fluctuation trajectory data points, thermal response gradient data points and power supply ripple instantaneous spectrum data points in the original signal stream, and outputs the photocurrent fluctuation-temperature gradient causal feature set and the photocurrent fluctuation-voltage ripple causal feature set.

[0033] Specifically, from each triplet, dynamic fluctuation trajectory data points of optical power (as photocurrent fluctuation data points), temperature gradient data points, and voltage ripple data points are extracted; these data points are then input into the PC algorithm to perform a triple conditional independence test. First, the conditional mutual information values ​​of photocurrent fluctuation data points and temperature gradient data points under the condition of fixed voltage ripple data points are calculated, and the conditional independence probability p1 is obtained through chi-square distribution transformation. Next, the conditional mutual information values ​​of the photocurrent fluctuation data points and voltage ripple data points under the condition of fixed temperature gradient data points are calculated, and the conditional independence probability p2 is obtained by conversion. Finally, the conditional mutual information values ​​of the temperature gradient data points and voltage ripple data points under the condition of fixed photocurrent fluctuation data points are calculated, and the conditional independence probability p3 is obtained by conversion. Compare p1, p2, and p3 with the causal significance threshold (determined by balancing false alarm rate and false negative rate based on the reliability standards and statistical confidence requirements of optical communication equipment through engineering verification data). When p1 < causal significance threshold, output the photocurrent fluctuation-temperature gradient causal characteristic group; when p2 < causal significance threshold, output the photocurrent fluctuation-voltage ripple causal characteristic group.

[0034] When p3 < causal significance threshold, i.e. temperature gradient and voltage ripple show statistical significance correlation, it belongs to sensor crosstalk (requires hardware maintenance) or power supply heat dissipation coupling (belongs to setting problem), and no causal feature group will be output.

[0035] To further explain, taking the photocurrent fluctuation data points and temperature gradient data points at a fixed voltage ripple data point as an example, the expression for calculating the conditional mutual information value of the photocurrent fluctuation data points and voltage ripple data points under the condition of a fixed temperature gradient data point is as follows: ; in, This represents a sequence of photocurrent fluctuation data. Represents a temperature gradient data sequence. This represents a voltage ripple data sequence. express and exist Conditional mutual information value under the given conditions The probability of voltage ripple state is used as a weighting factor to adjust the mutual information contribution under different voltage states. This indicates the specific values ​​of the voltage ripple data. This indicates the specific values ​​of the photocurrent fluctuation data. This indicates the specific values ​​of the temperature gradient data. Indicates in fixed conditions, and The joint probability, and They represent hour, and The conditional probability; The steps to convert conditional mutual information values ​​into conditional independence probabilities are as follows: First, calculate the chi-square statistic, expressed as: ; in, This represents the chi-square statistic. Indicates the number of data points in the sampling window (e.g.) ); Then, determine the degrees of freedom, expressed as: ; in, Indicates degrees of freedom. , and These represent the number of discretization intervals for the photocurrent fluctuation data sequence, temperature gradient data sequence, and voltage ripple data sequence, respectively (e.g., 10 intervals for each). Finally, the conditional independence probability is calculated using the following expression: ; in, Represents the probability of conditional independence. This represents the chi-square cumulative distribution function, which is implemented by looking up a table or numerical integration and is used to calculate the theoretical cumulative probability.

[0036] S2.2. Perform thermal carrier mobility calculation for each thermal response gradient data point to generate a temperature-power coupling characteristic value sequence; calculate carrier recombination rate for each optical power dynamic fluctuation trajectory data point to generate a carrier distortion index sequence. The photocurrent-temperature gradient feature set and the photocurrent-voltage ripple causal feature set are merged, and the temperature-power coupling feature value sequence and the carrier distortion index sequence are added to form a causal feature vector matrix.

