A three-phase time sequence control-based optical fiber fusion three-electrode discharge control method

Abstract

The application provides a three-phase time sequence control-based optical fiber fusion three-electrode discharge control method, and belongs to the technical field of optical fiber fusion electrode control, which adopts three annularly distributed main electrodes to generate three-phase independent adjustable PWM signals to drive a three-phase high-voltage unit, and generates a high-voltage pulse arc with a set phase difference. The discharge process sequentially enters a pre-fusion stage, a main fusion stage and a re-fusion stage, and the discharge-pause cycle mode is adopted in the main fusion stage. In the main fusion stage, arc emission spectrum data are collected from different azimuth angles by multiple spectrum probes, a three-dimensional arc thermal field temperature distribution is reconstructed in real time by using a tomographic reconstruction algorithm, and the circumferential temperature unevenness is extracted. The duty ratio and / or the phase difference of the PWM signals of each phase are adjusted according to the unevenness feedback, so that the arc circumferential thermal field tends to be uniform. The application realizes accurate control of the arc thermal field, significantly improves the optical fiber fusion quality, and is especially suitable for high-precision fusion of special optical fibers such as hollow core optical fibers and polarization maintaining optical fibers.

Description

Technical Field

[0001] This invention relates to the field of electrode control technology, and more specifically, to a three-electrode discharge control method for fiber optic fusion splicing based on three-phase timing control. Background Technology

[0002] This invention belongs to the field of optical fiber fusion splicing technology, specifically relating to a three-electrode discharge control method and device for optical fiber fusion splicing based on three-phase timing control and artificial intelligence adaptive compensation. With the development of optical communication towards high capacity and low latency, special optical fibers such as hollow-core fibers, multi-core fibers, and polarization-maintaining fibers are gradually becoming mainstream. Traditional general-purpose fusion splicers can no longer meet these high-precision, customized operational requirements. In particular, the special microstructure of hollow-core fibers requires extremely high alignment accuracy, which ordinary fusion splicing equipment cannot handle, requiring dedicated fusion splicers. This invention, based on traditional three-phase timing control, introduces AI real-time image analysis and arc brightness control technology, significantly improving the adaptability and success rate of fusion splicing for special optical fibers. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a three-electrode discharge control method for fiber optic fusion splicing based on three-phase timing control, comprising the following steps:

[0004] Step 1: Three main electrodes are used, which are distributed in a ring around the circumference at a preset angle interval; three PWM drive signals with independently adjustable phases are generated to drive the three-phase high voltage generating unit respectively, so that the three electrodes generate high voltage pulse arcs with a set phase difference.

[0005] Step 2: Control the discharge process to sequentially enter the pre-melting stage, the main melting stage, and the remelting stage, wherein the main melting stage adopts at least one discharge-pause cycle mode;

[0006] Step 3: During the main melting stage, multiple spectral probes arranged around the electrode ring are used to collect the emission spectrum data of the electric arc from different azimuth angles;

[0007] Step 4: Based on the collected emission spectrum data, the three-dimensional arc thermal field temperature distribution centered on the fiber axis is reconstructed in real time using a tomographic reconstruction algorithm. The tomographic reconstruction algorithm combines an algebraic iterative reconstruction algorithm with maximum entropy regularization to discretize the arc region into a three-dimensional grid. The emission coefficient of each voxel is iteratively solved using the projection data collected by each spectral probe to reconstruct the three-dimensional temperature field.

[0008] Step 5: Extract the circumferential temperature non-uniformity from the reconstructed three-dimensional arc thermal field temperature distribution; the circumferential temperature non-uniformity is defined as the difference between the maximum and minimum values ​​of the temperature at all circumferential nodes at a specified radial distance and axial position.

[0009] Step 6: Based on the circumferential temperature non-uniformity, adjust the duty cycle and / or phase difference of each phase PWM drive signal to make the circumferential thermal field of the arc more uniform; when the circumferential temperature non-uniformity exceeds the set threshold, identify the electrode phase corresponding to the circumferential node with the lowest temperature, increase the PWM duty cycle of the electrode phase, and at the same time decrease the PWM duty cycle of the opposite electrode phase.

[0010] Furthermore, the plurality of spectral probes consists of three, which are respectively arranged at 0 degrees, 180 degrees and 270 degrees azimuth angles around the electrode ring; each spectral probe includes a collimating lens and a branched optical fiber, with the two ends of the branched optical fiber respectively connected to two photodetectors configured with narrowband filters of different wavelengths.

[0011] Furthermore, it also includes an adaptive pre-sealing step for the hollow fiber end face: before the pre-fusion stage, the three-dimensional morphology of the hollow fiber end face is measured using a white light interferometric ranging module to extract the end face concavity depth, hole area ratio, and surface roughness; if the sealing standard is not met, at least one sealing arc pulse is applied through an auxiliary electrode, and the end face center height change rate and reflected light intensity are monitored in real time during the pulse. When the height change rate approaches zero and the reflected light intensity increases to a preset threshold, the current pulse is terminated, and the sealing quality is re-evaluated after each pulse until the sealing is qualified.

[0012] Furthermore, it also includes an abnormal arc sound pattern detection and protection step: during the discharge process, the arc sound signal is collected through a microphone, and the fundamental frequency amplitude variance, characteristic frequency band energy and pulse spike rate are extracted; when at least one of the above features exceeds the threshold corresponding to the normal baseline, an arc instability warning is triggered and a protection action is executed, the protection action including increasing the DC bus voltage and / or adjusting the PWM frequency.

[0013] Furthermore, the normal baseline is established based on the isolated forest model, which uses the acoustic feature vector of the pre-fusion stage in the normal welding process as the training sample; when the real-time detected feature value deviates from the normal baseline by a preset multiple of the standard deviation, it is judged as abnormal.

[0014] Furthermore, it also includes AI visual monitoring and adaptive compensation steps: using a convolutional neural network model to extract the fiber outer diameter profile, arc brightness distribution, end face alignment deviation, and microstructure integrity features of special optical fibers in the splice image in real time; at the same time, ambient temperature, humidity, and air pressure data are collected, and the pre-fusion time, main fusion current duty cycle, and refusion time are dynamically compensated based on the environmental data.

[0015] Furthermore, it also includes a polarization-maintaining fiber alignment step: before the pre-fusion stage, a lateral imaging system is used to acquire images of the polarization-maintaining fiber end face, and a semantic segmentation network is used to identify the pixel-level position of the fiber stress zone and calculate the polarization axis direction; a rotary motor is driven to adjust the fiber angle so that the polarization axis angle difference between the two fibers is less than 0.5 degrees and the center offset of the stress zone is less than 0.1 micrometers.

[0016] Furthermore, the three-phase high-voltage generating unit includes three independent high-voltage transformers, with the low-voltage sides of the three high-voltage transformers connected in a star configuration and the high-voltage sides connected in a delta configuration; the PWM drive signal is generated by the timer of the digital controller, and the phase difference of the three PWM signals is set by timer synchronization and initial counter value.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects:

[0018] (1) By reconstructing the three-dimensional electric arc thermal field through tomography and controlling the feedback of circumferential temperature non-uniformity, the uniformity adjustment of the circumferential thermal field of the electric arc was realized, and the circumferential temperature difference at the fusion point could be controlled within ±5℃.

[0019] (2) The three-stage discharge mode of pre-melting, main melting and remelting is adopted to optimize the heat input process, effectively prevent the fiber end face from overheating and promote the flow and filling of molten glass;

[0020] (3) It integrates modules such as adaptive pre-sealing of hollow fiber end face, arc sound pattern abnormality detection protection, AI visual monitoring and environmental compensation, which significantly improves the fusion splicing adaptability and success rate of special optical fibers such as hollow fiber and polarization-maintaining fiber, and can still maintain a high fusion qualification rate in extreme environments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the control system hardware configuration and connection in Example 1;

[0022] Figure 2 This is a schematic diagram of the three main electrode structure;

[0023] Figure 3 The waveform diagram shows the current duty cycle of the three stages of the discharge process;

[0024] Figure 4 This is a schematic diagram of the adaptive sealing process at the end face of a hollow optical fiber. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0026] Example 1

[0027] I. Control system hardware configuration and connection method, such as Figure 1 As shown:

[0028] 1. Electrode arrangement and special optical fiber adaptation

[0029] The main electrode uses three tungsten-cerium alloy electrodes with a diameter of 2.0 mm, the tips of which are ground into a 30-degree conical angle, and the radius of curvature of the discharge end is approximately 0.1 mm. Figure 2 As shown, the electrodes are arranged in a ring with a 180-degree and two 90-degree intervals on a circle with a diameter of 8 mm. The electrode tip is 2.5 mm away from the fiber axis, ensuring that the arc envelops the fiber in a hollow shape. The electrode holder uses a finely adjustable ceramic clamp, combined with a high-precision grating ruler, to ensure that the three electrodes are coplanar and perpendicular to the fiber, with an alignment accuracy better than 0.1 micrometers.

