An online dynamic optical module performance degradation adaptive compensation method

Abstract

This invention relates to the field of optical module communication technology and discloses an online dynamic adaptive compensation method for optical module performance degradation. It solves the technical problems of insufficient trend prediction capability, lack of adaptive dynamic adjustment capability, data dispersion, and lack of global coordination in existing optical module performance degradation monitoring and compensation methods. This invention acquires and preprocesses multi-source heterogeneous parameters, extracts core degradation features using a self-supervised model, and then uses a degradation trend prediction network to obtain core performance prediction values. This effectively captures the mutual influence between multiple processes and the evolution law of parameters, significantly improving the accuracy of performance degradation trend prediction. Furthermore, this invention automatically calculates the optimal compensation factor vector through a compensation decision network and uses proportional-integral-differential calculations to precisely adjust the production equipment. It can dynamically adjust the compensation amount according to the actual degradation state, and collects the actual values ​​and improvement rates after compensation to update each network, forming a closed-loop adaptive optimization.

Description

Technical Field

[0001] This invention relates to the field of optical module communication technology, and in particular to an online dynamic adaptive compensation method for optical module performance degradation. Background Technology

[0002] Currently, performance degradation monitoring and compensation during the optical module manufacturing process mainly rely on the following three methods:

[0003] I. Manual Experience Analysis Method. This method relies on senior engineers regularly reviewing test logs and historical data from each workstation, using their personal experience to determine if there are any abnormalities in the optical module's performance, and manually adjusting process parameters. This method is highly subjective, and the analysis results are greatly affected by the engineer's experience level. Furthermore, it is extremely inefficient when dealing with massive amounts of data from multiple workstations and multiple dimensions, making it difficult to detect hidden or gradual degradation trends.

[0004] Second, traditional statistical process control sets single-point thresholds at each workstation. When a parameter (such as optical power or extinction ratio) exceeds the preset upper or lower limit, an alarm is triggered. It can only capture instantaneous over-limit events and cannot analyze the slow drift pattern of parameters over time (such as gradual decay) or identify the complex degradation patterns caused by the mutual coupling between multiple workstations.

[0005] Third, independent workstation quality control involves setting up independent quality inspection points at each workstation, storing and analyzing data separately, and lacking data correlation and information exchange between workstations. When an abnormality at one workstation is transmitted to subsequent workstations through the process chain, it is impossible to trace the source of the degradation and the path of its spread.

[0006] In summary, the existing technology has the following technical problems:

[0007] 1. It is difficult to simultaneously analyze the changing patterns of equipment parameters over time and the mutual influence between different processes, making it impossible to accurately predict the trend of performance degradation. Often, problems can only be discovered after performance has significantly declined.

[0008] 2. It can only intervene after the parameters exceed the limits. The adjustment decision relies on human experience and cannot automatically calculate the optimal process correction amount based on the current operating status. It is also difficult to continuously optimize the correction strategy when the adjustment effect is not ideal.

[0009] 3. Data from each production stage is stored and managed independently, lacking a unified framework for data fusion and feature extraction. This makes it impossible to comprehensively depict the overall state of the production process, resulting in adjustment decisions being based only on local information and making it difficult to achieve coordinated optimization at the production line level. Summary of the Invention

[0010] The purpose of this invention is to solve the technical problems of insufficient trend prediction capability, lack of adaptive dynamic adjustment capability, data dispersion, and lack of global coordination in existing optical module performance degradation monitoring and compensation, and to provide an online dynamic adaptive compensation method for optical module performance degradation.

[0011] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0012] An online dynamic adaptive compensation method for optical module performance degradation includes the following sub-steps:

[0013] The core performance parameters, equipment operating parameters, environmental sensing parameters and process-related parameters are obtained by the AD converter and integrated into multi-source heterogeneous parameters. The multi-source heterogeneous parameters are preprocessed to obtain preprocessed data.

[0014] An initial feature set is obtained by extracting features from preprocessed data, and the initial feature set is processed by a self-supervised model to obtain core degradation features;

[0015] The degradation trend prediction network uses temporal and spatial correlations to process core degradation features and obtain core performance prediction values. The compensation decision network uses action space, spatial state and reward factors to process core degradation features and core performance prediction values ​​and obtain the optimal compensation factor vector.

[0016] The production equipment is compensated and adjusted by using the optimal compensation factor vector and proportional-integral-differential equations. The actual values ​​of the compensation factors and the performance parameters of the optical modules after compensation are collected to obtain the performance improvement rate. The performance improvement rate, the performance parameters of the optical modules after compensation, the actual values ​​of the compensation factors are updated to update the self-supervised model, the degradation trend prediction network, and the compensation decision network.

[0017] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the acquisition of core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters through an AD converter, and the integration into multi-source heterogeneous parameters, includes the following steps:

[0018] The MCU master node issues a data acquisition command, and the FPGA synchronous control chip generates a synchronous clock, triggering the AD converters of all production workstations to acquire data, obtaining core performance parameters, equipment operating parameters, environmental perception parameters, and process-related parameters.

[0019] The AD converter dynamically adjusts the sampling frequency of data acquisition based on the deviation of key performance parameters and standard values.

[0020] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the AD converter performs data acquisition to obtain core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters, includes the following sub-steps:

[0021] The core performance parameters of the optical module are collected in real time by testing instruments. These core parameters include the output optical power of the optical module, the extinction ratio of the optical signal, the jitter amplitude of the eye diagram, and the insertion loss during the transmission of the optical signal.

[0022] The equipment operating parameters of the testing and production optical module equipment are collected in real time by sensors at each workstation; the equipment operating parameters include the working voltage of the welding equipment, the applied pressure of the packaging equipment, the working current of the testing instruments, and the vibration amplitude of the production equipment.

[0023] The environmental sensing parameters of the optical module production environment are collected in real time by an environmental sensor array; the environmental sensing parameters include the real-time temperature, relative humidity, and atmospheric pressure of the production environment.

[0024] The process-related parameters are fed back in real time by the process execution equipment on the optical module production line. These parameters include the deviation between the optical chip mounting position and the standard position, the temperature curve that changes with the relative time of the welding process, and the execution time of the optical module packaging process.