[0037] Specifically, a reference mobility constant is read from a semiconductor material library, and temperature characteristic parameters are obtained. For each thermal response gradient data point in the photocurrent fluctuation-temperature gradient causal characteristic group, the semiconductor thermodynamic equation is used to calculate the hot carrier mobility, generating a temperature-power coupling characteristic value sequence, expressed as:

[0039] in, Indicates temperature The hot carrier mobility at that point Indicates the baseline mobility. This represents the i-th temperature gradient data point. Represents the temperature decay constant; It should be noted that calculating the thermal carrier mobility and generating the temperature-power coupling eigenvalue sequence refers to sequentially associating each mobility value with the power condition (such as the square of the electric field strength) of the corresponding thermal response gradient data point, and then outputting the results in series to finally generate the temperature-power coupling eigenvalue sequence.

[0040] Based on the dynamic fluctuation trajectory data points of optical power in the photocurrent fluctuation-voltage ripple causal characteristic set, the data is read from the photoelectric conversion parameter library. The recombination lifetime is calculated, and based on the photoelectric conversion principle, the carrier recombination rate is calculated, and a carrier distortion index sequence is constructed, expressed as: ; in, Indicates the first The carrier recombination rate at a given time characterizes the rate at which electron-hole pairs disappear per unit time and unit volume, reflecting the material's defect density. Indicates the first Photocurrent fluctuation data points at any given time. express Composite lifetime is a characteristic time constant determined by the defect states of the material; the smaller the value, the more composite centers there are. It should be noted that calculating the carrier recombination rate and constructing the carrier distortion index sequence refers to directly constructing the carrier distortion index sequence by sorting the carrier recombination rates calculated at each moment according to the time series. The numerical fluctuation of the carrier distortion index sequence directly reflects the real-time distortion effect of material defect density on the photoelectric conversion process.

[0041] The causal feature sets of photocurrent fluctuation-temperature gradient and photocurrent fluctuation-voltage ripple from the conditional independence test outputs are combined and temperature-power coupling feature values ​​and carrier distortion index sequences are added to form a causal feature vector matrix.

[0042] It should be noted that the semiconductor material library was obtained by measuring the carrier mobility curves at different temperatures using a Hall effect testing system; the temperature characteristic parameters were obtained by testing the laser output power as a function of temperature under constant pressure and fitting the attenuation curve using the least squares method; and the photoelectric conversion parameter library was obtained by constructing a set of calibration parameters for multi-source verification based on photoelectric device testing standards and Hall effect measured data. Recombination lifetime is determined by exciting carriers with a wavelength-tunable laser, measuring carrier lifetime curves using time-resolved photoluminescence decay, and extracting the characteristic time constant of the defect-dominated recombination process as... Composite lifespan.

[0043] S3. The causal feature vector matrix is ​​converted into a dynamic temporal graph structure. The edge weights are calculated through a temporal convolutional network. After the spatiotemporal features are aggregated through a graph attention mechanism, the damage pattern embedding vector is output.

[0044] S3.1. Based on the causal feature vector matrix, a basic time series graph topology is constructed. Within the sliding window, a three-layer causal convolution kernel with increasing dilation factor is used to calculate the spatiotemporal correlation strength between nodes. A cross-time period dependency weight matrix is ​​generated by cross-layer weighted summation.

[0045] Specifically, the eigenvectors in the causal eigenvector matrix at each time point are arranged in ascending order of timestamp; the eigenvectors corresponding to each timestamp are converted into graph nodes, and the graph node attributes are assigned the eigenvalues ​​of the corresponding time points in the causal eigenvector matrix; the initial graph structure is constructed using the timestamp values ​​as unique identifiers for the graph nodes; the graph nodes are sorted in ascending order of timestamp values ​​to form the basic time-series graph topology. On the time axis of the basic time series graph topology, continuous time intervals are defined, and a sliding window of fixed duration is set, covering multiple consecutive time nodes. A temporal convolutional network is then applied to these multiple consecutive time nodes within the sliding window. The first convolutional layer uses a causal convolutional kernel with an inflation factor of 1 (kernel size 3) to calculate the local spatiotemporal correlation strength between each node and its preceding neighbor, itself, and its following neighbor. The second convolutional layer uses a causal convolutional kernel with an inflation factor of 2 (kernel size 3) to calculate the mid-range spatiotemporal correlation strength between each node and its two preceding nodes, itself, and its two following nodes. The third convolutional layer uses a causal convolutional kernel with an inflation factor of 4 (kernel size 3) to calculate the long-range spatiotemporal correlation strength between each node and its four preceding nodes, itself, and its four following nodes. By integrating the local spatiotemporal correlation strength of the first convolutional layer, the mid-range spatiotemporal correlation strength of the second convolutional layer, and the long-range spatiotemporal correlation strength of the third convolutional layer, the cross-time-period dependency weights are calculated for each node pair, constructing a cross-time-period dependency weight matrix, expressed as: ; in, Indicates the current node with neighboring nodes Edge weights between them Indicates the convolutional layer index. This represents the activation function. Indicates the first The weight matrix of a convolutional layer. Represents a node eigenvectors, Represents a node eigenvectors, express and Feature splicing operation.