[0030] To address the specific requirements of hollow fiber fusion splicing, a set of adjustable auxiliary electrodes with a diameter of 1.5 mm is added around the electrode ring. Driven by a high-temperature micro motor, these electrodes are used to preheat and seal the end face of the hollow fiber before splicing, preventing the absorption peak of the gas inside the hollow fiber from affecting it and preventing external gas from entering the fiber microstructure.

[0031] 2. High-voltage generating unit configuration

[0032] The three-phase high-voltage transformer consists of three independent high-voltage transformers. Each transformer has a UY16 ferrite core, a primary winding of 20 turns of multi-strand insulated wire, and a secondary winding of 2200 turns, with a turns ratio of 1:110. The primary DC resistance is 0.05 ohms, and the secondary inductance is approximately 85 millihenries. The low-voltage sides of the three transformers are connected in a star configuration: the corresponding terminals of the three primary windings are connected to the A, B, and C phase drive outputs, respectively, while the non-corresponding terminals are connected together to form a star point, which is ungrounded and floating. The high-voltage sides are connected in a delta configuration: the tail end of the first transformer is connected to the head end of the second transformer, the tail end of the second transformer is connected to the head end of the third transformer, and the tail end of the third transformer is connected to the head end of the first transformer, forming a closed triangle. The three connection points are connected to the three electrodes via high-voltage leads. The entire high-voltage unit is vacuum-sealed with epoxy resin, with an insulation strength greater than 30 kV. Each secondary circuit is connected in series with a 15 picofarad / 15 kV high-voltage ceramic capacitor, forming a series resonance with the leakage inductance. The resonant frequency is about 90 kHz, which can realize zero-voltage switching and reduce losses.

[0033] 3. Driver circuit construction

[0034] A three-phase full-bridge inverter topology is adopted, sharing a DC bus voltage Vbus equal to 300V, obtained from the mains power through rectification, filtering, and active PFC circuitry, with a PFC efficiency of no less than 97%. Each phase bridge arm consists of four N-channel MOSFETs (model IXFH26N50P), each driven by an IR2113 high-voltage half-bridge driver chip. The power supply terminals of the driver chip are equipped with bootstrap diodes and bootstrap capacitors, with a capacitance of 10µF and a withstand voltage of 25V. The logic inputs of the IR2113 receive complementary PWM signals from the controller and automatically insert a dead time of approximately 500 nanoseconds. The outputs of the three full-bridges are connected to the three terminals of the low-voltage side of the three-phase high-voltage transformer in a star connection. Each MOSFET is equipped with an aluminum heatsink and supplemented with forced air cooling. A 4.7µF / 630V CBB snubber capacitor is connected in parallel between the positive and negative buses of the bridge arms to suppress voltage spikes. The DC bus inputs of the three full-bridges are shared, but the power circuits of each phase bridge arm are routed independently to reduce mutual interference.

[0035] To meet the energy requirements of ultra-fine discharge in special optical fibers such as hollow-core fibers, the drive circuit has reserved pads for SiC MOSFET model C2M0045170P. When the operating frequency is increased to above 200 kHz, the circuit can switch to a SiC solution to reduce switching losses and increase power density.

[0036] II. Generation and Setting of Three-Phase Drive Waveforms

[0037] 1. Basic Waveform Generation

[0038] The main controller uses the STM32H743 high-performance microcontroller, based on the Cortex-M7 core, with a main frequency of 480 MHz. It utilizes its advanced timers TIM1, TIM8, and TIM5 to generate three complementary PWM waves. The timer clock frequency is set to 480 MHz, with a prescaler value of 1 and an auto-reload value of 5999, resulting in a PWM frequency of 80 kHz and a period of 12.5 microseconds. Each timer is configured in center-aligned mode, with the output compare polarity set to active high and complementary output enabled. To achieve a preset phase difference, the timers are synchronized: TIM1 acts as the master timer, outputting a trigger signal TRGO to start TIM8 and TIM5; simultaneously, the initial counter values ​​for TIM8 and TIM5 are set to 4000 and 2000 respectively, with an auto-reload value of 5999, causing TIM8 to lead TIM1 by 180 degrees and TIM5 to lag TIM1 by 90 degrees. This generates three pairs of basic PWM waves with a preset phase difference.

[0039] It should be noted that the three-phase timing control used in this method is not standard three-phase AC, but rather refers to three independent and controllable single-phase high-voltage pulses, the phase difference of which can be arbitrarily set. In this embodiment, the relative phases of the three-phase pulses A, B, and C are set to 0°, 180°, and 90°, respectively, thereby generating a rotating asymmetrical arc between the ring electrodes to achieve circumferential thermal field regulation.

[0040] 2. Dead Zone Control

[0041] In the timer configuration, hardware dead-time insertion is directly enabled. The dead time is set to 96 timer clock cycles via the DTG bit of the TIMx_BDTR register, which is 96 divided by 480 MHz equals 200 nanoseconds. This is then combined with the approximately 500 nanosecond dead time internal to the IR2113, resulting in a total dead time of approximately 700 nanoseconds. This ensures that the upper and lower transistors on the same bridge arm will not shoot-through. Three complementary PWM waves are output from the MCU's six channels, two per phase, and after level conversion, are sent to the corresponding IR2113 chips.

[0042] 3. Phase Allocation

[0043] CH1 and CH1N are from the TIM1 output, corresponding to the A-phase full-bridge; CH2 and CH2N are from the TIM8 output, corresponding to the B-phase full-bridge; and CH3 and CH3N are from the TIM5 output, corresponding to the C-phase full-bridge. When the A-phase PWM is in positive phase, the B-phase lags by 180 degrees, and the C-phase leads by 90 degrees, forming a three-phase rotating voltage vector that drives the three high-voltage transformers to operate in the ABC phase sequence. A 10-ohm current-limiting resistor is connected in series between each IR2113 output and the MOSFET gate.

[0044] III. Establishment of programmed control logic and AI adaptive compensation for the discharge process

[0045] 1. Parametric settings and intelligent parameter library

[0046] A fusion parameter library is established in the MCU's internal Flash memory and stored in the form of a structure. In addition to the three levels of current and time parameters for pre-fusion, main fusion, and re-fusion, the following intelligent fields are added: special fiber type identifiers including SMF, MMF, PMF, HCF, MCF, etc.; AI model weight vector; arc brightness-temperature mapping coefficient; and ambient temperature and air pressure compensation coefficients.

[0047] For example, the parameter group established for G.652.D optical fiber is as follows: Prefusion current level I pre A value of 12 corresponds to an effective value of approximately 14 mA arc current, and the pre-melting time T pre For 30 milliseconds; Main melting current level I main 20 corresponds to approximately 18 mA, and the total discharge time T main The duration is 110 milliseconds, using a 2-cycle mode, with each discharge lasting 80 milliseconds and an interval of 40 milliseconds; the remelting current level is I.anneal 10 equals approximately 12 milliamperes, and the discharge time T anneal The current level is 300 milliseconds. Current level adjustment is achieved by changing the effective operating pulse width of the full-bridge inverter, outputting in pulse-group mode within each 12.5 microsecond cycle. Intermittent discharge is achieved through the control software enabling or disabling of timer outputs.

[0048] 2. Multi-stage discharge timing programming

[0049] The welding process is divided into three stages: pre-welding, main welding, and re-welding, implemented using a finite state machine. After initialization, it enters the IDLE state, and upon receiving a start command, it transitions to the PRE_FUSE state. In this state, the MCU enables three-phase PWM output and presses I... pre The duty cycle of the level is, for example, 44% to output an 80kHz square wave group, which lasts for 30 milliseconds before shutting off all PWM outputs and entering an interval state. After waiting for 5 milliseconds, it transitions to the MAIN_FUSE state. The main melting phase executes the loop logic: first, using I... main The PWM output is at a duty cycle of 58% for 80 milliseconds, then stops for 40 milliseconds; this constitutes one loop. The loop counter is checked to see if it has reached two iterations; if not, execution continues. After completion, a 10-millisecond interval is entered before transitioning to the ANNEAL state. The recirculation phase uses I... anneal The PWM output is continuously running for 300 milliseconds at a duty cycle of 36%, after which the output is completely turned off, completing the fusion process. All timing is managed by a high-precision Timer interrupt with an accuracy of 100 microseconds.

