A deep learning-based DPMZM modulator fast calibration system and method
By using a one-dimensional convolutional neural network with segmented decoupled scanning and multi-scale feature aggregation, combined with a physical perception hybrid loss function, the problems of slow speed and low accuracy of DPMZM calibration are solved, and fast and accurate calibration is achieved in high-speed optical communication systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA JILIANG UNIV
- Filing Date
- 2026-06-11
- Publication Date
- 2026-07-14
AI Technical Summary
Existing DPMZM calibration methods are slow and cannot meet the real-time control requirements of high-speed optical communication systems. Furthermore, deep learning calibration suffers from problems such as periodic multi-valued mapping, symmetric sign ambiguity, and loss function mismatch, resulting in low calibration accuracy.
A segmented decoupled scanning strategy is adopted to collect optical power data, construct a training dataset, and establish a mapping relationship between the output of the optical modulator and the control parameters through a one-dimensional convolutional neural network with multi-scale feature aggregation. A physical perception hybrid loss function is defined, and an adaptive optimization algorithm is used to ensure model convergence, thereby achieving fast and accurate calibration.
It has improved the calibration speed from seconds to milliseconds, resolved the physical property conflicts in deep learning calibration, improved calibration accuracy and model robustness, and met the engineering requirements of high-speed optical communication systems.
Smart Images
Figure CN122394664A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of optical communication device testing and control technology, and in particular relates to a fast calibration system and method for DPMZM modulators based on deep learning. Background Technology
[0002] The dual parallel Mach-Zehnder modulator (DPMZM), as a core photonic device in modern high-speed coherent optical communication systems, plays a decisive role in signal generation for high-order modulation formats such as QPSK and 16QAM. The DPMZM employs a nested topology, consisting of two sub-modulators and a master modulator connected in parallel. To ensure high-quality signal modulation performance, precise characterization and control of the device's key physical parameters are essential, specifically including three half-wave voltage parameters (…). ) and three DC bias operating points ( Accurately obtaining these parameters is of significant engineering importance.
[0003] Patent CN113395111A proposes a scheme for preliminary confirmation of the modulator's operating point based on the modulation phase without radio frequency input; however, its scaling speed is significantly slower than the traditional controlled variable method, which sequentially scans and adjusts the bias voltages of the three modulators to search for the maximum output optical power of the modulator. 3 While the scanning time is faster, the total time to complete the calibration process is still in the second range, making further improvement difficult and limiting the upper limit of calibration speed to some extent.
[0004] Some researchers have attempted to use deep neural networks (DNNs) or convolutional neural networks (CNNs) to establish a mapping relationship between scan curves and device parameters. However, directly using the raw voltage values as labels for end-to-end regression training yields extremely poor results in practical applications, mainly due to the conflict between the physical characteristics and the algorithm's principles.
[0005] The conflict between periodicity and gradient caused by "multi-valued mapping": The physical nature of optical phase is periodic. (Cycle). Bias voltage and The corresponding output spectra or scanning curves are completely identical. (Among them) This is the bias voltage. (where k is an integer and k is the half-wave voltage) For neural networks, this means that the same input feature corresponds to an infinite number of labels with different values. If trained directly, the network will try to fit the "average" of these different labels, causing the prediction to converge to invalid intermediate values (for example, the true values are 1V and 5V, but the network may output a value of 3V), which will produce serious periodic jump errors.
[0006] Symmetry-induced sign ambiguity: This phenomenon occurs at the orthogonal bias points of DPMZM, where the transmission curve often exhibits even function characteristics, physically biasing the phase. and At this time, it is often difficult to distinguish between positive and negative values based solely on the light intensity scan curve (as their spectral shapes are the same). Conventional neural networks cannot distinguish between positive and negative signs. During training, the loss function will oscillate violently between positive and negative values, eventually causing the model to fail to converge or the predicted value to be close to zero, completely losing its physical meaning.
[0007] The inapplicability of the Mean Squared Error (MSE) loss function: Conventional deep learning uses the mean squared error (MSE) as the default loss function. MSE assumes the parameter space is Euclidean space, but in optical phase control, the parameter space is manifold or toroidal, for example, with a phase error of 0.1. and 1.9 They are physically close (difference of only 0.2). However, from the perspective of MSE, 1.9 (The numerical error is huge). This mismatch between mathematical measurement and physical facts seriously hinders the convergence accuracy of the model.
[0008] Therefore, there is an urgent need for a calibration method that can be fast and accurate. Summary of the Invention
[0009] This invention addresses the technical shortcomings of existing DPMZM calibration methods by providing a deep learning-based fast calibration system and method for DPMZM modulators. This system achieves a calibration speed improvement from seconds to milliseconds, while simultaneously resolving issues such as periodic multi-valued mapping, symmetric sign ambiguity, and loss function mismatch in deep learning calibration. This improves calibration accuracy and meets the engineering requirements for DPMZM production calibration and online closed-loop control.