[0025] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the preprocessing of multi-source heterogeneous parameters to obtain preprocessed data includes the following sub-steps:

[0026] A bidirectional long short-term memory network based on an attention mechanism is used to repair missing values ​​of multi-source heterogeneous parameters.

[0027] After repairing the missing values, the multi-source heterogeneous parameters are normalized by adaptive maximum-minimum normalization to obtain normalized data;

[0028] The normalized data is denoised using a wavelet packet-attention joint method to obtain preprocessed data.

[0029] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the initial feature set is obtained by extracting features from preprocessed data, and the core degradation features are obtained by processing the initial feature set using a self-supervised model, includes the following sub-steps:

[0030] Automatically extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed data, and integrate them to obtain an initial feature set;

[0031] A self-supervised model is constructed based on a data augmentation module, an encoder, a momentum encoder, and a feature attention network. The self-supervised model processes the initial feature set to obtain the core degenerate features.

[0032] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the automatic extraction of time-domain features, frequency-domain features, and nonlinear features from preprocessed data, and the integration to obtain an initial feature set, includes the following steps:

[0033] Automatically extract the time-domain sequence of each time window from the preprocessed data, obtain the mean, variance, peak factor, kurtosis, and impulse factor of the time-domain sequence, and integrate them to obtain time-domain features;

[0034] Automatically extract frequency domain data from preprocessed data, obtain peak frequency, total harmonic distortion, and frequency band energy ratio of frequency domain data, and integrate to obtain frequency domain characteristics;

[0035] The system automatically extracts the data sequence for each time window from the preprocessed data, obtains the fractal dimension, approximate entropy, and sample entropy of the data sequence, and integrates them to obtain nonlinear features.

[0036] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the self-supervised model processes the initial feature set to obtain core degradation features, includes the following sub-steps:

[0037] The data augmentation module performs time stretching, amplitude perturbation, and window clipping enhancement operations on time-domain features, frequency-domain features, and nonlinear features respectively to obtain positive sample pairs.

[0038] The encoder maps positive sample pairs to obtain high-dimensional coding results of time-domain features, frequency-domain features, nonlinear features, and enhanced time-domain features, enhanced frequency-domain features, and enhanced nonlinear features;

[0039] Momentum encoders update their parameters using encoder parameters;

[0040] The contrastive loss functions for time-domain features, frequency-domain features, and nonlinear features are obtained by updating the momentum encoder and high-dimensional encoding results, and then integrated to obtain the total contrastive loss function.

[0041] The encoder is trained by the total contrastive loss function, and the encoder parameters are updated. After training, the encoder refines the time-domain features, frequency-domain features, and nonlinear features to obtain the refined high-dimensional features.

[0042] The purified high-dimensional features are adaptively weighted and fused using a feature attention network to obtain the core degenerate features.

[0043] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the degradation trend prediction network utilizes temporal and spatial correlation to process core degradation features and obtain core performance prediction values, includes the following sub-steps:

[0044] The input layer concatenates the core degradation features and the workstation association matrix to obtain spatiotemporal fusion features;

[0045] The temporal attention layer expands the spatiotemporal fusion features into a feature sequence by time step. The feature sequence is transformed linearly to obtain temporal query, temporal key, and temporal value. The temporal attention weights of the spatiotemporal fusion features are obtained using the temporal query and temporal key.

[0046] The spatial attention layer reorganizes the spatiotemporal fusion features into a workstation feature matrix according to the workstation dimension. The workstation feature matrix obtains spatial query, spatial key, and spatial value through linear transformation. The spatial attention weights of the spatiotemporal fusion features are obtained using the spatial query, spatial key, and workstation association matrix.

[0047] The hybrid attention layer weights and spatial attention weights are weighted and fused together, and spatial values ​​and temporal values ​​are weighted and fused together to obtain a hybrid attention weight and value matrix.

[0048] The output layer utilizes a hybrid attention weight and value matrix to output core performance predictions for future time intervals via a Transformer decoder.

[0049] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the compensation decision network uses action vectors, spatial vectors, and reward factors to process core degradation features and core performance prediction values ​​to obtain the optimal compensation factor vector includes the following sub-steps:

[0050] The state space layer combines the core performance prediction values, core performance standard values, core degradation features, and historical compensation records to obtain a space vector.

[0051] The motion space layer integrates the voltage compensation of the welding equipment, the pressure compensation of the packaging equipment, the welding temperature compensation, and the time compensation of the optical module packaging process to obtain the motion vector.

[0052] The reward function layer obtains multi-objective rewards by balancing the compensation effect, compensation cost, and security constraints;

[0053] The output module uses spatial vectors and action vectors to learn the optimal Q-function through a DQN network and obtain the optimal compensation factor vector.

[0054] Furthermore, an online dynamic adaptive compensation method for optical module performance degradation, wherein the production equipment is compensated and adjusted using an optimal compensation factor vector and proportional-integral-differential methods, the actual values ​​of the compensation factors and the performance parameters of the compensated optical modules are collected to obtain the performance improvement rate, and the self-supervised model, degradation trend prediction network, and compensation decision network are updated based on the performance improvement rate, the performance parameters of the compensated optical modules, and the actual values ​​of the compensation factors, includes the following sub-steps:

[0055] The optimal compensation factor vector is sent to the production equipment at each production station node through the MCU master node. The production equipment is adjusted by the proportional, integral and derivative functions of the servo controller to track the compensation command.

[0056] Collect actual values ​​of compensation factors and optical module performance parameters from all production workstations;

[0057] The performance improvement rate is obtained by comparing the optical module performance parameters before compensation, the optical module performance parameters after compensation, and the core performance standard values.

[0058] Feedback on performance improvement rate, compensated optical module performance parameters, and actual values ​​of compensation factors trigger the degradation trend prediction network to perform iterative parameter updates, the compensation decision network to perform experience replay to update parameters, and the attention weights of core degradation features to update.