[0050] 3. Cyclic and Intermittent Control

[0051] The discharge-pause cycle mode during the main melting stage is implemented using an internal software counter of the MCU. During discharge, the three-phase PWM outputs according to normal logic; during pause, the timer output enable bit is cleared, forcing all power transistors to turn off and extinguishing the arc. When re-enabled, the phase is restored, forming a continuous but intermittent heat input, preventing overheating of the fiber end face and allowing the molten glass time to flow and fill the micropores.

[0052] IV. AI-powered intelligent control and real-time process adjustment based on multi-physics field fusion perception

[0053] Based on conventional AI vision and spectral diagnostics, three new modules have been added: three-dimensional arc thermal field tomography reconstruction, white light interference control of hollow fiber end-face sealing quality, and arc acoustic signature anomaly detection and protection.

[0054] 1. Basic configuration of intelligent arc and visual monitoring system

[0055] This system is equipped with an integrated high-resolution CMOS image sensor and a multispectral arc diagnostic unit, including a silicon-based photodiode array. The wavelength response range is 200 nm to 1100 nm. The system allows for lateral observation of the fusion point and arc state via a lens group. The AI ​​image processing algorithm is based on the lightweight convolutional neural network MobileNetV3-Small, with a model size of only 5.2 megabytes and an inference time of less than 10 milliseconds. It extracts the following feature parameters in real time: fiber outer diameter profile, including diameter change rate and fused zone length; arc brightness distribution and thermal field uniformity; fiber end-face alignment deviation, i.e., micron-level offset of the X / Y axes; and special fiber microstructure integrity for monitoring hole collapse in hollow-core fibers.

[0056] Specifically, the MobileNetV3-Small model is used to extract image features during the fusion splicing process in real time, including fiber outer diameter profile, arc brightness distribution, end-face alignment deviation, and the integrity of the microstructure of special optical fibers. The training dataset used in this system consists of two parts. The first part is a standard dataset, containing 5000 fusion splicing process images collected in a controlled laboratory environment, covering four types of fiber: single-mode fiber G.652.D, multimode fiber, polarization-maintaining fiber, and hollow-core fiber. The second part is an extended dataset, containing 2000 images collected from actual field operating environments, including fusion splicing images under complex conditions such as high altitude, humidity, and vibration. All images are uniformly cropped to 224×224 pixels and undergo data augmentation processing, including random rotation, translation, scaling, and brightness adjustment.

[0057] The model was initialized using weights pre-trained on the ImageNet dataset, employing a transfer learning strategy. The optimizer was Adam, with an initial learning rate of 0.001, decreasing to 0.1 times the original rate every 5 epochs. Mean squared error loss was used for regression tasks (e.g., outer diameter, offset), and cross-entropy loss was used for classification tasks (e.g., microstructure integrity assessment). The training batch size was 32, and training lasted for 50 epochs. On the test set, the model achieved 98.5% accuracy in classifying fiber types, and the prediction error for end-face offset was less than 0.2 micrometers.

[0058] 2. Three-dimensional arc thermal field tomography reconstruction and circumferential isotropic temperature control

[0059] To achieve direct three-dimensional measurement of the rotating electric arc thermal field, this embodiment installs three sets of fiber optic spectral probes at azimuth angles of 180, 90, and 90 degrees around the electrode ring. Each probe set includes a collimating lens and a branched fiber, independently acquiring the arc spatial emission spectrum from the corresponding viewpoint, focusing on acquiring the 696.5 nm argon atom characteristic line or the 777.2 nm oxygen atom characteristic line. Using inverse Abelian transform and iterative tomography, the three-dimensional temperature field distribution T centered on the fiber axis is inverted in real time from the three sets of orthogonal projection data, where r is the radial coordinate, θ is the circumferential angle, z is the axial coordinate, and t is time. The spatial resolution can reach 50 μm radially, 5 degrees circumferentially, and 100 μm axially. The system calculates two key indicators: the circumferential temperature non-uniformity ΔT is equal to the difference between the maximum and minimum temperature values; and the axial temperature gradient dT / dz.

[0060] ΔTθ is used as the core controlled variable and fed back to the digital control system. When ΔT exceeds a set threshold, such as 30 degrees Celsius, the AI ​​module automatically adjusts the duty cycle combination and phase offset of the three-phase PWM. The specific control strategy is as follows: if the temperature in a certain circumferential region is too low, the duty cycle of the corresponding phase electrode is increased while the duty cycle of the opposite electrode is symmetrically decreased to maintain a relatively constant total heat input; if there is an overall uneven axial gradient, the common duty cycle of the three phases is finely adjusted in conjunction with the raising and lowering of the auxiliary electrodes. This control cycle is 5 milliseconds, ensuring that the uniformity of the arc circumferential thermal field is better than ±5 degrees Celsius.

[0061] It should be noted that in a shared bus voltage scheme, adjusting ΔT (e.g., increasing the duty cycle of a phase) is the preferred method. If the duty cycle adjustment reaches saturation (e.g., close to 100% or below the minimum sustaining voltage), and the thermal field is still uneven, then global or independent bus voltage regulation needs to be initiated as a backup. Alternatively, in a flyback boost module scheme, it can be explicitly stated that the main control module can independently control the drive voltage of each module and equip each phase with an independent feedback regulation loop to ensure the independence of each phase's output, thereby supporting the subsequent circumferential temperature equalization control logic.

[0062] In this embodiment, since only three sets of spectral probes are deployed (azimuth angles of 0°, 180°, and 270° respectively), it falls under the typical problem of Limited Angle CT (Limited Angle Computed Tomography). Directly applying the Abel Inversion method is theoretically only applicable to strictly axisymmetric distributions, and single-view data cannot invert circumferential inhomogeneities. To address this issue, the tomographic reconstruction algorithm in this system employs a combination of Algebraic Reconstruction Technique (ART) and maximum entropy regularization, as specifically implemented below:

[0063] 1) Discretization Model Construction

[0064] The arc region to be reconstructed is discretized into a three-dimensional mesh:

[0065] Radial (r): 0~300μm, step size 10μm, 31 nodes in total;

[0066] Circumferential (θ): 0~360°, step size 7.5°, 48 nodes in total;

[0067] Axial (z): Centered on the weld point, ±1.5mm, step size 50μm, a total of 61 nodes.

[0068] Each fiber optic spectral probe corresponds to one projected ray. The projection value p of each ray as it passes through the grid is... j Represented as:

[0069] ;

[0070] Where, x i w is the emission coefficient of the i-th voxel, which is positively correlated with temperature; ji Let be the path length of the j-th ray passing through the i-th voxel, calculated using the Siddon ray-driven algorithm; N is the total number of discrete voxels.

[0071] (2) Algebraic Iterative Reconstruction (ART) Formula

[0072] Simultaneous Algebraic Iterative Reconstruction (SIRT) strategy is employed to enhance robustness to noise:

[0073]

[0074] in, , Let be the emission coefficient of the i-th voxel (grid cell) during the k-th and k+1-th iterations. Let be the emission coefficient of the m-th voxel at the k-th iteration; k is the iteration number. The relaxation factor is initially set to 0.1 and decreases every 50 iterations; the projected data p j Data was collected simultaneously by three sets of probes (20kHz sampling rate). This represents the path length of the j-th ray through the m-th voxel. Projection data p j Data is acquired synchronously by three sets of probes (20kHz sampling rate). The acquisition by the three sets of spectral probes is synchronously initiated by the same hardware trigger signal, which is generated by the timer overflow event of the MCU at a frequency of 20kHz. Each channel of data is timestamped, and subsequent processing is performed according to time alignment.