[0010] This invention proposes a fast calibration method for DPMZM modulators based on deep learning, which includes the following steps: Step 101: Optical power data is acquired using a segmented decoupled scanning strategy.
[0011] Step 201: Construct a training dataset using the optical power transfer equation of DPMZM.
[0012] Step 301: Perform physical symmetry correction and tag preprocessing on the optical power data from step 101.
[0013] Step 401: A highly nonlinear mapping relationship between the output waveform of the optical modulator and the control parameters is established based on a one-dimensional convolutional neural network (Multi-scale 1D-CNN) with multi-scale feature aggregation. This network includes a hierarchical feature extraction layer, a global information aggregation layer, and a nonlinear regression prediction layer.
[0014] Step 501: Define a physical-sensing hybrid loss function. This loss function decouples the output parameters into "Euclidean space scalars" and "toroidal manifold state variables," and introduces a weighting strategy based on device physical sensitivity, as follows: ; in, and Preset loss weighting coefficients are used to balance the half-wave voltage loss term. With bias phase loss term Contribution to model training. Due to Using the mean square error form, while A cosine loss form based on device periodicity is adopted. The two types of loss terms differ in numerical scale and gradient change characteristics; therefore, different weighting coefficients are set to reconcile them. In some implementations, The value range can be 1.0 to 3.0. The value range can be 5.0 to 15.0. Through experimental comparison of multiple candidate parameter combinations, the optimal value is selected. =2.0, =10.0 to balance the prediction accuracy of half-wave voltage parameters and bias parameters, and to obtain better overall performance. For half-wave voltage loss, This is the offset phase cosine loss.
[0015] ; ; Step 601: Train the model and use an adaptive optimization algorithm to ensure that the model converges to the global optimum.
[0016] Step 701: Perform inverse mapping and restoration of physical parameters on the optimal solution output by the model.
[0017] Step 801: Export the trained model as a general open neural network exchange grid and embed it into the host computer.
[0018] This invention proposes a fast and accurate calibration system for DPMZM modulators based on deep learning. The system mainly consists of five parts: an optical link module, a photoelectric conversion and pre-processing circuit, a lower-level acquisition and data transmission module, an upper-level intelligent processing and human-computer interaction module, and a bias voltage driving and feedback module.
[0019] The optical link module consists of three parts: a laser, a DPMZM modulator, and a fiber coupler. The laser generates a continuous optical carrier signal, which is input to the DPMZM modulator. The DPMZM modulator, acting as the controlled object, receives the optical carrier from the laser and the electrical signal from the driver amplifier circuit, performing electro-optic modulation. The fiber coupler connects to the output of the DPMZM and sets the splitting ratio to 90:10. 90% of the main optical path signal is output to the subsequent communication link, while 10% of the monitoring optical path signal is extracted for feedback control.
[0020] The photoelectric conversion and preamplifier circuit consists of two parts: a photodetector and a preamplifier circuit. The photodetector is coupled to the monitoring port of the fiber optic coupler, which converts 10% of the optical power signal into a weak current signal. This weak current signal is then converted into a voltage signal and pre-amplified by the preamplifier circuit to match the input range of the STM32's internal analog-to-digital converter (ADC).
[0021] The host computer data acquisition and data transmission module is based on an STM32 microcontroller and is responsible for the digitization and communication transmission of the underlying data. The ADC data acquisition unit is connected to the output of the preamplifier circuit. Under the control of the scanning timing, it synchronously converts the monitored analog voltage to digital, obtains the digital sequence reflecting the current state of the DPMZM, and encapsulates the data link through the communication management unit, packs it into a data frame of composite communication protocol, and transmits the data to the host computer in real time through the USB physical interface. At the same time, it is also responsible for receiving control commands from the host computer.
[0022] The host computer intelligent processing and human-computer interaction module runs on a personal computer (PC) and is responsible for data analysis and decision-making. The data preprocessing unit receives raw data uploaded from the slave computer. It parses, normalizes, and reconstructs tensors to build a feature matrix that meets the input requirements of a neural network. The constructed feature matrix is then fed into a trained one-dimensional convolutional neural network model for feature extraction and forward inference, directly predicting the optimal voltage parameters required for the DPMZM modulator to operate at its best. The current optimal voltage value for the DPMZM modulator is simultaneously displayed on the host computer interface, and the predicted voltage value output by the inference engine is encapsulated into control commands and sent to the slave computer via the communication management unit.