[0059] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention obtains and preprocesses multi-source heterogeneous parameters through an AD converter, extracts core degradation features using a self-supervised model, and then uses a degradation trend prediction network to obtain core performance prediction values. This effectively captures the mutual influence between multiple processes and the evolution law of parameters, significantly improving the accuracy of performance degradation trend prediction. This invention automatically calculates the optimal compensation factor vector through a compensation decision network and uses proportional-integral-differential equations to precisely adjust the production equipment. It can dynamically adjust the compensation amount according to the actual degradation state, and collects the actual value after compensation and the improvement rate to update each network, forming a closed-loop adaptive optimization. This invention integrates scattered performance, equipment, environmental, and process parameters into unified multi-source heterogeneous parameters, and obtains core degradation features through a self-supervised model, realizing the effective fusion of data from each production link, enabling compensation decisions to be collaboratively optimized based on the global state of the production line. This invention integrates data acquisition, preprocessing, feature extraction, degradation prediction, compensation decision, execution adjustment, and feedback update into one complete closed loop of acquisition-processing-prediction-decision-execution-feedback, significantly reducing manual intervention and improving the stability and yield of the optical module production process. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 The flowchart shows an online dynamic adaptive compensation method for optical module performance degradation.

[0062] Figure 2This is a schematic diagram of the structure of a degradation trend prediction network.

[0063] Figure 3 A schematic diagram of the structure of a compensation decision network. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0065] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0066] like Figure 1 As shown, an online dynamic adaptive compensation method for optical module performance degradation includes the following steps:

[0067] S1: Obtain core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters through an AD converter, integrate them into multi-source heterogeneous parameters, preprocess the multi-source heterogeneous parameters, and obtain preprocessed data.

[0068] S11: Data is acquired via an AD converter. The sampling frequency is dynamically adjusted based on key performance parameters to obtain core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters. These are then integrated into multi-source heterogeneous parameters, as shown in the formula:

[0069] ;

[0070] in, For multi-source heterogeneous parameters, These are core performance parameters. For equipment operating parameters, For environmental sensing parameters, These are process-related parameters.

[0071] S111: The MCU master node issues a data acquisition command, the FPGA synchronous control chip generates a synchronous clock, and triggers the AD converters of all production workstations to acquire data, obtaining core performance parameters, equipment operating parameters, environmental perception parameters and process-related parameters.

[0072] It is important to note that the MCU master node controls all production station nodes (e.g., 8) via real-time Ethernet. Each production station node integrates a (16-bit high-precision) AD converter. The FPGA synchronization control chip generates a synchronization clock, thus achieving clock synchronization between the MCU master node and all production station nodes (slave nodes) and collecting synchronization errors. Real-time Ethernet transmission latency ≤10ms.

[0073] S1111: Real-time acquisition of core performance parameters of the optical module using testing instruments, the formula is:

[0074] ;

[0075] in, To output optical power to the optical module, The extinction ratio of the optical signal. The jitter amplitude of the eye diagram. This refers to the insertion loss during optical signal transmission.

[0076] S1112: Real-time acquisition of equipment operating parameters for testing and production optical modules via sensors at each workstation, using the following formula:

[0077] ;

[0078] in, This refers to the operating voltage of the welding equipment. Applying pressure to the packaging equipment, To test the operating current of the instrument, The vibration amplitude of the production equipment;

[0079] The production equipment includes welding equipment, packaging equipment, testing instruments, and other online production equipment.

[0080] Among the sensors at each workstation, the pressure sensor has the following accuracy: The temperature sensor accuracy is .

[0081] S1113: Real-time acquisition of environmental sensing parameters of the optical module production environment via an environmental sensor array, using the following formula:

[0082] ;

[0083] in, For the real-time temperature of the production environment, The relative humidity of the production environment. Atmospheric pressure in the production environment;

[0084] S1114: Process-related parameters are fed back in real time by the process execution equipment on the optical module production line, using the following formula:

[0085] ;

[0086] in, This refers to the deviation between the mounting position of the optical chip and the standard position. For the relative time of the welding process The changing temperature curve, This refers to the execution time of the optical module packaging process.

[0087] S112: The AD converter dynamically adjusts the data acquisition sampling frequency based on the deviation of key performance parameters and standard values, using the following formula:

[0088] ;

[0089] in, Let be the dynamic sampling frequency at time t, where t is the time index. As the reference sampling frequency, It is a natural exponential function. This is the sensitivity adjustment coefficient. These are the key performance parameters at time t. These are the standard values ​​for key performance parameters. This is the minimum sampling frequency.

[0090] In the embodiments, , , Key performance parameters selected: optical module output optical power .

[0091] S12: Perform preprocessing on the multi-source heterogeneous parameters, including missing value imputation, normalization, and denoising, to obtain the preprocessed data.

[0092] S121: A bidirectional long short-term memory network based on an attention mechanism, which repairs missing values ​​of multi-source heterogeneous parameters, using the following formula:

[0093] ;

[0094] in, The missing values ​​after repair at time t. For the bidirectional long short-term memory network operation of the attention mechanism, For target parameters (e.g., optical module output optical power) The time-series window data, The time window length, For target-related parameters at time t (e.g., optical module output optical power) When time t is missing, the extinction ratio of the light signal at time t. Working voltage of welding equipment (As a target-related parameter to aid prediction).

[0095] In the embodiments, .

[0096] S122: After repairing missing values, the multi-source heterogeneous parameters are normalized using adaptive max-min normalization to obtain the normalized data. The formula is as follows:

[0097] ;

[0098] in, The data at time t is the normalized data. The original data at time t (which may be the data after missing values ​​have been repaired) Any parameter within it, such as the extinction ratio of the optical signal after repairing missing values). To obtain the minimum value, From The original data sequence from time t to time t. The length of the sliding window. To obtain the maximum value, For the normalized upper limit, This is the lower bound for normalization.

[0099] In the embodiments, , , .

[0100] S123: Denoise the normalized data using a wavelet packet-attention joint method to obtain the preprocessed data, as shown in the formula:

[0101] ;

[0102] in, Let be the preprocessed (denoised) data at time t, and let i be the index of the wavelet packet decomposition layer. The total number of wavelet packet decomposition layers. Let be the dynamic attention weights of the i-th layer decomposition result at time t. , For wavelet packet transform, This represents the decomposition level of the wavelet packet transform.