[0046] S3.2. Input the cross-time dependency weight matrix into the graph attention mechanism, calculate the attention coefficients between nodes, and aggregate the neighbor features in a weighted manner to generate a spatiotemporal fusion vector. The vector is then processed alternately by a gated temporal recurrent unit and a graph convolutional network, and residual connections and pooling dimensionality reduction are performed to output the damage pattern embedding vector.

[0047] Specifically, the cross-time-period dependency weight matrix is ​​input into the graph attention mechanism processing unit; an independent attention calculation unit is assigned to the graph node corresponding to each timestamp; the attention coefficient between the current node and its neighboring nodes is calculated, with the expression as follows: ; in, Indicates the current node with neighboring nodes Attention coefficient This represents a learnable attention vector used to map high-dimensional features to a scalar score space. Indicates the current node The set of neighbors defines the receptive field of graph attention. This represents an improved activation function that addresses the vanishing gradient problem with negative inputs. Represents the neighbor node index variable. Representing neighboring nodes eigenvectors; Based on the attention coefficient weighted aggregation of neighbor node features, an initial spatiotemporal fusion feature vector is generated. The spatiotemporal fusion feature vector is then input into the gated temporal recurrent unit processing layer: the hidden state is initialized as a zero vector, and neighbor nodes are processed one by one in chronological order. For each neighbor node, the update gate, reset gate, and candidate hidden state are calculated, and the new hidden state is output. After all neighbor nodes are processed, the hidden state matrix is ​​output. The hidden state matrix is ​​input into the graph convolutional network layer: an adjacency matrix with self-connection is constructed and the degree matrix is ​​calculated. Graph convolution operation is performed to generate the first graph convolution output. The first graph convolution output is input into the gated temporal recurrent unit again, and the second hidden state matrix is ​​output according to the same process. Finally, the initial spatiotemporal fusion feature vector is added element-wise to the second graph convolution output to generate the node feature vector after residual connection. Max pooling is performed on the node feature vectors after residual concatenation along the time axis to extract the maximum value of each feature dimension and form a pooling vector. The pooling vector is then input into a fully connected layer for linear transformation and dimensionality reduction, and the damage mode embedding vector is output.

[0048] S4. Based on the damage pattern embedding vector, use Monte Carlo perturbation samples to screen for difficult cases, and integrate the knowledge distillation mechanism to generate an optimized diagnostic model.

[0049] S4.1. Based on the damage pattern embedding vector and the pre-stored physical verification damage sample set, initialize the deep neural network architecture and iteratively optimize it to convergence through backpropagation of cross-entropy loss to generate a benchmark diagnostic model.

[0050] Specifically, after receiving the damage pattern embedding vector, a pre-stored physical verification damage sample set is loaded; the pre-stored physical verification damage sample set contains laboratory-calibrated damage type labels and damage severity quantification values; the weight parameters and network layer connections of the deep neural network architecture are initialized; the damage pattern embedding vector is input into the input layer of the deep neural network architecture; the deep neural network architecture parses the damage pattern embedding vector layer by layer and outputs the damage probability distribution vector; the difference between the predicted probability distribution of the damage probability distribution vector and the real damage label is calculated based on the cross-entropy loss function; the weight parameters of the deep neural network architecture are updated through the backpropagation algorithm; and a benchmark diagnostic model is generated after iterative optimization until the loss function converges.

[0051] S4.2. In the real-time raw signal stream of the optical module, random offset is superimposed on the optical power trajectory, Gaussian noise is injected into the temperature gradient, and harmonic distortion is added to the ripple spectrum to generate a disturbance sample set. After three-level transformation of synchronous capture, causal feature generation and temporal graph modeling, the disturbance damage mode embedding vector is output.