[0023] The bias voltage drive and feedback module is responsible for executing control commands, converting digital signals into physical voltages that are applied to the modulator to form a closed loop. The digital-to-analog converter (DAC) data output unit receives the predicted voltage command from the host computer sent by the serial communication module, converts the digital control code into a corresponding analog DC voltage signal, and then amplifies and impedance-matches the analog voltage output by the DAC through a drive amplifier circuit. This directly drives each DC bias port of the DAC, thereby locking the modulator's operating point.
[0024] Compared with the prior art, the present invention has the following advantages: The scaling speed is improved by orders of magnitude: through a segmented decoupled scanning strategy, the number of scaling samples is reduced from the traditional quadratic O(N) to O(N) times. 2 The linearity is reduced to O(3N), and the calibration time is shortened from the second level of existing methods to the millisecond level, meeting the engineering requirements of DPMZM online real-time control; To resolve the physical conflict in deep learning calibration: label preprocessing with periodic folding and absolute value symmetry mapping reduces the impact of multi-valued mapping and symbol ambiguity of optical phase on model training; the designed physical-aware hybrid loss function adapts to the Euclidean space of half-wave voltage and the toroidal manifold space of bias voltage, improving the mismatch problem of traditional mean square error loss function. High calibration accuracy, meeting engineering requirements: The model exhibits excellent predictive performance on the test set, with high coefficients of determination R for all half-wave voltage parameters. 2 All reach above 0.95, R of the bias voltage of the I and Q arms 2 R at the orthogonal offset point of the main phase arm reaches 0.98 or higher. 2 The phase error is increased to over 0.915; after converting the predicted voltage into physical phase, the phase error is only 6.5°~10.5°, which is completely controlled within the ±15° DPMZM operating point locking engineering requirements. The model is robust and adaptable to real hardware environments: By normalizing the input data, injecting Gaussian noise, and employing a random deactivation strategy for neurons, the model can resist interference from real-world environments such as light source power drift, photodiode thermal noise, and ADC quantization noise. Even if the input light power fluctuates by more than 50%, the prediction results can still be kept stable. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.
[0026] Figure 1To use the method proposed in this invention to adjust the minimum transmission point bias voltage of the Q-arm sub-modulator ( Predict experimental results; Figure 2 To achieve the minimum transmission bias voltage of the I-arm sub-modulator using the method proposed in this invention ( Predict experimental results; Figure 3 To use the method proposed in this invention to adjust the quadrature bias point voltage of the main phase arm ( Predict experimental results; Figure 4 To use the method proposed in this invention to measure the half-wave voltage of the I-arm sub-modulator ( Predict experimental results; Figure 5 To use the method proposed in this invention to measure the half-wave voltage of the main phase arm ( Predict experimental results; Figure 6 To use the method proposed in this invention to measure the half-wave voltage of the Q-arm sub-modulator ( Predict experimental results; Figure 7 These are the experimental results without using the absolute value mapping strategy; Figure 8 This is the result of the error distribution; Figure 9 The results of the speed calculation experiment; Figure 10 This invention proposes a framework diagram for a fast calibration system for a DPMZM modulator based on deep learning. Figure 11 The present invention proposes a fast calibration method for DPMZM modulators based on deep learning. Flowchart of the method is provided. Detailed Implementation
[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.
[0028] This invention proposes a fast calibration method for DPMZM modulators based on deep learning, which includes the following steps: Step 101: Optical power data is acquired using a segmented decoupled scanning strategy.
[0029] Specifically, three independent one-dimensional voltage scanning strategies were designed for the I-arm sub-modulator, Q-arm sub-modulator, and main phase arm of the DPMZM. Scanning voltages were applied sequentially, and the output optical power curves were acquired, with a total sampling count of 3N. A simulation model was constructed based on the DPMZM optical power transfer equation, generating a dataset containing 50,000 samples. Each sample contained six physical parameter labels, namely the half-wave voltage of the I-arm sub-modulator (…). ), Q-arm sub-modulator half-wave voltage ( ), main phase arm half-wave voltage ( ), and the minimum transmission bias voltage of the I-arm sub-modulator ( ), Q-arm submodulator minimum transmission point bias voltage ( ), main phase arm quadrature bias point voltage ( ); The DPMZM has a nested structure (including two sub-MZMs and one main MZM), and its output optical power is simultaneously affected by six physical parameters (three half-wave voltages). and three bias voltages The nonlinear coupling effect of the optical power fluctuations can be significant. If only a single random scan is performed, the network may struggle to distinguish which arm caused the power change; for example, bias drift in the I-arm and bias drift in the Q-arm could lead to similar power fluctuations.