[0103] S2: Extract features from preprocessed data to obtain an initial feature set, and process the initial feature set using a self-supervised model to obtain core degenerate features;

[0104] S21: Automatically extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed data, and integrate them to obtain an initial feature set. The formula is as follows:

[0105] ;

[0106] in, For the initial feature set, For time domain characteristics, For frequency domain characteristics, It exhibits nonlinear characteristics.

[0107] S211: Automatically extract the time-domain sequence of each time window from the preprocessed data (i.e., the sequence composed of time-domain data of each time window), obtain the mean, variance, peak factor, kurtosis, and impulse factor of the time-domain sequence, and integrate them to obtain the time-domain features. The formula is:

[0108] ;

[0109] in, The mean of the time-domain sequence. Let Variance be the variance of the time-domain sequence. The peak factor of the time-domain sequence, For the kurtosis of the time-domain sequence, is the pulse factor of the time-domain sequence.

[0110] S212: Automatically extract frequency domain data from the preprocessed data (obtained by performing a Fourier transform on the time series), acquire the peak frequency, total harmonic distortion, and frequency band energy ratio of the frequency domain data, and integrate them to obtain frequency domain characteristics. The formula is:

[0111] ;

[0112] in, The peak frequency of the frequency domain data. For the total harmonic distortion of frequency domain data, This represents the frequency band energy percentage of the frequency domain data.

[0113] It is important to note that peak frequency refers to the frequency with the highest energy in the power spectral density, harmonic distortion refers to the ratio of the total harmonic amplitude to the fundamental amplitude, and band energy ratio refers to the ratio of the energy within the target band to the total energy.

[0114] S213: Automatically extract the data sequence for each time window from the preprocessed data, obtain the fractal dimension (obtained through box counting), approximate entropy, and sample entropy of the data sequence, and integrate them to obtain nonlinear features. The formula is:

[0115] ;

[0116] in, Let fractal dimension be the data sequence. The approximate entropy of the data sequence. denoted as the sample entropy of the data sequence.

[0117] It should be noted that the data sequence is derived from the time-domain sequence. This indicates the complexity of the data sequence. This indicates the degree of irregularity in the data sequence. It is less sensitive to the length of the data sequence.

[0118] S22: A self-supervised model is constructed based on a data augmentation module, an encoder, a momentum encoder, and a feature attention network. The self-supervised model processes the initial feature set to obtain the core degenerate features.

[0119] S221: The data augmentation module performs time stretching, amplitude perturbation, and window clipping enhancement operations on time-domain features, frequency-domain features, and nonlinear features respectively to obtain positive sample pairs. , , ;

[0120] in, To enhance time-domain features, To enhance frequency domain characteristics, To enhance nonlinear characteristics;

[0121] S222: The encoder maps positive sample pairs to obtain high-dimensional coding results of time-domain features, frequency-domain features, nonlinear features, and enhanced time-domain features, enhanced frequency-domain features, and enhanced nonlinear features. The formula is as follows:

[0122] ;

[0123] in, The result is a high-dimensional encoding of time-domain features. Mapping for the encoder (consisting of a 3-layer convolutional neural network and a 2-layer fully connected network), To enhance the high-dimensional coding results of time-domain features, The high-dimensional coding result of frequency domain features To enhance the high-dimensional coding results of frequency domain features, This is a high-dimensional encoding result for nonlinear features. To enhance the high-dimensional coding results of nonlinear features.

[0124] S223: The momentum encoder updates its parameters using encoder parameters, as shown in the formula:

[0125] ;

[0126] in, For updated momentum encoder parameters, The momentum coefficient, For momentum encoder parameters, For encoder parameters;

[0127] In the embodiments, .

[0128] S224: The contrastive loss functions for time-domain features, frequency-domain features, and nonlinear features are obtained through the updated momentum encoder and high-dimensional encoding results. The total contrastive loss function is then obtained by ensembling these functions. The formula is as follows:

[0129] ;

[0130] in, The contrastive loss function is for temporal features. To enhance the high-dimensional encoding results of temporal features mapped by the momentum encoder, The total number of negative samples. Index for negative samples For temperature coefficient, The dynamic encoding feature of the k-th negative sample in the time domain. The contrastive loss function is for frequency domain features. To enhance the high-dimensional encoding results of frequency domain features mapped by the momentum encoder, The dynamic coding feature of the k-th negative sample in the frequency domain. For the contrastive loss function of nonlinear features, To enhance the high-dimensional encoding results of nonlinear features mapped by the momentum encoder, The dynamic encoding feature of the k-th negative sample is nonlinear. The total contrastive loss function is... This is a vector inner product multiplication.

[0131] In the embodiments, , .

[0132] S225: The encoder is trained using the total contrastive loss function, and the encoder parameters are updated. After training, the encoder refines the time-domain features, frequency-domain features, and nonlinear features to obtain the refined high-dimensional features. The formula is as follows:

[0133] ;

[0134] in, For the purified high-dimensional time-domain features, For encoder mapping after training, For the purified high-dimensional frequency domain features, This refers to the purified high-dimensional nonlinear features.

[0135] S226: The purified high-dimensional features are adaptively weighted and fused using a feature attention network to obtain the core degenerate features. The formula is as follows:

[0136] ;

[0137] in, As the core degradation feature, For feature attention network operations, for The corresponding attention weights .

[0138] S3: The degradation trend prediction network uses temporal and spatial correlations to process core degradation features and obtain core performance prediction values. The compensation decision network uses action vectors, spatial vectors and reward factors to process core degradation features and core performance prediction values ​​and obtain the optimal compensation factor vector.

[0139] S31: A degradation trend prediction network is constructed based on the input layer, temporal attention layer, spatial attention layer, hybrid attention layer, and output layer. The degradation trend prediction network processes the core degradation features to obtain the core performance prediction value in the future time.

[0140] like Figure 2 As shown, the first output of the input layer is connected to the input of the temporal attention layer, the second output of the input layer is connected to the input of the spatial attention layer, the output of the temporal attention layer is connected to the first input of the hybrid attention layer, the output of the spatial attention layer is connected to the second input of the hybrid attention layer, and the output of the hybrid attention layer is connected to the input of the output layer.

[0141] S311: The input layer concatenates the core degradation features and the workstation association matrix to obtain spatiotemporal fusion features, as shown in the formula:

[0142] ;

[0143] in, As a feature of spatiotemporal fusion, For splicing, This is the workstation correlation matrix. , Let 8 be the set of real numbers, and 8 be the total number of production workstations.