[0075] To accelerate iterative convergence and avoid local minima, this system employs the inverse Abelian transform under the assumption of axisymmetry to generate the initial temperature field T. (0)Using only the projection data p(ρ) acquired by a spectral probe located at a 0° azimuth angle (assuming the arc is uniformly distributed circumferentially), the radial temperature distribution T0(r) is calculated via inverse Abelian transformation. Then, T0(r) is copied circumferentially (0°–360°) to all grid voxels, i.e., T (0) (r,θ,z)=T0(r). Although this initial field cannot reflect circumferential inhomogeneity, it provides a reasonable radial distribution prior, which enables subsequent ART iterations to converge quickly to a physically consistent solution under finite angular projection.

[0076] (3) Regularization and constraints

[0077] To prevent artifacts and divergence in finite-angle CT reconstruction, Tikhonov regularization and physical constraints are introduced:

[0078] Regularization parameters: The Tikhonov regularization coefficient α is set to 0.01 and determined in offline simulation using the L-curve method. The regularization matrix employs a second-order difference operator to penalize the Laplace operator modulus of the temperature field.

[0079] Physical constraints:

[0080] Temperature T≥3000 K (Minimum excitation temperature of electric arc);

[0081] Temperature T≤12000 KT≤12000 K (upper limit of ionization of optical fiber materials);

[0082] The axial temperature distribution is continuous, and the second derivative is bounded.

[0083] (4) Iteration termination criterion

[0084] The following two termination conditions are set; iteration stops when either one is met:

[0085] Residual change rate:

[0086]

[0087] Where p is the projection measurement vector, which is the emission intensity data (20kHz sampling rate) synchronously acquired by three sets of fiber optic spectral probes. Each element p j The measured value corresponding to the j-th ray; A is the system matrix (projection operator), describing the contribution weight of each voxel to each projected ray, and its element A ji =w ji The path length of the j-th ray through the i-th voxel; and Let be the reconstructed vectors for the k-th and k+1-th iterations; preferably, the maximum number of iterations reaches 200.

[0088] (5) Spatial resolution verification

[0089] Verification was performed under typical arc conditions (80kHz, 18mA) by injecting a known Gaussian temperature field:

[0090] Radial resolution: better than 50 μm (based on modulation transfer function MTF=0.1 standard);

[0091] Circumferential resolution: better than 5° (directly supported by 48 circumferential grids);

[0092] Axial resolution: better than 100μm (corresponding to a discretization step size of 50μm, which is satisfied by Nyquist sampling).

[0093] The computation time for reconstructing a complete three-dimensional temperature field is related to the number of iterations. In this embodiment, the maximum number of iterations is set to 25, with each iteration taking approximately 0.2ms (based on the STM32H743 @480MHz floating-point performance). The total time to complete one full three-dimensional temperature field reconstruction is approximately 5ms. This reconstruction cycle matches the system's feedback control cycle (5ms), ensuring that the latest three-dimensional temperature field data is obtained within each control cycle, thereby achieving real-time feedback control of the arc thermal field.

[0094] 3. Real-time monitoring and adaptive sealing of hollow fiber end-face sealing quality using white light interferometry

[0095] For the HC-1550-02 hollow-core optical fiber, this embodiment integrates a low-coherence white-light interferometric ranging module in the fiber's axis-to-axis optical path, with a center wavelength of 1310 nm and a bandwidth of 50 nm. This module calculates the microstructure of the fiber end-face surface in real time by measuring the interference pattern of the reflection spectrum, achieving an axial resolution of 10 nm. Figure 4 As shown, the specific sealing process logic is as follows:

[0096] The first step is initial measurement. Before the pre-sealed arc discharge, the original indentation depth d0 of the hollow fiber end face and the presence of open holes are measured, and the initial value of the end face roughness Ra0 is recorded.

[0097] The second step is dynamic sealing. The system calls the special optical fiber parameter library and automatically switches to a low-energy, long preheating mode. A low-energy arc pulse is applied, with a current level of I. pre The setting is 8, and the pulse width is 0.5 milliseconds. During the discharge process, the relative height change Δd of the end face center point is measured every 0.2 milliseconds. Simultaneously, the change in the intensity of the reflected light at the end face center is monitored.

[0098] The third step is endpoint determination. When Δd changes from a rapid increase to a stable plateau, and the intensity of the reflected light at the center of the end face increases to a preset threshold, indicating that a smooth and dense glass sealing layer has been formed, the current discharge pulse is immediately terminated. If the sealing standard is not met after the first pulse, the system automatically increases the current level of the next pulse to 10 or extends the pulse width to 0.8 milliseconds until the condition is met. The total number of pulses in the entire adaptive sealing process does not exceed 5 to avoid excessive thermal effects.

[0099] Step 4: Quality Verification. After sealing, the end face morphology is scanned again. If residual open holes or surface roughness Ra greater than 20 nanometers are detected, a local repair pulse is automatically executed. The pulse width is 0.3 milliseconds, the current level is 6, and it only acts on the electrode phase corresponding to the hole area. After sealing is completed, the system records the sealing parameters of the optical fiber to the parameter library for reuse in subsequent optical fibers of the same batch.

[0100] 4. Abnormal discharge pattern recognition and active protection based on arc acoustic signature characteristics

[0101] In this embodiment, a miniature electret microphone with a frequency response range of 10 Hz to 50 kHz is installed on the inner wall of the welding chamber. The signal is pre-amplified and then sent to the high-speed ADC of the MCU with a sampling rate of 200 kHz. A short-time Fourier transform is performed on the acoustic signal to extract the following features:

[0102] First, there is the fundamental frequency stability, which refers to the amplitude jitter variance of the three-phase PWM frequency of 80 kHz and its harmonic components. Under normal operation, this variance is small; however, it increases significantly when the electrodes are contaminated or when there are air pressure fluctuations.

[0103] Second, characteristic frequency band energy, which calculates the total energy in the 1 kHz to 10 kHz frequency band. This frequency band is related to plasma sheath oscillation and reflects the confinement state of the electric arc.

[0104] Third is the pulse spike rate, which counts the number of transient strong pulses exceeding a preset amplitude threshold per millisecond, corresponding to micro-arc or local breakdown events.

[0105] A lightweight isolated forest model is trained, using acoustic signature characteristics during normal welding as a benchmark. When the harmonic amplitude variance exceeds three times the normal standard deviation or the pulse spike rate is greater than 5 times per millisecond, the system predicts that arc extinction or instability may occur within the next 10 milliseconds. At this point, active protection is immediately implemented: the current main welding procedure is interrupted, switching to re-welding mode and temporarily increasing the bus voltage by 10% for 20 milliseconds to stabilize the arc; if the abnormality is not eliminated, the entire welding process is paused and an alarm is issued to prompt the operator to check the electrodes or environment. This acoustic signature monitoring module performs decision-level fusion with visual and spectral information, triggering protection only when there are no obvious abnormalities in visual and spectral data but abnormal acoustic signatures, achieving redundancy among multiple physical quantities.

[0106] Specifically, the isolated forest model is used for anomaly detection of arc acoustic signature features, and the establishment of a normal baseline is crucial. The normal baseline data acquisition process is as follows: Under a standard laboratory environment (temperature 25℃±2℃, relative humidity 40%±5%, atmospheric pressure 1013 hPa±20 hPa), the same fusion splicer was used to perform normal fusion splicing operations on 200 standard single-mode optical fibers G.652.D. In the last 10 milliseconds of the pre-fusion stage of each splice, microphone acoustic signals were acquired at a sampling rate of 200 kHz. Short-time Fourier transform was performed on each 1-millisecond frame of acoustic signal to extract three features: the 80 kHz fundamental frequency amplitude sliding variance, the total energy of the 1 kHz to 10 kHz frequency band, and the pulse spike rate.

[0107] A total of 2000 frames of normal voiceprint feature vectors were collected (10 frames contributed by each fiber). These 2000 feature vectors were used to train an isolated forest model, with 100 trees, a maximum tree depth of 8, and 256 sample subsamples. After training, the model outputs an anomaly score for each sample. The normal threshold was set to the 95th percentile of all training sample anomaly scores. Based on this, the 3x standard deviation rule used in the scheme is an engineered approximation of this threshold. Specifically, the mean of the variance of the 80 kHz fundamental frequency amplitude in the normal baseline data is μ, and the standard deviation is σ. The system sets the threshold for triggering an early warning to... Experiments have shown that this threshold corresponds to a false alarm rate of approximately 2.3% for the 95th percentile threshold of the isolated forest model, which meets practical requirements.