[0030] In order to train a deep neural network to accurately identify the complex parameter space of DPMZM, this invention adopts a segmented decoupled scanning strategy to obtain the voltage scanning sequence, with a total input dimension of (N,3,41).
[0031] The segmented decoupling scanning strategy includes the following three stages.
[0032] Phase 1: I-arm submodulator scanning During this phase, the control voltages of the Q-arm and the main phase arm are kept constant, and a linearly increasing scan voltage is applied only to the bias port of the I-arm; the scan range covers a range greater than [missing information]. (For example, 0V to 10V), with 41 sampling points. This curve mainly reflects the extinction ratio and periodicity characteristics of the I-arm, from which the network can extract key data. and The characteristic of this is that the influence of other parameters is temporarily suppressed at this time.
[0033] Second stage, Q-arm submodulator scanning During this stage, the control voltages of the I-arm and the main phase arm are kept constant, and a linearly increasing scan voltage is applied only to the bias port of the Q-arm. This section of the curve mainly reflects the physical characteristics of the Q-arm, from which the network can extract key information. and Its features enable decoupling from the I-arm parameters.
[0034] Phase 3: Main Phase Arm Scan During this stage, the voltages of the I-arm and Q-arm are kept constant (usually at the maximum transmission point), and a linear scanning voltage is applied only to the main phase arm. This section of the curve reflects the interference between the I-path optical signal and the Q-path optical signal. Using this data, combined with information from the previous two sections, the deep learning network can accurately lock the interference. and .
[0035] Since defining the bias voltage of a single modulator and monitoring the output optical power requires N scans, traditional global traversal or two-dimensional coupled scanning methods typically require... Multiple samplings are required to cover the search space. In contrast, this invention utilizes the decoupling characteristics of optical power response to transform the complex joint scanning into independent one-dimensional scans for the I, Q, and P ports. This requires only a simple superposition of three scanning processes, significantly reducing the total number of samplings to 3N. This method, from a quadratic (O( The dimensionality reduction from sampling to linear level (O(N)) significantly improves the scaling efficiency of the DPMZM bias voltage while ensuring scaling accuracy.
[0036] Step 201: Construct a training dataset using the optical power transfer equation of DPMZM.
[0037] Specifically, based on input optical power Output optical power The training dataset is constructed by mapping the data to various control voltages and device parameters; the mapping relationships are as follows: ; in, The insertion loss factor of the device; , , These are the three half-wave voltage parameters to be extracted; , , For the DC bias voltage to be extracted, these parameters determine the position of the modulator's transmission curve, corresponding to the minimum transmission point voltage of the I-arm ( ), minimum transmission point voltage of Q arm ( ) and the quadrature phase point voltage of the main arm ( ); , , This refers to the active control voltage applied during the scanning phase. The three terms in the formula represent: the I-path light intensity, the Q-path light intensity, and the interference phase of the I / Q paths, respectively.
[0038] To cover all possible device states, 50,000 samples were acquired through scanning. The six label parameters of each sample were randomly and uniformly distributed within the following ranges: Half-wave voltage ( ): Randomly selected between 3.0V and 8.0V.
[0039] Scan voltage ( , , ): Scanning is performed in steps of 0.5V from -10.0V to 10.0V; (41 scan points for one waveform).
[0040] bias voltage ( ):exist to It is randomly selected from among them.
[0041] Step 301: Perform physical symmetry correction and tag preprocessing on the optical power data from step 101.
[0042] Because the photoelectric transmission of DPMZM exhibits significant periodicity and symmetry, the original voltage labels suffer from a severe "multi-value mapping" problem (i.e., the physical state is the same but the voltage values are different). Directly using the original voltage labels as training targets will cause the neural network to fail to converge. Therefore, this invention introduces a feature mapping mechanism based on physical laws, specifically including the following three sub-steps: Step 3011: Normalize the optical power data. In this step, the input features are forcibly mapped to the dimensionless interval [0,1], so that the model only focuses on the shape features of the power curve rather than the absolute value. Even if the input optical power fluctuates by more than 50%, the shape of the normalized curve remains unchanged, ensuring the stability of the prediction results.
[0043] Step 3012: For the NULL point bias voltage tags of the I and Q paths, use their corresponding half-wave voltages. Map it to the principal value range centered at 0. Mapped bias voltage Specifically: ; Through the above transformation, the infinite voltage solution space is compressed into a single physical principal period, transforming the "ill-posed problem" into a "well-posed problem," which significantly reduces the training difficulty of the model.