[0144] In the embodiment, the workstation association matrix Inner element Representative production station With production station The correlation strength is initialized by the production process logic. For example, the correlation strength between the welding station and the packaging station is higher than that between the welding station and the testing station, and is dynamically updated through data feedback.

[0145] S312: The temporal attention layer expands the spatiotemporal fusion features into a feature sequence over time steps. The feature sequence undergoes a linear transformation to obtain the temporal query, temporal key, and temporal value. The temporal attention weights of the spatiotemporal fusion features are then obtained using the temporal query and temporal key, as shown in the formula:

[0146] ;

[0147] in, For time-series queries, This is the weight matrix for time-series queries. For timing key, This is the weight matrix for the time series keys. These are timing values. The weight matrix for time series values, For temporal attention weights, For normalized activation functions, For matrix transpose, For feature dimensions.

[0148] In the embodiments, .

[0149] S313: The spatial attention layer reorganizes the spatiotemporal fusion features into a workstation feature matrix H according to the workstation dimension. The workstation feature matrix obtains spatial queries, spatial keys, and spatial values ​​through linear transformation. The spatial attention weights of the spatiotemporal fusion features are obtained using the spatial queries, spatial keys, and workstation association matrix, as shown in the formula:

[0150] ;

[0151] in, For spatial queries, This is the weight matrix for spatial queries. For space keys, This is the weight matrix of the spatial bonds. For spatial values, The weight matrix for the spatiotemporal values, For spatial attention weights.

[0152] S314: The hybrid attention layer weights and spatial attention weights are weighted and fused together, and spatial and temporal values ​​are weighted and fused together to obtain a hybrid attention weight and value matrix, as shown in the formula:

[0153] ;

[0154] in, For mixed attention weights, These are the time-series weighting coefficients. It is a value matrix.

[0155] In the embodiments, It can be dynamically adjusted according to the actual degradation mode, but always Features have already undergone dimensional alignment or projection before fusion.

[0156] S315: The output layer utilizes a hybrid attention weight and value matrix to output the core performance prediction values ​​for future time intervals through the Transformer decoder, as shown in the formula:

[0157] ;

[0158] in, for Real-time performance core prediction values Indexing future time steps For Transformer decoder decoding operations, This is a position encoding used to supplement temporal position information. This is the output layer weight matrix. This is the output layer bias vector. for The predicted output optical power value of the time-of-flight optical module. for The predicted extinction ratio of the optical signal at any given time. for The predicted jitter amplitude of the eye diagram at any given time.

[0159] It is important to note that the insertion loss during optical signal transmission is not predicted because it mainly depends on static coupling and connector loss in the optical path, and is usually almost constant or changes very little.

[0160] S32: A compensation decision network is constructed based on the state space layer, action space layer, reward function layer, and output module. The compensation decision network processes the core performance prediction value and core degradation feature to obtain the optimal compensation factor vector.

[0161] like Figure 3As shown, the output of the state space layer is connected to the first input of the output module, the output of the action space layer is connected to the second input of the output module, the output of the reward function layer is connected to the third input of the output module, and the output of the output module outputs the optimal compensation factor vector.

[0162] S321: The state space layer obtains a space vector from the set of performance core predicted values, performance core standard values, core degradation features, and historical compensation records, using the following formula: ;

[0163] in, For spatial vectors, As the core performance standard value, For historical compensation records, The standard value for output optical power of the optical module. This is the standard value for the extinction ratio of the optical signal. This is the standard value for the jitter amplitude of the eye diagram.

[0164] In the embodiments, Store the most recent 10 optimal compensation factor vectors and their corresponding performance improvement rates.

[0165] S322: The action space layer integrates the voltage compensation of the welding equipment, the pressure compensation of the packaging equipment, the welding temperature compensation, and the time compensation of the optical module packaging process to obtain the action vector, with the formula as follows:

[0166] ;

[0167] in, For action vectors, This refers to the voltage compensation amount for the welding equipment. This refers to the pressure compensation amount for the packaging equipment. This is the welding temperature compensation amount. This is the time compensation amount for the optical module packaging process.

[0168] It is important to note that each compensation amount has a set safety threshold range to avoid overcompensation. .

[0169] S323: The reward function layer obtains multi-objective rewards by balancing compensation effects, compensation costs, and safety constraints. The formula is as follows:

[0170] ;

[0171] in, Let t be the multi-objective reward at time t. The weighting coefficients for the compensation effect, The weighting coefficient for compensating for costs, To compensate for the amplitude, For the weighting coefficients of safety constraints, For indicator functions, when Beyond safe range hour, ,otherwise , This refers to the lower threshold of core performance parameters. This refers to the upper limit threshold of core performance parameters. The lower limit threshold for the output optical power of the optical module. This is the lower limit threshold of the extinction ratio of the optical signal. This is the lower limit threshold for the jitter amplitude of the eye diagram. The upper limit threshold for the output optical power of the optical module. This represents the upper limit threshold of the extinction ratio of the optical signal. This is the upper limit threshold for the jitter amplitude of the eye diagram.

[0172] In the embodiments, , , .

[0173] It is important to note that using As a supervisory signal, the parameters of the DQN network are updated through the DQN algorithm, enabling the Q function to accurately evaluate the cumulative reward of the state-action pair.

[0174] S324: The output module uses spatial vectors and action vectors to learn the optimal Q-function through a DQN network, obtaining the optimal compensation factor vector. The formula is as follows:

[0175] ;

[0176] in, This represents the optimal compensation factor vector. In action space Finding action vectors , making The value is the largest. For the Q function, For the parameters of the DQN network, Used to evaluate spatial vectors Execution action vector Accumulated rewards.

[0177] S4: The production equipment is compensated and adjusted by the optimal compensation factor vector and proportional-integral-differential equation. The actual value of the compensation factor and the performance parameters of the optical module after compensation are collected to obtain the performance improvement rate. The performance improvement rate, the performance parameters of the optical module after compensation, and the actual value of the compensation factor are fed back to update the self-supervised model, the degradation trend prediction network, and the compensation decision network.