[0052] Specifically, the process involves: capturing the original signal stream of the optical module in the real-time data acquisition channel; performing a perturbation operation on the dynamic fluctuation trajectory of optical power in the original signal stream: superimposing a random offset on the original value of the dynamic fluctuation trajectory of optical power to generate a perturbed optical power trajectory; performing a perturbation operation on the thermal response gradient data: superimposing Gaussian white noise including the mean and standard deviation on the original temperature gradient value to generate a perturbed temperature gradient; performing a perturbation operation on the instantaneous spectrum of the power supply ripple: injecting random harmonic distortion of the fundamental component into the original spectrum to generate a perturbed ripple spectrum; combining the perturbed optical power trajectory, temperature gradient, and ripple spectrum into a perturbed sample set; converting the perturbed sample set into a time-stamped triplet sequence in the synchronous acquisition process of S1; converting the time-stamped triplet sequence into a perturbed causal feature vector matrix in the causal feature generation process of S2; and inputting the perturbed causal feature vector matrix into the temporal graph network modeling process in S3 to generate a perturbed damage mode embedding vector.

[0053] S4.3. Based on the perturbation damage pattern embedding vector, difficult examples are screened by generative adversarial network, and a pre-stored physical verification damage sample set is merged to form an enhancement set. Feature weighted training and knowledge distillation are used to output an optimized diagnostic model.

[0054] Specifically, the baseline diagnostic model is fixed as the discriminator component of a generative adversarial network; a screening operation is performed on the input perturbation damage pattern embedding vector: first, the discriminator calculates the prediction confidence of each sample, and marks samples with prediction confidence below the confidence threshold as potential hard examples; the potential hard examples are input into the generator to reconstruct the feature vector and then input into the discriminator again to calculate the mean absolute error between the reconstructed feature vector and the damage pattern embedding vector; samples with mean absolute error greater than the error threshold are retained to form the final hard example sample set, and type labels are added to the final hard example sample set, outputting the labeled final hard example sample set; To further clarify, the confidence threshold is set based on the lower confidence limit requirement for optical module damage detection; the error threshold is set based on the upper limit of the normalized tolerance of the feature vector dimension.

[0055] The pre-stored physical verification damage sample set and the labeled final hard example sample set are merged to form an enhanced training set: the information gain method is used to evaluate the discriminative contribution of each feature dimension in the enhanced training set to damage classification, the feature weight values ​​are calculated and normalized to the interval of 0 to 1; the cross-entropy loss function is adjusted based on the feature weight values ​​to increase the weight ratio of high-contribution features in the loss calculation. Filter low-confidence samples in the enhanced training set; freeze the parameters of the baseline diagnostic model as the teacher model; use the teacher model with half the number of layers as the student model; combine the true label loss with the output distribution loss of the teacher model; set a temperature coefficient and control the loss weight ratio to output the trained student model; It should be noted that freezing the baseline diagnostic model parameters as the teacher model means that the weights of the frozen baseline diagnostic model no longer participate in gradient updates, which is a standard operation of knowledge distillation.

[0056] It should be noted that the loss is combined with the loss of the real label and the output distribution of the teacher model; a temperature coefficient is set and the weight ratio of the loss is controlled; the output of the trained student model refers to the difference between the predicted probability distribution of the student model output and the real damage label, and the classification error is quantified by the cross-entropy loss function; the softened probability distribution of the teacher model output (the probability vector after being scaled by the temperature coefficient) is extracted simultaneously, and the KL divergence loss between the student model output and the teacher softened distribution is calculated. The KL divergence loss carries the inter-class association information learned by the teacher model; the temperature coefficient is introduced to adjust the smoothness of the softened distribution (the larger the value, the smoother the distribution), which enhances the learning signal of hard samples; the weight ratio of the real label loss and the teacher distribution loss is set, and the final distillation loss function is generated by weighted summation; backpropagation is performed based on the final loss, and knowledge transfer is achieved by updating only the student model parameters.