[0044] Step 3013: Map the labels of the Quad points to the positive semi-axis [0, ... Ignoring the phase sign, the mapped Quad point voltage... Specifically: ; Step 401: A highly nonlinear mapping relationship between the output waveform of the optical modulator and the control parameters is established based on a one-dimensional convolutional neural network (Multi-scale 1D-CNN) with multi-scale feature aggregation. This network includes a hierarchical feature extraction layer, a global information aggregation layer, and a nonlinear regression prediction layer.
[0045] Specifically, the hierarchical feature extraction layer extracts features from low-level waveform contours to high-level semantic states step by step through stacked convolutional blocks (ConvBlocks).
[0046] Input layer: Tensor input with network receiving dimension (B,3,L); where B is the batch size and 3 represents three parallel monitoring channels, which monitor the optical power of I, Q and P respectively; it can directly fuse multiple optical signals and use the correlation between channels to solve the phase ambiguity problem that may exist in a single signal.
[0047] Convolutional Block 1: Kernel size is 5, output channels are 64; it uses activation functions (BatchNorm, ReLU) and a max-pooling layer. The large kernel provides a large receptive field, capable of covering half or a complete cycle segment of the optical power curve. This layer is primarily responsible for extracting the macroscopic contour features of the waveform, ignoring high-frequency noise. The max-pooling layer introduces local translation invariance, ensuring the extracted feature map remains stable even with slight shifts in the scanning start point.
[0048] Convolutional Block 2: The convolutional kernel size is 2, with 128 output channels; the small convolutional kernel focuses on capturing local abrupt changes and subtle gradient changes in the signal, especially the sharpness features at peak and trough inflection points. Doubling the number of channels is to extend low-dimensional contour information to a higher-dimensional feature space, increasing the diversity of feature combinations.
[0049] Convolutional Block 3: The kernel size is 3, and the output channels are 256. It fuses the features of the first two layers through deep convolution to form a high-dimensional abstract semantic vector that can uniquely represent the modulator's operating state. The output of this layer is no longer a simple waveform shape, but an implicit encoding containing physical information such as bias voltage position and half-wave voltage drift.
[0050] The global information aggregation layer employs one-dimensional adaptive average pooling to compress scan sequences of arbitrary length into a 256-dimensional fixed feature vector, integrating information from the entire scan cycle and suppressing the influence of measurement noise.
[0051] The nonlinear regression prediction layer consists of three fully connected (Linear) layers (256→128→6). A Dropout (0.2) function is introduced to randomly deactivate 20% of the neurons in the fully connected layers; this simulates sensor noise loss or signal jitter in a real-world hardware environment, forcing the network to avoid over-reliance on a few specific neuron paths, thereby improving the model's robustness in real-world conditions. The final output layer does not include an activation function and directly outputs the six predicted normalized parameters.
[0052] Step 501: Define a physical sensing hybrid loss function. This loss function decouples the output parameters into "Euclidean space scalars" and "toroidal manifold state variables," and introduces a weighting strategy based on device physical sensitivity. The total loss function is... Specifically as follows: ; in, and Preset loss weighting coefficients are used to balance the half-wave voltage loss term. With bias phase loss term Contribution to model training. Due to Using the mean square error form, while A cosine loss form based on device periodicity is adopted. The two types of loss terms differ in numerical scale and gradient change characteristics; therefore, different weighting coefficients are set to reconcile them. In some implementations, The value range can be 1.0 to 3.0. The value range can be 5.0 to 15.0. Through experimental comparison of multiple candidate parameter combinations, the optimal value is selected. =2.0, =10.0 to balance the prediction accuracy of half-wave voltage parameters and bias parameters, and to obtain better overall performance; its performance index is shown in Table 1. For half-wave voltage loss, This is the offset phase cosine loss.
[0053] Table 1 Performance Indicators for Different Weighting Coefficients
[0054] half-wave voltage loss In the middle, due to the half-wave voltage ( The mean square error (MSE) is an inherent physical property characterizing the electro-optic conversion efficiency of a modulator, belonging to a linear and continuous Euclidean space. Since its value directly corresponds to the physical order of magnitude and lacks periodicity, its absolute numerical accuracy can be precisely constrained using the mean square error (MSE). ; In the formula, , These are the predicted and actual values of the half-wave voltage of the k-th arm, respectively.
[0055] This ensures that the model can accurately pinpoint the "scale" of the device, serving as a benchmark for subsequent phase calculations.
[0056] Bias phase cosine loss In this approach, a strategy combining cosine metric and component sensitivity weighting is adopted. First, the voltage error is normalized to a phase error. Then, the periodicity of the cosine function is used to eliminate the gradient error caused by "phase entanglement," as detailed below: ; In the formula, This represents the cosine loss term of the bias voltage of the k-th arm. , These are the predicted and actual values of the bias voltage of the k-th arm, respectively. This represents the true value of the half-wave voltage of the k-th arm.