[0178] S41: The optimal compensation factor vector is sent as a compensation command to the production equipment at each production station node via the MCU master node. The production equipment is adjusted using the proportional-integral-derivative (PI) function of the servo controller to track the compensation command. The formula is:

[0179] ;

[0180] in, The control signal of the servo controller at time t (the servo controller outputs a control signal to drive the actuator, such as a stepper motor in welding equipment or a servo motor in packaging equipment). The proportionality coefficient is the derivative of the proportional integral. The integral coefficient of the derivative of the proportional integral. The differential coefficients of the proportional integral are... The compensation factor deviation at time t, Let t be the actual value of the compensation factor. The integral term of the deviation, It is a time variable that changes continuously from 0 to t. This is the differential term of the deviation.

[0181] It is important to note that the optimal compensation factor vector needs to be converted into a binary data string that the device can recognize according to the Modbus-RTU protocol format, and a CRC-16 checksum needs to be added to the end to ensure reliable and error-free data transmission.

[0182] S42: Collect the actual values ​​of the compensation factor and optical module performance parameters from all production workstations. The formula is:

[0183] ;

[0184] in, Here are the performance parameters of the optical module after compensation at time t. The output optical power of the optical module after compensation at time t. Let be the extinction ratio of the compensated optical signal at time t. Let be the jitter amplitude of the eye diagram after compensation at time t. Let be the actual voltage compensation amount of the welding equipment at time t. Let t be the actual pressure compensation amount of the packaging equipment at time t. This is the actual compensation amount for welding temperature. This represents the actual time compensation amount for the optical module packaging process.

[0185] S43: The performance improvement rate is obtained by using the optical module performance parameters before compensation, the optical module performance parameters after compensation, and the core performance standard value. The formula is as follows:

[0186] ;

[0187] in, Let be the optical module performance parameters before compensation at time t. The output optical power of the optical module before compensation at time t. Let be the extinction ratio of the optical signal before compensation at time t. Let t be the jitter amplitude of the eye diagram before compensation. For performance improvement rate.

[0188] It should be noted that if If the compensation effect is poor, the baseline sampling frequency is increased to twice the original frequency, and the compensation decision network update is shortened to 50 sampling cycles.

[0189] S44: Feedback on performance improvement rate, compensated optical module performance parameters, and actual value of compensation factor triggers the degradation trend prediction network to perform iterative parameter updates, the compensation decision network to perform experience replay to update parameters, and updates the attention weights of core degradation features.

[0190] It also includes iterative training of the degradation trend prediction network, which adopts an incremental learning + regularization iterative strategy. The degradation trend prediction network is updated every 100 sampling periods, including the following sub-steps:

[0191] S51: Introducing a sliding window dataset, Z samples are drawn from the sliding window set at each update, with a focus on penalizing large-bias predictions, to obtain an improved weighted average absolute error loss for the degradation trend prediction network, as shown in the formula:

[0192] ;

[0193] in, The improved weighted average absolute error loss is used for the degradation trend prediction network, where Z is the total number of samples drawn, and z is the sample number. The weight of the z-th sample. Let z be the core performance prediction value of the z-th sample. This represents the actual performance core value of the z-th sample. As the core performance standard value, The optical module outputs the predicted optical power value for the z-th sample. Let z be the predicted extinction ratio of the optical signal of the z-th sample. Let be the predicted jitter amplitude value of the eye diagram for the z-th sample. The optical module outputs optical power for the z-th sample. Let z be the extinction ratio of the light signal of the z-th sample. Let be the jitter amplitude of the eye diagram of the z-th sample;

[0194] It should be noted that the sliding window size is set to 1000 samples.

[0195] S52: Obtain the gradient of the improved weighted average absolute error loss with respect to the degradation trend prediction network parameters through backpropagation. Update the degradation trend prediction network parameters using stochastic gradient descent and regularization, as shown in the formula:

[0196] ;

[0197] in, To update the parameters of the degradation trend prediction network, Predict network parameters to show the degradation trend before updating. For learning rate, To predict the gradient of network parameters in the degradation trend before the update using the improved weighted average absolute error loss, The L2 regularization coefficient is... To predict network parameters for degradation trends Find the gradient.

[0198] In the embodiments, , .

[0199] It is important to note that only the latest sampled data should be used for iterative training to avoid degradation trend prediction network drift caused by outdated historical data.

[0200] The beneficial effects of this invention are as follows:

[0201] 1. This invention uses the core performance parameters, equipment operating parameters, environmental perception parameters, and process-related parameters obtained by the testing instruments, sensors at each workstation, environmental sensor arrays, and process execution equipment to form a complete data acquisition system covering four dimensions: performance, equipment, environment, and process. Moreover, the process-related parameters are fed back in real time.

[0202] 2. This invention employs a bidirectional long short-term memory network based on an attention mechanism to repair missing values. It utilizes temporal window data before and after the target parameter, as well as contemporaneous information of other relevant parameters, to fully explore temporal and cross-parameter correlations, thereby improving filling accuracy. This invention uses adaptive max-min normalization, taking the local extrema within the sliding window as the normalization benchmark, so that the normalization parameters are adjusted in real time according to the data distribution, adapting to the non-stationary characteristics of the production process. This invention uses wavelet packet attention joint denoising, dividing the signal into multiple frequency bands through wavelet packet decomposition, and then using the attention mechanism to dynamically assign weights to each frequency band, with high weights for signal-dominant frequency bands and low weights for noise-dominant frequency bands, to achieve adaptive denoising.

[0203] 3. This invention automatically extracts three complementary features—time domain, frequency domain, and nonlinearity—covering three dimensions: signal amplitude statistics, frequency energy distribution, and system complexity, thus forming a more complete description of the degradation process. Furthermore, this invention utilizes a self-supervised model constructed through data augmentation, encoders, momentum encoders, and feature attention networks. This model can automatically learn the most robust and discriminative core degradation features from initial features, overcoming the limitations of manual feature design and significantly improving the accuracy and generalization ability of subsequent predictions and decisions.