[0057] Based on the trained learning model, the historical gradient data during the training process is backtracked; the squared mean of the historical gradient data is calculated for each student model parameter to generate a parameter importance score; the parameter importance score is input into the weight update algorithm; a regularization constraint term is added during gradient descent to limit the adjustment range of the parameter importance score; after the regularization constraint training is completed, the optimized weight matrix and bias vector are output, and the importance matrix is ​​generated simultaneously, which are combined into the optimized diagnostic model.

[0058] It should be noted that inputting parameter importance scores into the weight update algorithm; adding regularization constraints during gradient descent to limit the adjustment range of parameter importance scores refers to initializing the regularization constraint controller after receiving the parameter importance scores; dividing the parameter groups into high-importance and low-importance groups based on their importance scores; configuring strong regularization coefficients for the high-importance parameter groups to limit the weight adjustment range; configuring weak regularization coefficients for the low-importance parameter groups to increase the weight adjustment range; calculating the original gradient of the loss function for the current batch during gradient descent iterations; applying regularization constraints to the original gradients: the gradients of high-importance parameters are decayed proportionally (e.g., gradient value × 0.05), while the gradients of low-importance parameters remain unchanged; inputting the constrained gradients into the optimizer to perform weight updates; and outputting the updated student model weight matrix and bias vector.

[0059] S5. When the damage probability output by the diagnostic model exceeds the warning threshold, a causal intervention decision is triggered, a level-three alarm instruction is generated, and the device health record database is updated.

[0060] S5.1. Based on the optimized diagnostic model, the damage probability value is calculated in real time. When the warning threshold is exceeded, the causal intervention engine is activated to generate the theoretical value range by performing thermodynamic-photoelectric dual constraints, and to calculate the divergence values ​​of driving current and temperature.

[0061] Specifically, based on the weight matrix, bias vector, and importance matrix in the optimized diagnostic model, the weight matrix is ​​loaded into the convolution kernel parameter register of the deep neural network architecture; the bias vector is written into the bias register of the fully connected layer; the importance matrix is ​​configured to the regularization constraint controller; the perturbation damage pattern is embedded into the vector input of the optimized diagnostic model; local damage features are extracted through the convolutional layer; the features are fused through the fully connected layer and the damage probability distribution vector is output, and finally the damage probability value (range 0~1) at the current time is calculated. Initialize the warning status register to "normal"; continuously read the damage probability value output by the optimized diagnostic model; when the damage probability value exceeds the warning threshold, update the warning status register to "pending activation"; generate a causal intervention engine trigger pulse signal; It should be noted that the warning threshold is set based on the confidence level requirement for optical module damage detection.

[0062] Upon receiving the trigger pulse signal, the causal intervention engine is activated, accesses the optical module specification parameter library, extracts the allowable range of laser drive current (the specific value is defined by the optical module model document), extracts the ambient temperature operating range (the specific value is defined by the optical module thermal specification), and loads the hot carrier mobility (obtained through Hall effect testing) from the semiconductor material library and the quantum efficiency (obtained through photoelectric response testing) from the photoelectric conversion parameter library. It should be noted that the optical module specification parameter library is a standardized database built by integrating equipment manufacturers' documents (including model specifications and thermal specifications), semiconductor material test reports, and photoelectric conversion calibration data. It is initialized before the equipment leaves the factory and continuously calibrated and updated based on the measured data after deployment.

[0063] Based on the measured values ​​of input optical power and bias voltage collected at the current moment, the nonlinear decay relationship of current with temperature is derived using the semiconductor thermodynamic equation and real-time temperature gradient data points. The current fluctuation range is limited according to quantum efficiency. The theoretical range of driving current is generated by integrating thermodynamic and photoelectric constraint boundaries. Temperature distribution modeling is performed simultaneously, and a steady-state temperature gradient distribution is generated by using Fourier's law in the thermal specification. The theoretical distribution range is adjusted by combining the thermal resistance coefficient of the heat sink (obtained from the mechanical structure document) to generate a theoretical normal temperature distribution. It should be noted that limiting the current fluctuation range restricts the driving current to the linear operating region of photoelectric conversion through quantum efficiency constraints, avoiding laser output power saturation or carrier overflow due to current exceeding the limit, thereby ensuring that the theoretical value range can accurately reflect the physical damage boundary of semiconductor devices.