[0057] When the predicted value differs from the actual value by 2 At that time, the loss value is 0, which is consistent with physical facts.
[0058] Furthermore, since DPMZM includes three key bias control points: two sub-modulator biases and one main phase bias, and because the main phase bias directly determines the orthogonality of the I / Q signals, even a small deviation in it can lead to severe scatter plot distortion. To strengthen the model's focus on this key parameter, it is given a weight of 2.5 when calculating the average value, specifically expressed as follows: ; In the formula, This is the total loss term due to bias voltage. To expand the mean.
[0059] This strategy forces the network to prioritize optimizing the main phase offset, which has the greatest impact on the system's bit error rate.
[0060] Step 601: Train the model and use an adaptive optimization algorithm to ensure that the model converges to the global optimum.
[0061] Specifically, in order to overcome the “threshold difference” between simulation data and the real optical link environment (which has photodiode thermal noise, ADC quantization noise, etc.), a dynamic noise injection mechanism is introduced before the data is sent into the network.
[0062] In each training round, zero-mean Gaussian white noise is superimposed on the input optical power feature vector X to obtain the noisy input feature vector. : ; Set noise standard deviation =0.002 (corresponding to 0.2% of the normalized amplitude). This small perturbation forces the model to learn the topological relationships between features instead of relying on specific single numerical features, which significantly enhances the model's robustness to interference in low signal-to-noise ratio environments.
[0063] The training process employs a "rapid descent, stable convergence" strategy, with the following specific configuration: The optimizer chosen is the AdamW optimizer. Compared to the traditional Adam optimizer, the AdamW optimizer decouples weight decay from gradient updates, more effectively preventing overfitting, making it particularly suitable for scenarios involving a large number of physical parameter regressions in this task. Initial learning rate. Weight decay coefficient .
[0064] The learning rate scheduling is combined with an adaptive decay strategy during plateau periods. When the loss value no longer decreases within 10 consecutive epochs, the learning rate is automatically decayed to 0.5 times the current value. This "dynamic annealing" mechanism allows the model to quickly find the solution space in the early stages of training and then finely adjust it with small steps in the later stages, thereby accurately locking the operating point of DPMZM.
[0065] Step 701: Perform inverse mapping and restoration of physical parameters on the optimal solution output by the model.
[0066] Specifically, the model directly outputs It is a normalized and dimensionless value that must be restored to the real physical voltage value through inverse mapping in order to drive the DAC.
[0067] First, the numerical values are restored based on the scaling factor (LABEL_SCALE) recorded during dataset preprocessing. This represents the original voltage obtained by restoring the normalized predicted value output by the model to a preset scaling factor: ; Since the bias voltage has a periodic physical meaning, it needs to be processed according to the device characteristics: Half-wave voltage ( ): Directly adopt As a final result, it is an absolute physical quantity within Euclidean space.
[0068] bias voltage ( ): Utilizing the characteristics of the cosine loss function, the model's prediction essentially reflects the phase state. To prevent excessive voltage drift from exceeding the DAC range, [the following is used]... The periodicity folds the voltage back into the main cycle. Indicates to The final physical voltage obtained after further processing: ; Step 801: Export the trained model into a general open neural network exchange format and embed it into the host computer.
[0069] Specifically, the one-dimensional convolutional neural network model trained in the Python environment is exported to a general open neural network exchange format. In the Qt host computer, the model is embedded through an integrated model inference runtime environment. Upon startup of the host computer software, the model file with the .onnx extension is automatically loaded and the inference session is initialized. Then, the preprocessed feature tensors are input into the inference engine to perform forward propagation calculations. This process is entirely executed on the PC and does not consume resources on the lower-level machine. After the model processes the data, the output layer is directly mapped to six floating-point values: the half-wave voltage corresponding to each path and the optimal DC bias voltage required for the DPMZM to reach its optimal operating point.
[0070] like Figure 11 As shown, this invention proposes a fast and accurate calibration system for DPMZM modulators based on deep learning. The system mainly consists of five parts: an optical link module, a photoelectric conversion and pre-processing circuit, a lower-level acquisition and data transmission module, an upper-level intelligent processing and human-computer interaction module, and a bias voltage driving and feedback module.
[0071] The optical link module consists of three parts: a laser, a DPMZM modulator, and a fiber coupler. The laser generates a continuous optical carrier signal, which is input to the DPMZM modulator. The DPMZM modulator, acting as the controlled object, receives the optical carrier from the laser and the electrical signal from the driver amplifier circuit, performing electro-optic modulation. The fiber coupler connects to the output of the DPMZM and sets the splitting ratio to 90:10. 90% of the main optical path signal is output to the subsequent communication link, while 10% of the monitoring optical path signal is extracted for feedback control.