[0204] 5. This invention adds peak factor, kurtosis, and impulse factor to the time-domain features to enhance the ability to perceive abnormal impacts and waveform distortions; at the same time, it introduces peak frequency, total harmonic distortion, and frequency band energy ratio from the frequency-domain features to locate the source of harmonic interference and energy drift; and it introduces fractal dimension, approximate entropy, and sample entropy from the nonlinear features to quantify the complexity of the optical module. The time-domain features, frequency-domain features, and nonlinear features work together to construct a complete feature system from amplitude statistics and frequency structure to nonlinear dynamics.

[0205] 6. This invention uses a data augmentation module to perform time stretching, amplitude perturbation, and window clipping on time-domain features, frequency-domain features, and nonlinear features respectively, generating positive sample pairs and enriching the diversity of feature representations; it uses an encoder to map the original and augmented features to a high-dimensional space, and then uses a momentum encoder to update the momentum through encoder parameters, stabilizing the target of contrastive learning.

[0206] 7. The combined method of data augmentation, contrastive learning purification and attention fusion in this invention effectively removes redundancy and noise in the initial features, making the core degradation features more sensitive and accurate to the degradation trend of the optical module. The feature attention network can adaptively adjust the weights of each dimension of features according to the dynamic changes of the degradation mode, thereby improving the performance of degradation prediction and compensation decision.

[0207] 8. The mechanism of the present invention, which combines temporal attention and spatial attention, enables the degradation trend prediction network to simultaneously utilize historical change patterns and cross-workstation coupling information for prediction, significantly improving prediction accuracy and robustness.

[0208] 9. This invention integrates the predicted performance core value, the standard performance core value, the core degradation features, and historical compensation records into a spatial vector through a state space layer, providing complete current state and historical experience information for compensation decisions. Through an action space layer, it integrates the voltage compensation of the welding equipment, the pressure compensation of the packaging equipment, the welding temperature compensation, and the time compensation of the optical module packaging process into an action vector, clarifying the adjustable action dimensions. Through a reward function layer, it balances the compensation effect (the closeness of the predicted value to the standard value), the compensation cost (the magnitude of each compensation amount), and the safety constraints (whether the predicted value exceeds the safety range), calculating a multi-objective reward as a learning signal for reinforcement learning. The output module uses the spatial vector and action vector to learn the optimal Q-function through a DQN network, obtaining the optimal compensation factor vector that maximizes the cumulative reward. It can automatically learn the optimal compensation strategy under different states, achieving a multi-objective balance between compensation effect, cost, and safety. Furthermore, it continuously optimizes decisions using historical compensation records, significantly improving the accuracy and adaptability of compensation.

[0209] 10. This invention distributes the optimal compensation factor vector to the production equipment at each production station node via the MCU master node. The production equipment uses a servo controller for proportional-integral-derivative closed-loop adjustment to track compensation commands in real time, ensuring that the compensation amount is executed accurately. The actual values ​​of the compensation factors and the performance parameters of the optical modules after compensation are collected from all production station nodes. The performance improvement rate is calculated using the performance parameters before and after compensation, along with the core performance standard value, quantifying the compensation effect. The performance improvement rate, the performance parameters of the optical modules after compensation, and the actual values ​​of the compensation factors are fed back to the system, triggering the degradation trend prediction network to iteratively update parameters (making predictions more accurate) and the compensation decision network to perform experience replay to update parameters (making decisions better), while simultaneously updating the attention weights of the core degradation features. This invention's closed-loop feedback mechanism enables online iterative optimization using actual execution results, while PID closed-loop adjustment ensures the accurate execution of compensation commands, achieving a complete closed loop of decision-making-execution-collection-feedback-update.

[0210] 11. This invention establishes a master-slave architecture by connecting the MCU master node with each production station node. The MCU master node uniformly issues acquisition commands and generates a synchronization clock in conjunction with the FPGA synchronization control chip, triggering the AD converters of all production station nodes to simultaneously acquire data. This achieves precise time synchronization of data from each station and breaks down data silos between stations. This invention dynamically adjusts the sampling frequency based on the deviation of key performance parameters from standard values. When parameters are normal, low-frequency sampling is used to save resources, and when parameters are abnormal, the sampling frequency is increased to capture details. This avoids the generation of redundant data and prevents the loss of key data.

[0211] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0212] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An online dynamic adaptive compensation method for optical module performance degradation, characterized in that, Includes the following sub-steps: The core performance parameters, equipment operating parameters, environmental sensing parameters and process-related parameters are obtained by the AD converter and integrated into multi-source heterogeneous parameters. The multi-source heterogeneous parameters are preprocessed to obtain preprocessed data. An initial feature set is obtained by extracting features from preprocessed data, and the initial feature set is processed by a self-supervised model to obtain core degradation features; The degradation trend prediction network uses temporal and spatial correlations to process core degradation features and obtain core performance prediction values. The compensation decision network uses action space, spatial state and reward factors to process core degradation features and core performance prediction values ​​and obtain the optimal compensation factor vector. The production equipment is compensated and adjusted by using the optimal compensation factor vector and proportional-integral-differential equations. The actual values ​​of the compensation factors and the performance parameters of the optical modules after compensation are collected to obtain the performance improvement rate. The performance improvement rate, the performance parameters of the optical modules after compensation, the actual values ​​of the compensation factors are updated to update the self-supervised model, the degradation trend prediction network, and the compensation decision network.

2. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The process of acquiring core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters through an AD converter and integrating them into multi-source heterogeneous parameters includes the following steps: The MCU master node issues a data acquisition command, and the FPGA synchronous control chip generates a synchronous clock, triggering the AD converters of all production workstations to acquire data, obtaining core performance parameters, equipment operating parameters, environmental perception parameters, and process-related parameters. The AD converter dynamically adjusts the sampling frequency of data acquisition based on the deviation of key performance parameters and standard values.

3. The online dynamic optical module performance degradation adaptive compensation method according to claim 2, characterized in that, The AD converter performs data acquisition to obtain core performance parameters, equipment operating parameters, environmental sensing parameters, and process-related parameters, including the following sub-steps: The core performance parameters of the optical module are collected in real time by testing instruments. These core parameters include the output optical power of the optical module, the extinction ratio of the optical signal, the jitter amplitude of the eye diagram, and the insertion loss during the transmission of the optical signal. The equipment operating parameters of the testing and production optical module equipment are collected in real time by sensors at each workstation; the equipment operating parameters include the working voltage of the welding equipment, the applied pressure of the packaging equipment, the working current of the testing instruments, and the vibration amplitude of the production equipment. The environmental sensing parameters of the optical module production environment are collected in real time by an environmental sensor array; the environmental sensing parameters include the real-time temperature, relative humidity, and atmospheric pressure of the production environment. The process-related parameters are fed back in real time by the process execution equipment on the optical module production line. These parameters include the deviation between the optical chip mounting position and the standard position, the temperature curve that changes with the relative time of the welding process, and the execution time of the optical module packaging process.

4. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous parameters to obtain preprocessed data includes the following sub-steps: A bidirectional long short-term memory network based on an attention mechanism is used to repair missing values ​​of multi-source heterogeneous parameters. After repairing the missing values, the multi-source heterogeneous parameters are normalized by adaptive maximum-minimum normalization to obtain normalized data; The normalized data is denoised using a wavelet packet-attention joint method to obtain preprocessed data.

5. The online dynamic adaptive compensation method for optical module performance degradation according to claim 1, characterized in that, The process of extracting features from preprocessed data to obtain an initial feature set, and then processing the initial feature set using a self-supervised model to obtain core degradation features, includes the following sub-steps: Automatically extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed data, and integrate them to obtain an initial feature set; A self-supervised model is constructed based on a data augmentation module, an encoder, a momentum encoder, and a feature attention network. The self-supervised model processes the initial feature set to obtain the core degenerate features.

6. The online dynamic optical module performance degradation adaptive compensation method according to claim 5, characterized in that, The automatic extraction of time-domain features, frequency-domain features, and nonlinear features from preprocessed data, and the integration to obtain an initial feature set, includes the following steps: Automatically extract the time-domain sequence of each time window from the preprocessed data, obtain the mean, variance, peak factor, kurtosis, and impulse factor of the time-domain sequence, and integrate them to obtain time-domain features; Automatically extract frequency domain data from preprocessed data, obtain peak frequency, total harmonic distortion, and frequency band energy ratio of frequency domain data, and integrate to obtain frequency domain characteristics; The system automatically extracts the data sequence for each time window from the preprocessed data, obtains the fractal dimension, approximate entropy, and sample entropy of the data sequence, and integrates them to obtain nonlinear features.

7. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The self-supervised model processes the initial feature set to obtain core degradation features, including the following sub-steps: The data augmentation module performs time stretching, amplitude perturbation, and window clipping enhancement operations on time-domain features, frequency-domain features, and nonlinear features respectively to obtain positive sample pairs. The encoder maps positive sample pairs to obtain high-dimensional coding results of time-domain features, frequency-domain features, nonlinear features, and enhanced time-domain features, enhanced frequency-domain features, and enhanced nonlinear features; Momentum encoders update their parameters using encoder parameters; The contrastive loss functions for time-domain features, frequency-domain features, and nonlinear features are obtained by updating the momentum encoder and high-dimensional encoding results, and then integrated to obtain the total contrastive loss function. The encoder is trained by the total contrastive loss function, and the encoder parameters are updated. After training, the encoder refines the time-domain features, frequency-domain features, and nonlinear features to obtain the refined high-dimensional features. The purified high-dimensional features are adaptively weighted and fused using a feature attention network to obtain the core degenerate features.

8. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The degradation trend prediction network utilizes temporal and spatial correlations to process core degradation features and obtain core performance prediction values, including the following sub-steps: The input layer concatenates the core degradation features and the workstation association matrix to obtain spatiotemporal fusion features; The temporal attention layer expands the spatiotemporal fusion features into a feature sequence by time step. The feature sequence is transformed linearly to obtain temporal query, temporal key, and temporal value. The temporal attention weights of the spatiotemporal fusion features are obtained using the temporal query and temporal key. The spatial attention layer reorganizes the spatiotemporal fusion features into a workstation feature matrix according to the workstation dimension. The workstation feature matrix obtains spatial query, spatial key, and spatial value through linear transformation. The spatial attention weights of the spatiotemporal fusion features are obtained using the spatial query, spatial key, and workstation association matrix. The hybrid attention layer weights and spatial attention weights are weighted and fused together, and spatial values ​​and temporal values ​​are weighted and fused together to obtain a hybrid attention weight and value matrix. The output layer utilizes a hybrid attention weight and value matrix to output core performance predictions for future time intervals via a Transformer decoder.

9. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The compensation decision network uses action vectors, spatial vectors, and reward factors to process core degradation features and core performance prediction values, and obtains the optimal compensation factor vector through the following sub-steps: The state space layer combines the core performance prediction values, core performance standard values, core degradation features, and historical compensation records to obtain a space vector. The motion space layer integrates the voltage compensation of the welding equipment, the pressure compensation of the packaging equipment, the welding temperature compensation, and the time compensation of the optical module packaging process to obtain the motion vector. The reward function layer obtains multi-objective rewards by balancing the compensation effect, compensation cost, and security constraints; The output module uses spatial vectors and action vectors to learn the optimal Q-function through a DQN network and obtain the optimal compensation factor vector.

10. The online dynamic optical module performance degradation adaptive compensation method according to claim 1, characterized in that, The process of adjusting the production equipment through the optimal compensation factor vector and proportional-integral-differential equations, collecting the actual values ​​of the compensation factors and the performance parameters of the optical modules after compensation, obtaining the performance improvement rate, and updating the self-supervised model, degradation trend prediction network, and compensation decision network with the performance improvement rate, the performance parameters of the optical modules after compensation, and the actual values ​​of the compensation factors includes the following sub-steps: The optimal compensation factor vector is sent to the production equipment at each production station node through the MCU master node. The production equipment is adjusted by the proportional, integral and derivative functions of the servo controller to track the compensation command. Collect actual values ​​of compensation factors and optical module performance parameters from all production workstations; The performance improvement rate is obtained by comparing the optical module performance parameters before compensation, the optical module performance parameters after compensation, and the core performance standard values. Feedback on performance improvement rate, compensated optical module performance parameters, and actual values ​​of compensation factors trigger the degradation trend prediction network to perform iterative parameter updates, the compensation decision network to perform experience replay to update parameters, and the attention weights of core degradation features to update.

Images