[0108] After each successful fusion, the system will supplement the baseline database with the acoustic signature feature vector of the current fusion pre-fusion stage, and perform an incremental update on the isolation forest model every 50 fusions to adapt to the slow changes caused by factors such as electrode aging.

[0109] 5. AI-driven comprehensive adaptive compensation algorithm

[0110] Building upon three-dimensional thermal field, sealing, and acoustic fingerprint protection, this system further integrates traditional PID control with an AI decision-making module based on real-time loss prediction and dynamic temperature, humidity, and altitude compensation. The specific process is as follows:

[0111] Step 1: Multi-dimensional feature fusion. At the beginning of the main melting stage, the AI ​​model simultaneously processes CMOS images, multispectral data, three-dimensional thermal field reconstruction data, and acoustic signature features to calculate the actual diameter change rate of the welded area, the spatial temperature gradient of the arc plasma, the circumferential temperature non-uniformity, and the acoustic signature stability index.

[0112] Step two, deviation diagnosis and collaborative compensation. If the detected expansion of the fiber outer diameter exceeds 5% of the nominal value of 125 micrometers, the cause is determined through multispectral and three-dimensional thermal field analysis: if the arc center temperature is too high, the AI ​​module automatically reduces the three-phase common duty cycle; if arc skew causes uneven heating, the three-phase phase offset is adjusted first; if the acoustic signature shows frequent arc jitter, the bus voltage is increased and the PWM frequency is fine-tuned. The three compensation methods can be executed simultaneously or with priority.

[0113] Step 3, Environmental Adaptation. The AI ​​module continuously collects data from internal temperature, humidity, and air pressure sensors at a sampling rate of 1 Hz, dynamically correcting discharge parameters. For example, in thin air environments above 3000 meters altitude, it automatically extends the pre-melting time by 15% and increases the main melting current by 8%, while adjusting the acoustic fingerprint anomaly threshold to reduce false alarm rates, ensuring a high pass rate for welding in high-altitude or underwater environments.

[0114] 6. Supplement to Special Fiber Optic Dedicated Fusion Splicing Modes

[0115] In addition to the adaptive sealing of the hollow-core fiber mentioned above, this embodiment also implements a bridging mode and a polarization-maintaining fiber alignment mode. When the hollow-core fiber is used for pigtail splicing with a conventional single-mode fiber, the system automatically calls the graded-index multimode fiber (GIMF) bridging mode. The AI ​​vision system identifies the position of the bridging fiber and precisely controls the discharge energy distribution to adapt to different thermal fusion characteristics. For the polarization-maintaining fiber, the system uses a higher-precision lateral imaging system with a resolution of 0.05 micrometers per pixel. The AI ​​algorithm identifies the polarization axis position of the fiber stress zone in real time, and automatic alignment is achieved by fine-tuning the motor before splicing, with a core offset of less than 0.1 micrometers.

[0116] 7. Real-time weld quality assessment and remelting repair

[0117] After fusion splicing is completed, the AI ​​system comprehensively calculates and estimates the splice loss based on image features, including splice smoothness, number of air bubbles, core diameter offset, historical 3D thermal field data, and acoustic signature records. If the estimated loss exceeds a preset threshold (e.g., 0.03 dB), the system automatically triggers a refusion mode. Without removing the fiber, it recalibrates the discharge parameters and performs a 120-millisecond refusion to repair the splice. The estimation error is less than 0.005 dB. After all operations are completed, the system packages and stores key parameters from the fusion process, such as temperature curves, current waveforms, 3D thermal field reconstruction data, historical arc brightness, acoustic signature sequences, and the final loss estimate, into the device log for subsequent big data analysis.

[0118] Example 2

[0119] After the operator places the optical fiber into the fusion splicer and selects the fiber type, the system executes the following steps in sequence:

[0120] The first step is to collect and preload environmental parameters.

[0121] The system first invokes the environmental sensing module. This module contains three digital sensors: a DS18B20 temperature sensor with a measurement range of -55°C to +125°C, an accuracy of ±0.5°C, and a resolution of 0.0625°C; an SHT30 humidity sensor with a measurement range of 0 to 100% relative humidity, an accuracy of ±3% relative humidity; and a BMP388 bar pressure sensor with a measurement range of 300 hPa to 1250 hPa, an accuracy of ±50 Pa, and a resolution of 0.18 Pa. All three sensors are connected to the I2C bus, and the system continuously reads the data 10 times at a sampling rate of 1 Hz, taking the arithmetic mean as the current ambient temperature, relative humidity, and atmospheric pressure values.

[0122] After completing the environmental measurements, the system reads the corresponding reference fusion splicing parameter structure from the parameter library of the MCU's internal Flash memory, based on the fiber type selected by the operator. This structure contains the following fields: pre-fusion current duty cycle D. pre Dimensionless, ranging from 10% to 60%; pre-melting time T pre Unit: milliseconds; Main melting current duty cycle D main The duration of a single discharge in the main melting process, T. on The unit is milliseconds; the interval T between two consecutive discharges of the main melt. off Unit: milliseconds; Number of main melting cycles N cycle Dimensionless; Remelting current duty cycle D anneal Remelting time T anneal The unit is milliseconds. For G.652.D single-mode fiber, the reference parameter is D. pre Equals 44%, T pre Equal to 30 milliseconds, D main Equals 58%, T on Equal to 80 milliseconds, T off Equals 40 milliseconds, N cycle Equals 2, D anneal Equals 36%, T anneal This equals 300 milliseconds. For HC-1550-02 type hollow fiber, the reference parameter is D. pre Equals 30%, T pre Equal to 50 milliseconds, D main Equals 48%, T on Equal to 120 milliseconds, T off Equals 60 milliseconds, N cycle Equals 1, D anneal Equals 30%, T anneal It equals 400 milliseconds.

[0123] The system performs linear compensation on the baseline parameters based on the current environmental parameters. Temperature compensation rule: When the ambient temperature is below 20 degrees Celsius, the pre-melting time T is reduced for every 1 degree Celsius decrease. pre Increase by 2%, main melting current duty cycle Dmain Increase by 0.5%. Humidity compensation rule: When the relative humidity is higher than 60%, for every 5% increase, the main melting current duty cycle D is increased. main Increase by one percent, remelting time T anneal Increase by 2%. Atmospheric pressure compensation rule: When the atmospheric pressure is lower than 1013 hPa standard atmosphere, for every 10 hPa decrease, the main melting current duty cycle D... main Increase by 0.8%, pre-melting time T pre Increase the threshold by 1.5%, and simultaneously adjust the pulse spike rate threshold for voiceprint anomaly detection from 5 times per millisecond to 4 times per millisecond to reduce the false alarm rate. All compensated parameters are stored in the working parameter set in RAM for subsequent steps. The compensation result is limited to a safe range, i.e., D. main No more than 70%, T pre The time limit is no more than 80 milliseconds to prevent overcompensation from damaging the fiber optic cable.

[0124] The second step, for hollow optical fibers, is to perform initial white light interferometry measurements and adaptive pre-sealing until the sealing quality meets the standards.

[0125] When the fiber type is identified as HC-1550-02 hollow fiber, the system automatically executes this step. The white light interferometric ranging module consists of a superluminescent diode (SLED) light source with a center wavelength of 1310 nm and a bandwidth of 50 nm, a 2×2 fiber coupler, a reference arm reflector, a measuring arm collimator mounted on a precision displacement stage, and a spectral analysis module. During measurement, the measuring arm collimator, driven by a micro stepper motor, performs a grating scan of the fiber end face along the X and Y directions. The scanning range is a circular area with a diameter of 200 μm, with a step interval of 2 μm. At each scanning point, the spectral analysis module acquires the interference spectrum in the range of 1520 nm to 1580 nm at a resolution of 0.1 nm, and calculates the relative height of that point using a fast Fourier transform. Finally, a three-dimensional topographic image of the end face is reconstructed, with an axial resolution of 10 nm and a lateral resolution of 2 μm.

[0126] The system extracts three key parameters from the topography data: the initial indentation depth d0 in the central region of the end face, defined as the difference between the average height of all points within a 20-micrometer radius of the geometric center of the end face and the average height of the annular region at the outer edge of the fiber; and the area ratio A of the voids. hole , is defined as the total area of ​​connected regions with a depth exceeding 500 nanometers in the topography image divided by the total scan area; surface roughness Ra is defined as the arithmetic mean of the absolute values ​​of the differences between the height of each sampling point and the average height.