[0064] To further explain, the thermal conductivity parameters of the optical module material are obtained from the thermal specification document; based on the material thermal conductivity parameters and combined with the real-time acquired chip heat source power data, a three-dimensional heat conduction differential equation is established; the initial temperature gradient distribution field is obtained by solving the three-dimensional heat conduction differential equation; the thermal resistance coefficient of the heat sink in the mechanical structure document is read, and boundary condition correction is performed on the initial temperature gradient distribution; the gradient change rate of the overheated region in the initial temperature gradient distribution field is adjusted through iterative optimization, and finally the steady-state temperature gradient distribution that satisfies the thermal equilibrium condition is output.

[0065] It should be noted that establishing a three-dimensional thermal conduction differential equation based on material thermal conductivity parameters and real-time chip heat source power data involves reading the thermal conductivity parameters of each layer of the optical module material (such as the thermal conductivity of aluminum nitride in the chip layer and FR4 in the substrate layer); calculating the anisotropic thermal conduction tensor matrix of the material based on the thermal conductivity parameters; synchronously receiving real-time chip heat source power data (calculating instantaneous power consumption through onboard current sensors and voltage sampling circuits); mapping the heat source power data to the heat source location coordinates in a three-dimensional spatial mesh; establishing a basic heat flow balance equation based on the Fourier law of thermal conduction: the heat capacity per unit volume multiplied by the rate of temperature change equals the divergence of the dot product of the thermal conductivity tensor and the temperature gradient plus the heat source power density; refining the heat capacity term calculation by combining material specific heat capacity parameters and density parameters; performing finite element mesh discretization on the three-dimensional space (the mesh size is adaptively adjusted according to the rate of change of the thermal gradient); establishing a discretized set of thermal conduction differential equations at the mesh nodes; applying heat sink boundary conditions (calculating the boundary heat flux density based on the thermal resistance coefficient); and finally generating a three-dimensional thermal conduction partial differential equation containing spatial coordinates and time variables.

[0066] Extract the drive current measurement value corresponding to the current timestamp and the ambient temperature measurement value at the same timestamp from the raw data stream, and verify that the timestamp is aligned with the theoretical simulation reference time; discretize the theoretical current value range into a probability distribution, and construct a Gaussian distribution centered on the drive current measurement value, and calculate the divergence value of the drive current, expressed as: ; in, Indicates drive current The JS divergence value is used to quantify the degree of difference between the theoretical distribution and the actual distribution (the larger the value, the more significant the anomaly). KL divergence is used to measure the information loss between two probability distributions. Indicates drive current Theoretical distribution, This represents the arithmetic mean of the theoretical and measured distributions of the driving current. Indicates drive current The measured distribution is a Gaussian distribution centered on the measured value of the driving current. The JS divergence value of temperature is calculated synchronously. The formula for the JS divergence value of temperature is the same as that for the JS divergence value of the driving current.

[0067] S5.2. Generate dynamic thresholds based on historical false alarm rates, determine physical layer root causes by exceeding JS divergence limits, generate three-level alarm commands by matching multi-level alarm rules, and update health record records synchronously.

[0068] Specifically, the system queries the historical operation database of the equipment to obtain the false alarm rate statistical curve; it uses the least squares method to fit the false alarm rate time series data to generate a dynamic threshold; when the Jensen-Shannon divergence value of the drive current or temperature exceeds the dynamic threshold, it is determined to be a physical layer abnormal root cause, and the root cause location identifier (such as "drive current deviation" or "temperature exceeds the limit") is output. Perform rule matching based on the input root cause identifier: Level 1 alarm: Single flag activation (e.g., only "drive current offset") → trigger code ALARM_1; Level 2 alarm: Two flags are activated simultaneously (e.g., "drive current offset" + "temperature exceeds limit") → trigger code ALARM_2; Level 3 alarm: Dual flag activation and meeting the characteristic mutation conditions (such as containing multiple anomaly types and the divergence value increasing rapidly) → trigger code ALARM_3; The corresponding maintenance work order is generated based on the alarm level: Level 1 alarm work order suggests parameter calibration, Level 2 alarm work order suggests component inspection, and Level 3 alarm work order suggests immediate shutdown for maintenance. Extract the characteristic parameters of the current damage event: including the damage probability value output by the diagnostic model, the divergence value calculated by the causal intervention engine, and the alarm level generated by the alarm command; write the damage event record table (including timestamp, damage type, and severity fields) into the device health record database; update the cumulative damage index based on the damage probability value (e.g., new index = original index + damage probability × time weight), and output the updated health record record.