[0072] The photoelectric conversion and preamplifier circuit consists of two parts: a photodetector and a preamplifier circuit. The photodetector is coupled to the monitoring port of the fiber optic coupler, which converts 10% of the optical power signal into a weak current signal. The preamplifier circuit then converts the weak current signal into a voltage signal and amplifies it initially to match the input range of the STM32's internal ADC.
[0073] The host computer data acquisition and data transmission module is based on an STM32 microcontroller and is responsible for the digitization and communication transmission of the underlying data. The ADC data acquisition unit is connected to the output of the preamplifier circuit. Under the control of the scanning timing, it synchronously converts the monitored analog voltage to digital, obtains the digital sequence reflecting the current state of the DPMZM, and encapsulates the data link through the communication management unit, packs it into a data frame of composite communication protocol, and transmits the data to the host computer in real time through the serial bus (USB) physical interface. At the same time, it is also responsible for receiving control commands from the host computer.
[0074] The host computer intelligent processing and human-computer interaction module runs on the PC and is responsible for data analysis and decision-making. The data preprocessing unit receives raw data uploaded from the slave computer. It parses, normalizes, and reconstructs tensors to build a feature matrix that meets the input requirements of the neural network. The constructed feature matrix is then fed into a trained one-dimensional convolutional neural network model for feature extraction and forward inference, directly predicting the optimal voltage parameters required for the DPMZM modulator to operate at its best. The current optimal voltage value for the DPMZM modulator is simultaneously displayed on the host computer interface, and the predicted voltage value output by the inference engine is encapsulated into control commands and sent to the slave computer via the communication management unit.
[0075] The bias voltage drive and feedback module is responsible for executing control commands, converting digital signals into physical voltages that are applied to the modulator to form a closed loop. The DAC data output unit receives the predicted voltage command from the host computer sent by the serial communication module, converts the digital control code into a corresponding analog DC voltage signal, and then uses a drive amplifier circuit to amplify and impedance match the analog voltage output by the DAC, directly driving each DC bias port of the DPMZM, thereby locking the modulator's operating point.
[0076] like Figure 11 The diagram shows the workflow of this invention, which achieves fully automatic control from power-on startup to optimal operating point locking. The specific steps include system initialization, feature scanning and data acquisition, data preprocessing, 1D-CNN neural network model inference and execution, and model locking.
[0077] After the system powers on, the STM32 main control chip first performs hardware initialization, configuring the system clock, ADC sampling parameters, DAC output channels, and communication interfaces. Simultaneously, it loads a pre-trained and quantized 1D-CNN neural network model from Flash memory into the RAM memory pool, initializing the neural network engine. To obtain the physical characteristics of the current DPMZM modulator, the system sequentially performs three voltage scans of specific waveforms, constructing three sets of optical power response sequences corresponding to the I, Q, and P branches, denoted as SI, SQ, and SP, respectively. The lower-level machine then processes the three sets of raw optical power response sequences (… The data is encapsulated according to a predetermined communication protocol. After packaging, it is sent to the host computer via serial communication for data preprocessing. This preprocessing includes filtering and noise reduction (removing high-frequency electronic noise), normalization (mapping voltage values to the [0,1] interval), and tensor construction (stacking the three processed sequences into an input tensor of shape (1,3, sequence length). The constructed input tensor is then fed into the neural network inference engine, which performs convolution, pooling, and fully connected operations, outputting the required six scaling data points: the optimal bias voltage and half-wave voltage. The data is then packaged and sent to the lower-level computer via serial communication. The STM32 in the lower-level computer converts the model-predicted optimal bias voltage into an analog signal via a DAC output control unit. After amplification by the driver amplifier circuit, the signal is transmitted to the corresponding port of the DPMZM, where the modulator is locked at its optimal operating point.
[0078] The model constructed based on this invention achieves significantly better results than traditional methods on the test set: High-precision regression: like Figure 4 , Figure 5 , Figure 6 The scatter plot showing the predicted and actual values indicates all half-wave voltage parameters. coefficient of determination All reached 0.95; like Figure 1 , Figure 2 As shown, the most difficult to predict and Point bias voltage It reached 0.98.
[0079] Solving symmetry fuzziness: Through an absolute value mapping strategy, the orthogonal bias point |QUAD_P| The prediction accuracy has been significantly improved. For example... Figure 7 As shown, without using an absolute value mapping strategy, its The value is 0.770, so the model is unusable; after using the absolute value mapping strategy, such as... Figure 3 As shown, its The accuracy was improved to 0.9137, and the mean absolute error (MEA) was only 0.196, effectively eliminating the interference of multiple solutions caused by periodicity.