[0127] If both d0 and A are satisfied simultaneously hole If the value is zero and Ra is less than 20 nanometers, the end face is considered to be properly sealed, and the pre-sealing step is skipped. Otherwise, the adaptive sealing cycle is entered.

[0128] The hardware actuator for the adaptive sealing cycle is the aforementioned liftable auxiliary electrode. The system first lowers the auxiliary electrode to a position 2.0 mm from the fiber axis, coplanar with the main electrode ring. Sealed discharge parameter initialization: Assume the current pulse number k equals 1, and the maximum allowable number of pulses K... max The value is 5. The execution process of each pulse is as follows:

[0129] The system is based on current level Iseal k Apply an electric arc pulse. Iseal k The conversion relationship with duty cycle is as follows: Iseal1 corresponds to a duty cycle of 3% multiplied by Iseal plus 5%. For the first pulse, Iseal1 equals 8, corresponding to a duty cycle of 29%; for the second pulse, Iseal2 equals 10, corresponding to a duty cycle of 35%; for the third pulse and subsequent pulses, Iseal... k A value of 12 corresponds to a duty cycle of 41%. The pulse width is initially set to 0.5 milliseconds, but the system dynamically terminates based on real-time feedback. During the discharge process, the white light interferometer module continuously measures the height h of the center point of the end face at a frequency of once every 0.2 milliseconds. center and the reflected light intensity I in the central region ref The reflected light intensity is normalized to the initial value I. ref0 The system calculates the height change rate dh / dt in real time, using the linear regression slope estimate from the three most recent sampling points.

[0130] White light interferometry has two working modes:

[0131] Real-time single-point monitoring mode for dynamic feedback during the sealing process:

[0132] The measurement beam is fixedly focused at the center point of the end face (a region with a diameter of 10 μm) and no scanning is performed.

[0133] A sampling rate of 5kHz (i.e., an altitude value is output every 0.2 milliseconds) is used to monitor the altitude change rate dh / dt in real time.

[0134] Offline morphology reconstruction mode for pre- and post-sealing quality assessment:

[0135] The entire 200μm diameter circular area is scanned in 2μm increments, with approximately 7850 scan points. A single full scan takes about 2 seconds.

[0136] It is performed only once each before the start of sealing, after the end of sealing, and during the welding quality evaluation stage.

[0137] When the following two conditions are met simultaneously, the system immediately turns off the discharge output of this pulse: Condition 1, the absolute value of dh / dt is less than 0.5 nanometers per millisecond, indicating that the height tends to be stable; Condition 2, I ref divided by I ref0 is greater than 2.5, indicating that a smooth specular reflection has been formed on the glass surface. If the above conditions are not met when the pulse duration reaches the preset maximum width of 1.0 millisecond, this pulse is forcibly terminated.

[0138] After each pulse ends, the system waits for a 10-millisecond cooling interval, then rescans the end face topography to update d0, A hole and Ra. If d0 is less than 100 nanometers, A hole is equal to zero percent and Ra is less than 20 nanometers, the sealing is determined to be qualified and the loop is exited. If the conditions are not met and k is less than K max , then k is increased by 1, the next pulse is executed, and the pulse width is increased by 0.1 millisecond compared with the previous pulse. If k reaches K max and the sealing standard is still not met, the system issues an audible and visual alarm to prompt the operator to check the cleanliness of the optical fiber end face or replace the optical fiber, and aborts the welding process.

[0139] After the sealing is qualified, the system writes the number of pulses used in this sealing, the current level and actual duration of each pulse into the temporary log of the optical fiber batch, for subsequent quick invocation of the same sealing parameters for optical fibers of the same batch.

[0140] The third step, optical fiber alignment and polarization-maintaining axis alignment if required, then start the pre-fusion stage

[0141] The optical fiber alignment process is divided into two stages: coarse alignment and fine alignment. In the coarse alignment stage, the system drives the X-axis and Y-axis stepper motors to make the position deviation of the cores of the two sections of optical fibers in the CMOS image less than 5 micrometers. In the fine alignment stage, the system switches to the high-resolution imaging mode, the image sensor outputs 2048×2048 pixels, and each pixel corresponds to an actual size of 0.1 micrometer. The outer contour of the optical fiber is extracted through an edge detection algorithm, and the center of the fiber core is located through the grayscale centroid method. The alignment target is that the deviation in the X direction is less than 0.1 micrometer, the deviation in the Y direction is less than 0.1 micrometer, and the end face gap is adjusted to 30 micrometers ± 2 micrometers through the Z-axis motor. The optical fiber angle is adjusted by the rotary motor so that the angle deviation of the optical fibers on both sides is less than 0.5 degrees.

[0142] If the fiber type is polarization-maintaining fiber, such as panda-type PMF, the system automatically activates the polarization-maintaining alignment mode after fine alignment. The lateral imaging system uses a high-magnification microscope and a telecentric lens, with a resolution of 0.05 micrometers per pixel. The AI ​​model, based on the U-Net semantic segmentation network, performs pixel-level segmentation of the stress areas in the image, identifying the center coordinates of two circular stress areas. The angle between the centers of the two stress areas is calculated as the polarization axis direction. The system drives a rotary motor to fine-tune the fiber angle, ensuring that the difference in polarization axis angle between the two sides is less than 0.5 degrees, while maintaining the center offset of the stress area is less than 0.1 micrometers. The entire alignment process is completed within 500 milliseconds.

[0143] Specifically, the U-Net network is dedicated to polarization axis identification of polarization-maintaining fibers, and the dataset source and annotation method should be clearly stated here. The training dataset consists of 2000 end-face microscopic images of panda-type polarization-maintaining fibers. All images were acquired in situ within the fusion splicer using a high-resolution CMOS sensor, with a resolution of 0.05 micrometers per pixel. Image annotation was independently completed by three professional engineers with over five years of experience in fiber fusion splicing. The annotators used the Labelme tool to mark the two circular stress zones in the images as target regions pixel by pixel. After annotation, the Dice coefficient was used to evaluate annotation consistency, with the consistency index requiring an average Dice coefficient among the three annotators to be greater than 0.95. For samples with a consistency below 0.9, a fourth expert arbitrated the data before re-annotation. The dataset was randomly divided into a training set of 1600 images (80%), a validation set of 200 images (10%), and a test set of 200 images (10%).

[0144] The network was trained end-to-end, with the encoder using a ResNet-34 pre-trained on ImageNet as the backbone to initialize weights. The optimizer was Adam with an initial learning rate of 0.001. To accommodate small object segmentation, a combination of Dice loss and cross-entropy loss was used, with weighting coefficients of 0.7 and 0.3, respectively. The training batch size was 8, and the training lasted for 100 epochs. Performance metrics on the test set were: mean intersection-over-union ratio (IoU) of 96.2%, pixel accuracy of 98.8%, and a standard deviation of 0.03 micrometers for stress zone center localization error.

[0145] During inference, the system uses the centroid coordinates of the two segmented stress zones as the center of the stress zone. The polarization axis direction is determined by calculating the angle between the two center points. The system drives a rotary motor to ensure that the angle difference between the polarization axes of the two optical fibers is less than 0.5 degrees and the offset of the stress zone center is less than 0.1 micrometers.

[0146] After alignment, the system initiates the pre-melting stage. The pre-melting parameters use the environmentally compensated D parameters from the first step. pre and T pre For example, in a standard environment, Dpre Forty-four percent, T pre The duration is 30 milliseconds. The system enables three-phase PWM output, outputting an 80 kHz square wave group according to the set duty cycle. During the pre-melting process, the multispectral arc diagnostic unit monitors the radiation intensity of the arc at the 696.5 nm characteristic wavelength with a sampling rate of 5 kHz. If the real-time value of the radiation intensity deviates from the average value from the 5th to the 10th millisecond after the start of pre-melting by more than ±5%, the system fine-tunes the DC bus voltage, adjusting it by ±1 volt per step, with a maximum adjustment range of ±10 volts, to stabilize the arc. The pre-melting duration reaches T. pre Afterward, the system shuts down all PWM outputs and waits for a 5-millisecond interval to allow the molten end face to initially take shape.

[0147] The fourth step involves entering the main melting stage, where three-dimensional thermal field tomography reconstruction control, acoustic anomaly monitoring, and AI visual assistance are activated to adjust discharge parameters in real time.