[0069] It should be noted that the health record database is a pre-installed equipment lifecycle management database on the optical communication equipment operation and maintenance platform. Basic records are created when the equipment is put into production and are continuously updated.

[0070] This embodiment also provides a machine learning-based optical module damage detection system, including: The synchronous acquisition module synchronously captures and receives dynamic fluctuation trajectory of power output, thermal response gradient data, and instantaneous spectrum of power ripple, and generates the original signal stream after alignment by the timestamp engine; The causal feature module, based on the original signal stream, identifies causal feature groups through conditional independence tests and generates a causal feature vector matrix by combining semiconductor thermodynamic equations and photoelectric conversion principles. The graph modeling module converts the causal feature vector matrix into a dynamic temporal graph structure, calculates the edge weights through a temporal convolutional network, aggregates the spatiotemporal features through a graph attention mechanism, and outputs the damage pattern embedding vector. The incremental training module uses Monte Carlo perturbation samples to screen difficult examples based on the damage pattern embedding vector and integrates a knowledge distillation mechanism to generate an optimized diagnostic model. The decision center module triggers causal intervention decisions when the damage probability output by the diagnostic model exceeds the warning threshold, generates a three-level alarm command, and updates the device health record database.

[0071] This embodiment also provides a computer device applicable to the machine learning-based optical module damage detection method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based optical module damage detection method proposed in the above embodiment.

[0072] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0073] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the machine learning-based optical module damage detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0074] In summary, this invention eliminates sampling time difference by using an FPGA atomic clock timestamp engine and clock domain cross-control; triple conditional independence test (PC algorithm) filters out spurious correlations between temperature gradient and voltage ripple, reducing the feature selection misjudgment rate; thermal resistance attenuation characteristics are calculated based on the carrier mobility equation, and SRH composite lifetime quantification material defects are superimposed to generate temperature-power coupling characteristics and carrier distortion index; Monte Carlo perturbation samples are used for adversarial training to generate a difficult example set, and knowledge distillation is combined to compress the volume.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based optical module damage detection method, characterized in that: include, The system simultaneously captures the dynamic fluctuation trajectory of light power, thermal response gradient data, and instantaneous spectrum of power ripple, and generates the original signal stream after alignment by the timestamp engine. Based on the original signal stream, causal feature groups are identified through conditional independence tests, and a causal feature vector matrix is ​​generated by combining semiconductor thermodynamic equations and photoelectric conversion principles. The causal feature vector matrix is ​​converted into a dynamic temporal graph structure. The edge weights are calculated through a temporal convolutional network. After the spatiotemporal features are aggregated through a graph attention mechanism, the damage pattern embedding vector is output. Based on the damage pattern embedding vector, difficult cases are screened using Monte Carlo perturbation samples, and an optimized diagnostic model is generated by integrating a knowledge distillation mechanism. When the damage probability output by the optimized diagnostic model exceeds the warning threshold, a causal intervention decision is triggered, generating a level-three alarm command and updating the device health record database.

2. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The generation of the original signal stream includes the following steps: The system captures the dynamic fluctuation trajectory of optical power, thermal response gradient data, and instantaneous spectrum data of power supply ripple. After the hardware coprocessor superimposes microsecond-level time stamps, it outputs time-stamped data points of dynamic fluctuation trajectory of optical power, thermal response gradient data points, and instantaneous spectrum data points of power supply ripple. The clock domain crossover control logic is used to process time-stamped optical power dynamic fluctuation trajectory data points, thermal response gradient data points, and power supply ripple instantaneous spectrum data points to generate the original signal stream.

3. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The generation of the causal feature vector matrix includes the following steps. The PC algorithm is used to perform a triple conditional independence test on the optical power dynamic fluctuation trajectory data points, thermal response gradient data points and power supply ripple instantaneous spectrum data points in the original signal stream, and outputs the photocurrent fluctuation-temperature gradient causal feature set and the photocurrent fluctuation-voltage ripple causal feature set. For each thermal response gradient data point, thermal carrier mobility is calculated to generate a temperature-power coupling characteristic value sequence. For each optical power dynamic fluctuation trajectory data point, carrier recombination rate is calculated to generate a carrier distortion index sequence. The causal feature sets of photocurrent fluctuation-temperature gradient and photocurrent fluctuation-voltage ripple from the conditional independence test outputs are combined and temperature-power coupling feature values ​​and carrier distortion index sequences are added to form a causal feature vector matrix.

4. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The output damage pattern embedding vector includes the following steps. Based on the causal feature vector matrix, a basic time series graph topology is constructed. Within the sliding window, a three-layer causal convolution kernel with increasing dilation factor is used to calculate the spatiotemporal correlation strength between nodes. A cross-time period dependency weight matrix is ​​generated by cross-layer weighted summation. The cross-time dependency weight matrix is ​​input into the graph attention mechanism to calculate the attention coefficients between nodes. The neighbor features are then weighted and aggregated to generate a spatiotemporal fusion vector. This vector is then processed alternately by a gated temporal recurrent unit and a graph convolutional network, and residual connections and pooling dimensionality reduction are performed to output the damage pattern embedding vector.

5. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The process of generating the optimized diagnostic model includes the following steps. Based on the damage pattern embedding vector and the pre-stored physical verification damage sample set, a deep neural network architecture is initialized and iteratively optimized to convergence through backpropagation using cross-entropy loss, generating a benchmark diagnostic model. In the real-time raw signal stream of the optical module, random offset is superimposed on the optical power trajectory, Gaussian noise is injected into the temperature gradient, and harmonic distortion is added to the ripple spectrum to generate a disturbance sample set. After three-level transformation of synchronous capture, causal feature generation and temporal graph modeling, the disturbance damage mode embedding vector is output. Based on the perturbation loss pattern embedding vector, a difficult example is selected by generative adversarial network, and a pre-stored physical verification damage sample set is merged to form an enhancement set. Feature weighted training and knowledge distillation are used to output an optimized diagnostic model.

6. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The generation of the Level 3 alarm command includes the following steps. The damage probability value is calculated in real time based on the optimized diagnostic model. When the damage probability value is exceeded, the causal intervention engine is activated to generate the theoretical value range by performing thermodynamic-photoelectric dual constraints and calculate the divergence value of driving current and temperature. Dynamic thresholds are generated based on historical false alarm rates. Physical layer root causes are determined by exceeding JS divergence limits, and three-level alarm commands are generated by matching multi-level alarm rules.

7. The machine learning-based optical module damage detection method as described in claim 1, characterized in that: The update of the device health record database refers to writing the damage probability value output by the optimized diagnostic model, the divergence value calculated by the causal intervention engine, and the alarm level generated by the three-level alarm command into the damage event record table of the device health record database.

8. A machine learning-based optical module damage detection system, based on the machine learning-based optical module damage detection method according to any one of claims 1 to 7, characterized in that: include, The synchronous acquisition module synchronously captures and receives dynamic fluctuation trajectory of power output, thermal response gradient data, and instantaneous spectrum of power ripple, and generates the original signal stream after alignment by the timestamp engine; The causal feature module, based on the original signal stream, identifies causal feature groups through conditional independence tests and generates a causal feature vector matrix by combining semiconductor thermodynamic equations and photoelectric conversion principles. The graph modeling module converts the causal feature vector matrix into a dynamic temporal graph structure, calculates the edge weights through a temporal convolutional network, aggregates the spatiotemporal features through a graph attention mechanism, and outputs the damage pattern embedding vector. The incremental training module uses Monte Carlo perturbation samples to screen difficult examples based on the damage pattern embedding vector and integrates a knowledge distillation mechanism to generate an optimized diagnostic model. The decision center module triggers causal intervention decisions when the damage probability output by the diagnostic model exceeds the warning threshold, generates a three-level alarm command, and updates the device health record database.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based optical module damage detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based optical module damage detection method according to any one of claims 1 to 7.

Images