[0080] Extremely low phase error: After converting the predicted voltage to the physical phase, the error distribution is as follows: Figure 8 , Figure 9 As shown, the converted physical phase error is only 6.5-10.5 degrees. This is significant for DPMZM's operating point locking (which is typically required at...). Within 15 degrees (the temperature is within 15 degrees) it fully meets the engineering requirements.
[0081] Fast calculation speed: like Figure 10 As shown, compared to the several seconds required by iterative fitting algorithms, deep learning reduces the time to milliseconds, making it suitable for real-time monitoring and rapid calibration.
Claims
1. A fast calibration method for DPMZM modulators based on deep learning, characterized in that, The method includes the following steps: Step 101: Optical power data is acquired using a segmented decoupled scanning strategy; Step 201: Construct a training dataset using the optical power transfer equation of DPMZM; Step 301: Perform physical symmetry correction and tag preprocessing on the optical power data from step 101; Step 401: Establish a highly nonlinear mapping relationship between the output waveform of the optical modulator and the control parameters based on a one-dimensional convolutional neural network with multi-scale feature aggregation; Step 501: Determine the physical sensing hybrid loss function based on the weighted strategy of device physical sensitivity; Step 601: Train the model and use an adaptive optimization algorithm to ensure that the model converges to the global optimum. Step 701: Perform inverse mapping and restoration of physical parameters on the optimal solution output by the model.
2. The fast calibration method for DPMZM modulator based on deep learning according to claim 1, the method further includes step 801, exporting the trained model into the general ONNX format and embedding it into the host computer.
3. In the fast calibration method for DPMZM modulator based on deep learning according to claim 1, in step 101, three independent one-dimensional voltage scanning strategies are designed for the I-arm sub-modulator, Q-arm sub-modulator, and main phase arm of the DPMZM, and scanning voltages are applied sequentially and the output optical power curves are acquired.
4. The fast calibration method for DPMZM modulators based on deep learning according to claim 1, wherein the physical symmetry correction and label preprocessing includes the following three sub-steps: Step 3011: Normalize the optical power data; Step 3012: For the NULL point bias voltage tags of the I and Q paths, use their corresponding half-wave voltages. Map it to the principal value range centered at 0. ]; Step 3013: Map the labels of the Quad points to the positive semi-axis [0, ... The phase sign is ignored.
5. The fast calibration method for DPMZM modulators based on deep learning according to claim 1, wherein in step 501, the physical sensing mixing loss function is: ； In the formula, The value range can be 1.0 to 3.
0. The value range can be 5.0 to 15.
0. Through experimental comparison of multiple candidate parameter combinations, the optimal value is selected. =2.0, =10.0 to balance the prediction accuracy of half-wave voltage parameters and bias parameters, and to obtain better overall performance; For half-wave voltage loss, For bias phase cosine loss; where, Half-wave voltage drop for: ； In the formula, , These are the predicted and actual values of the half-wave voltage of the k-th arm, respectively. Bias phase cosine loss for: ； In the formula, To expand the mean, for: ； In the formula, , These are the predicted and actual values of the bias voltage of the k-th arm, respectively. This represents the true value of the half-wave voltage of the k-th arm.
6. The fast scaling method for DPMZM modulators based on deep learning according to claim 1, in step 701, numerical restoration is performed based on the scaling factor LABEL_SCALE recorded during dataset preprocessing: ； In the formula, For predicted values, This is the restored value.
7. A fast calibration system for a DPMZM modulator based on deep learning, characterized in that, The system for implementing the method according to any one of claims 1-6 comprises five parts: an optical link module, a photoelectric conversion and pre-processing circuit, a lower-level acquisition and data transmission module, an upper-level intelligent processing and human-computer interaction module, and a bias voltage drive and feedback module. Optical link module: It consists of a laser, a DPMZM, and a fiber coupler. The laser generates a continuous optical carrier, and the fiber coupler splits the output light of the DPMZM in a 90:10 ratio. Photoelectric conversion and preamplifier circuit: Composed of a photodetector and a preamplifier circuit, it converts the monitored optical signal into a current signal, and then converts and amplifies it into a voltage signal that matches the range of the ADC; Lower-level acquisition and data transmission module: Based on STM32 microcontroller, it completes ADC analog-to-digital conversion, data encapsulation and transmission, and upper-level control command reception; The host computer intelligent processing and human-computer interaction module completes data preprocessing, inference calculation of ONNX format models, and realizes parameter display and control command sending. Bias voltage drive and feedback module: Composed of DAC data output unit and drive amplifier circuit, it converts digital control code into analog voltage and amplifies it to drive each DC bias port of DPMZM.