[0148] Before the main melting phase begins, the system loads the D parameter set from the working parameter set. main T on T off and N cycle D under standard conditions main It is 58%, T on For 80 milliseconds, T off For 40 milliseconds, N cycle The value equals 2. The system initializes a cycle counter. cnt It is 0.

[0149] Entering the main melting cycle. First, the electron discharge phase is executed: the system outputs a three-phase PWM with a duty cycle of D. main Duration T on During this period, the 3D thermal field tomography reconstruction module, the acoustic signature monitoring module, and the AI ​​vision assistance module operate simultaneously.

[0150] The hardware of the three-dimensional thermal field tomography reconstruction module consists of three sets of fiber optic spectral probes, located outside the electrode ring at azimuth angles of 0°, 180°, and 270°. Each probe includes a collimating lens and a branched fiber. The two ends of the branched fiber are connected to two photomultiplier tubes, the former equipped with a 696.5 nm narrowband filter and the latter with a 777.2 nm narrowband filter. The signals from the three sets of probes, totaling six channels, are simultaneously acquired at a sampling rate of 20 kHz. Based on the emission intensity projections from the three perspectives, the system uses the inverse Abelian transform algorithm combined with maximum entropy iterative reconstruction to output a three-dimensional temperature field centered on the fiber axis with a period of 1 millisecond. The space is discretized into 15 nodes in the radial direction (10 micrometers per node), 24 nodes in the circumferential direction (15 degrees per node), and 21 nodes in the axial direction (±1 mm range centered on the fusion splice) (100 micrometers per node). The system calculates two key indicators from the temperature field: the circumferential temperature non-uniformity ΔTθ is equal to the difference between the maximum and minimum temperatures of all circumferential nodes at a radial distance of 50 micrometers and an axial position of 0; the axial temperature gradient dT / dz is equal to the maximum absolute value of the ratio of the temperature difference between adjacent nodes in the axial direction to the distance between them.

[0151] The control logic is as follows: The system reads the current ΔT every 5 milliseconds. If ΔT is greater than or equal to 30 degrees Celsius, thermal field equalization control is initiated. First, the electrode phase corresponding to the circumferential node with the lowest temperature is identified and designated as phase P. min Then, the PWM duty cycle of this phase is increased by an adjustment amount Δ based on the current D_main, and the phase P is adjusted accordingly. min The PWM duty cycle of the opposite electrode is reduced by half, Δ. The initial value of Δ is 2%, with each adjustment step being 0.5%, and the maximum cumulative adjustment not exceeding 8%. After adjustment, wait 5 milliseconds to recalculate ΔT. If ΔT drops below 20 degrees Celsius, the current duty cycle combination is maintained; if ΔT continues to increase beyond 40 degrees Celsius, all duty cycles are reset to D. main And restart the adjustment. If the axial temperature gradient exceeds 1000 degrees Celsius per millimeter, the system will automatically adjust the T... on Extend by 10 percent to provide more time for heat diffusion.

[0152] The voiceprint anomaly monitoring module acquires microphone signals at a sampling rate of 200 kHz and calculates a set of features every 1 millisecond. These features include: the sliding variance of the 80 kHz fundamental frequency amplitude, calculated over 10 consecutive 1-millisecond frames; the total energy in the 1 kHz to 10 kHz frequency band, calculated using a Butterworth bandpass filter; and the pulse spike rate, defined as the number of events per millisecond where the signal amplitude exceeds five times the absolute deviation of the median. The system compares these three features with the normal baseline within each frame. The normal baseline is a stable value statistically obtained in the last 10 milliseconds of the pre-melting stage. If both the 80 kHz fundamental frequency amplitude variance and the pulse spike rate are greater than three times the normal baseline and greater than five times per millisecond, an arc instability warning is triggered. Instead of immediately interrupting the main melting process, rapid compensation is performed: the DC bus voltage is increased from 300 V to 315 V for 10 milliseconds, while the PWM frequency is adjusted down by 2 kHz from 80 kHz to 78 kHz. If the acoustic signature returns to normal after 10 milliseconds, the original voltage and frequency are restored; if the abnormality persists for more than 30 milliseconds, the system will actively terminate the current electron discharge stage and switch to the emergency remelting procedure.

[0153] The AI ​​vision-assisted module acquires splice images at 50 frames per second and calculates the fiber outer diameter profile in real time. Canny edge detection is used to extract the upper and lower edges of the fiber, and the outer diameter value is calculated after fitting a straight line. The outer diameter expansion rate is defined as the difference between the current outer diameter and the reference outer diameter of 125 micrometers divided by 125 micrometers. If the expansion rate exceeds 5%, the system combines three-dimensional thermal field data to determine the cause: if the expansion rate is high and the center temperature exceeds 2000 degrees Celsius, the D value is reduced. main Two percent; if the expansion rate is high but the temperature is normal and ΔT is large, the thermal field equalization logic will be invoked first. If the expansion rate exceeds eight percent, the system will immediately terminate the main melt discharge and issue an alarm.

[0154] After the electron discharge phase is completed, the system enters the interval sub-phase: all PWM outputs are turned off for a duration of T. off During the interval, the system maintains three-dimensional thermal field and acoustic signature monitoring but does not make adjustments, allowing the molten glass to flow naturally.

[0155] After one complete main melt cycle is completed, cycle cnt Add 1. If cycle cnt Less than Ncy cle If so, then continue with the next discharge-interval cycle. If cycle cnt Equal to Ncy cle If the main melting stage is interrupted, the process will exit and enter a 10-millisecond transition interval before transitioning to the remelting stage.

[0156] Fifth, after the main melting is completed, remelting is performed while continuously monitoring the morphology of the weld joint.

[0157] Lower power parameters are used during the remelting stage: Danneal It is 36 percent, T anneal The duty cycle is 300 milliseconds, which may be adjusted to 34% to 40% and 300 to 345 milliseconds after environmental compensation. The system outputs a three-phase PWM with a duty cycle of D. anneal Continuous output without interruption. During the remelting process, the three-dimensional thermal field control is downgraded to monitoring mode and no longer actively adjusts the duty cycle; the acoustic fingerprint monitoring switches to recording-only mode and does not trigger active protection.

[0158] The weld joint morphology monitoring continuously acquires images at a frequency of 10 frames per second, focusing on monitoring the smoothness of the molten zone. The smoothness index S is defined as the ratio of the standard deviation of the outer diameter measured every 10 micrometers within a range of ±200 micrometers centered on the weld joint to the average diameter. The system calculates S every 100 milliseconds. If S is greater than 0.05, it indicates unevenness in the weld joint profile, and the system automatically extends the remelting time by 20 milliseconds. S is re-evaluated after each extension, with a maximum of 5 extensions allowed, meaning the total remelting time does not exceed 500 milliseconds. If S drops below 0.03 within the extended time, remelting is terminated prematurely. After remelting, the system shuts down all PWM outputs and cools for 50 milliseconds to allow the weld joint to fully solidify. Figure 3 The diagram shows the current duty cycle waveforms for the three stages of the discharge process. The three stages and their internal cycles are arranged in chronological order, with the width of the color blocks representing the duration and the vertical axis indicating the duty cycle values.

[0159] Step 6: Quality assessment. If the quality is satisfactory, the process ends; otherwise, remelting or an alarm will be triggered.

[0160] After cooling is complete, the system initiates a quality assessment procedure. The assessment is based on four quantitative indicators, calculated jointly by the AI ​​vision system and the high-precision imaging module.

[0161] The first item is the expansion coefficient R of the weld joint diameter. exp The system measures the outer diameter every 5 micrometers within a 100-micrometer range on either side of the center of the weld joint. The largest outer diameter value is divided by 125 micrometers and then subtracted by 1 to obtain the expansion rate. Acceptance requirement R exp Less than 8%.

[0162] The second item is the number of bubbles, N. bubble The system employs high dynamic range imaging and adaptive threshold segmentation to identify grayscale anomalies within the fusion splice area. For each candidate region, its area and roundness are calculated. Regions with an area greater than 0.5 square micrometers and a roundness greater than 0.7 are counted as one bubble. The acceptable requirement is that within a 200-micrometer length along the fiber axis, the number of bubbles should not exceed two, and the maximum size of a single bubble should be less than 2 micrometers.

[0163] The third item is the core diameter offset Δcore. The system locates the center coordinates of the left and right fiber cores through image segmentation, calculates the absolute values ​​of the differences between the two in the X and Y directions, and takes the larger value as Δcore. The acceptable requirement is that Δcore is less than 0.5 micrometers.

[0164] The fourth item is the surface roughness Ra of the weld joint. s The system acquires topographic data of the weld area from the white light interferometry module and calculates the roughness. Acceptance requirement: Ra s Less than 50 nanometers.

[0165] The AI ​​comprehensive loss prediction model uses a linear weighted formula: the loss prediction equals 0.1 multiplied by R. exp Divide by 8% and add 0.3 multiplied by N bubble Divide by 2, add 0.4 multiplied by Δcore, divide by 0.5 micrometers, add 0.2 multiplied by Ra. s Divide by 50 nanometers. Each term is normalized by dividing by its acceptable threshold. Loss estimation is in decibels (dB). If the estimated loss is less than 0.03 dB, it is considered acceptable, the system records an acceptable log, and the splicing process ends. If the estimated loss is between 0.03 and 0.08 dB, it is considered repairable, and the system automatically triggers re-fusion mode. If the estimated loss is greater than 0.08 dB, it is considered unrepairable, the system issues an audible and visual alarm and displays the estimated loss value, prompting the operator to re-prepare the fiber endface.

[0166] The specific operation of the remelting mode is as follows: the system does not move the optical fiber or change its relative position; it directly reloads the main fusion parameters but adjusts the power reduction. The remelting current duty cycle D... remelt Set as the original D main Ninety percent of the remelting single discharge time T remelt The timer is set to 60 milliseconds, with only one loop executed and no interval. No three-dimensional thermal field adjustment is performed during remelting; a simple constant power output is used to avoid excessive parameter complexity. After remelting, the quality assessment of this step is repeated. If the second assessment is successful, a successful remelting log is recorded; if it still fails, an alarm is triggered and manual intervention is requested.

[0167] Before refusion, the system does not move the fiber but re-performs an alignment check, evaluating the offset only using electron microscopy images without motor adjustments. If the offset exceeds 0.5 μm, the system abandons refusion and immediately triggers an alarm. For hollow-core fibers, pre-sealing is not repeated before refusion because the original sealing layer has already been fused into the splice after the first fusion. After refusion, a complete quality assessment (including end-face morphology inspection) is performed again. If the refusion still fails, a maximum of one retry is allowed (two refusions in total); if the second refusion also fails, an alarm is triggered.

[0168] All fusion splicing process data is automatically stored to an SD card. The log file name format is YYYYMMDD_HHMMSS_fiber type.txt. The log content includes: ambient temperature, humidity, and air pressure; operating parameters after initial compensation; for hollow-core fibers, recording the sealing pulse sequence and the real-time height and reflected light intensity curves of each pulse; recording ΔTθ and adjustment actions every 5 milliseconds during the main fusion stage; timestamps of acoustic signature anomalies and their handling measures; the outer diameter expansion rate sequence for each frame; the values ​​of the four final quality assessment indicators and the loss estimate; whether refusion occurred and the refusion result. This log can be used for subsequent big data analysis and process optimization.

[0169] With the detailed parameters and logical descriptions of the above six steps, those skilled in the art can implement the high-performance fiber optic fusion splicing control method described in this embodiment without creative effort.

[0170] Through the above hardware configuration, control logic, and multi-physics field fusion sensing method, this embodiment achieves high-quality, high-success-rate fusion splicing of ordinary single-mode optical fiber and special hollow-core optical fiber, and can maintain a pass rate of better than 99.99% even in extreme environments such as high altitude and humidity.

[0171] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A three-electrode discharge control method for optical fiber fusion splicing based on three-phase timing control, characterized in that, Includes the following steps: Step 1: Use three main electrodes, which are arranged in a ring around the circumference at a preset angle. Three PWM drive signals with independently adjustable phases are generated to drive the three-phase high-voltage generating unit respectively, so that the three electrodes generate high-voltage pulse arcs with a set phase difference. Step 2: Control the discharge process to sequentially enter the pre-melting stage, the main melting stage, and the remelting stage, wherein the main melting stage adopts at least one discharge-pause cycle mode; Step 3: During the main melting stage, multiple spectral probes arranged around the electrode ring are used to collect the emission spectrum data of the electric arc from different azimuth angles; Step 4: Based on the collected emission spectrum data, the three-dimensional arc thermal field temperature distribution centered on the fiber axis is reconstructed in real time using a tomographic reconstruction algorithm. The tomographic reconstruction algorithm combines an algebraic iterative reconstruction algorithm with maximum entropy regularization to discretize the arc region into a three-dimensional grid. The emission coefficient of each voxel is iteratively solved using the projection data collected by each spectral probe to reconstruct the three-dimensional temperature field. Step 5: Extract the circumferential temperature non-uniformity from the reconstructed three-dimensional arc thermal field temperature distribution; the circumferential temperature non-uniformity is defined as the difference between the maximum and minimum values ​​of the temperature at all circumferential nodes at a specified radial distance and axial position. Step 6: Based on the circumferential temperature non-uniformity, adjust the duty cycle and / or phase difference of each phase PWM drive signal to make the circumferential thermal field of the arc more uniform; when the circumferential temperature non-uniformity exceeds the set threshold, identify the electrode phase corresponding to the circumferential node with the lowest temperature, increase the PWM duty cycle of the electrode phase, and at the same time decrease the PWM duty cycle of the opposite electrode phase.

2. The method according to claim 1, characterized in that, The plurality of spectral probes are three in number, arranged at 0 degrees, 180 degrees and 270 degrees azimuth angles around the electrode ring respectively; each spectral probe includes a collimating lens and a branched optical fiber, with the two ends of the branched optical fiber connected to two photodetectors configured with narrowband filters of different wavelengths respectively.

3. The method according to claim 1, characterized in that, It also includes an adaptive pre-sealing step for the hollow fiber end face: before the pre-fusion stage, the three-dimensional morphology of the hollow fiber end face is measured using a white light interferometric ranging module to extract the end face recess depth, hole area ratio and surface roughness; if the sealing standard is not met, at least one sealing arc pulse is applied through an auxiliary electrode, and the end face center height change rate and reflected light intensity are monitored in real time during the pulse. When the height change rate approaches zero and the reflected light intensity increases to a preset threshold, the current pulse is terminated, and the sealing quality is re-evaluated after each pulse until the sealing is qualified.

4. The method according to claim 1, characterized in that, It also includes an abnormal arc sound pattern detection and protection step: during the discharge process, the arc sound signal is collected by a microphone, and the fundamental frequency amplitude variance, characteristic frequency band energy and pulse spike rate are extracted; when at least one of the above features exceeds the threshold corresponding to the normal baseline, an arc instability warning is triggered and a protection action is executed, the protection action includes increasing the DC bus voltage and / or adjusting the PWM frequency.

5. The method according to claim 4, characterized in that, The normal baseline is established based on the isolated forest model, which uses the acoustic signature feature vector of the pre-fusion stage in the normal welding process as the training sample. When the real-time detected feature value deviates from the normal baseline by a preset multiple of the standard deviation, it is judged as abnormal.

6. The method according to claim 1, characterized in that, It also includes AI visual monitoring and adaptive compensation steps: using a convolutional neural network model to extract in real time the fiber outer diameter profile, arc brightness distribution, end face alignment deviation and special fiber microstructure integrity features from the fusion splice image; Simultaneously, ambient temperature, humidity, and air pressure data are collected, and the pre-melting time, main melting current duty cycle, and remelting time are dynamically compensated based on the environmental data.

7. The method according to claim 1, characterized in that, It also includes a polarization-maintaining fiber alignment step: before the pre-fusion stage, a lateral imaging system is used to acquire images of the polarization-maintaining fiber end face, and a semantic segmentation network is used to identify the pixel-level position of the fiber stress zone and calculate the polarization axis direction; a rotary motor is driven to adjust the fiber angle so that the polarization axis angle difference between the two fibers is less than 0.5 degrees and the center offset of the stress zone is less than 0.1 micrometers.

8. The method according to claim 1, characterized in that, The three-phase high-voltage generating unit includes three independent high-voltage transformers. The low-voltage sides of the three high-voltage transformers are connected in a star configuration, and the high-voltage sides are connected in a delta configuration. The PWM drive signal is generated by the timer of the digital controller, and the phase difference of the three PWM signals is set by timer synchronization and initial counter value.

Images