Multi-sensor information fusion method for sewage pipeline detection robot
By using a quantum-inspired neural network fusion model, the problems of low detection accuracy and poor environmental adaptability in the multi-sensor multi-modal information fusion of sewage pipeline inspection robots are solved, achieving rapid transfer and efficient computation, thereby improving detection accuracy and environmental adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-03-31
AI Technical Summary
Existing methods for fusing multi-sensor, multi-modal information in sewage pipeline inspection robots suffer from low detection accuracy, poor environmental adaptability, difficulty in model transfer, and low computational efficiency, making it difficult to meet real-time detection requirements.
A quantum-inspired neural network fusion model is adopted. Multimodal data is collected synchronously through a master clock to build a dataset and train and optimize the model. Feature extraction is performed by combining the characteristics of quantum superposition and entanglement. Transfer learning and model compression techniques are used to deploy the model on a sewage pipeline inspection robot for parallel computing.
It improves detection accuracy, enhances environmental adaptability, enables rapid model transfer and efficient computation, and can quickly identify pipeline defects in complex environments.
Smart Images
Figure CN121765618A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of sewage pipeline inspection technology, specifically a multi-sensor information fusion method for a sewage pipeline inspection robot. Background Technology
[0002] With urban development, sewage pipeline systems are becoming increasingly large and complex, and their safe and stable operation is crucial to the normal functioning of cities. Sewage pipeline inspection robots, as a highly efficient inspection tool, can penetrate deep into the pipelines to obtain critical information; however, they currently have many shortcomings in multi-sensor, multi-modal information fusion.
[0003] Traditional multi-sensor fusion methods mainly focus on simple fusion at the data layer, feature layer, or decision layer, making it difficult to fully explore the complex correlations and potential features between different modalities. For example, when processing visual images and ultrasonic data, it is impossible to effectively combine the texture information in the image with the distance information of the ultrasonic wave, resulting in low detection accuracy for pipeline defects.
[0004] Existing fusion algorithms have poor adaptability to complex environments. When there is strong noise, uneven lighting, or complex pipe structure in the sewage pipeline, the reliability and stability of the fusion results decrease significantly. Moreover, the detection data of sewage pipelines in different regions and of different types vary, and existing methods lack the ability to effectively handle these differences, making it difficult to achieve rapid transfer and application of detection models.
[0005] In addition, traditional methods also have bottlenecks in computational efficiency, making it difficult to meet the needs of real-time detection and unable to analyze multimodal data in a timely and accurate manner during the rapid movement of the robot. Summary of the Invention
[0006] To address the above problems, this invention provides a multi-sensor information fusion method for sewage pipeline inspection robots, which solves the problems of low detection accuracy, poor environmental adaptability, difficulty in model transfer, and low computational efficiency in existing multi-sensor multi-modal information fusion methods for sewage pipeline inspection robots.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A multi-sensor information fusion method for a sewage pipeline inspection robot includes the following steps:
[0009] S1 Sensor Synchronous Acquisition: Trigger signals are sent to multiple sensors of the sewage pipeline inspection robot via the master clock to control the multiple sensors to acquire multimodal data at the same time. The multimodal data includes pipeline inner wall images, wall thickness distance, and harmful gas concentration.
[0010] S2 Data Preparation: Collect multi-sensor multi-modal data of various types of sewage pipes, construct a dataset, and divide the dataset into a training set, a validation set, and a test set, with the ratio of the training set, validation set, and test set being 7:2:1;
[0011] S3 Model Training: A quantum-inspired neural network fusion model is trained on the training set. The quantum-inspired neural network fusion model uses the superposition and entanglement characteristics of qubits to extract multimodal data features.
[0012] S4 Model Optimization: The parameters of the quantum-inspired neural network fusion model were optimized using the Adam optimization algorithm;
[0013] S5 Transfer Learning: The quantum-inspired neural network fusion model is pre-trained and fine-tuned using transfer learning, wherein the pre-training is performed on a public sewage pipe inspection dataset and the fine-tuning uses a small amount of sample data from the target domain.
[0014] S6 Model Compression: The quantum-inspired neural network fusion model trained in step S5 is pruned and quantized. Pruning uses L1 regularization to remove connections and neurons that are less than or equal to a threshold. Quantization converts the floating-point parameters in the model into low-precision fixed-point representations.
[0015] S7 Deployment and Operation: A parallel computing framework is deployed on the hardware platform of the sewage pipeline inspection robot. The parallel computing acceleration of the quantum-inspired neural network fusion model is realized by using the CUDA programming model of the graphics processing unit, and the allocation of computing resources is dynamically adjusted according to the robot's motion state and the complexity of the inspection task.
[0016] Furthermore, the feature extraction of the quantum-inspired neural network fusion model includes mapping sample data to a quantum feature space to obtain quantum states. The mapping function is The sample data is quantum state .
[0017] Furthermore, the quantum state The Bloch is represented as ,in It is the Bloch representation of a positive quantum bit. It is the Bloch representation of a negative qubit, where C is the correlation parameter; when the quantum state is in a product state, the quantum characteristics simplify to .
[0018] Furthermore, the mapping function τ makes the quantum state The information loss function that preserves the maximum feature information of the sample data. Defined as:
[0019] Where N is the number of samples. It is the i-th original sample. It is the i-th new sample reconstructed from the quantum state.
[0020] Furthermore, the Adam optimization algorithm includes the following sub-steps:
[0021] Initialize model parameters: initial model weights First moment (momentum): Second moment (adaptive learning rate): Number of iterations: Hyperparameter: learning rate attenuation coefficient Numerical stability term ;
[0022] Parameters are updated iteratively, and the following operations are performed during each iteration:
[0023] The iteration counter increments, and the value of t increases by 1: ;
[0024] Gradient calculation: based on current model parameters Calculate the loss function using randomly selected sample batches. For model parameters gradient ,in ;
[0025] Update the first moment: according to Update the first-order moment estimate;
[0026] Update the second moment: according to Update the second-order moment estimate, where The square of the gradient;
[0027] First-order moment deviation correction: Calculate the corrected first-order moment ;
[0028] Second-order moment deviation correction: Calculate the corrected second-order moment ;
[0029] Parameter update: based on Update all weight parameters of the model.
[0030] Furthermore, in the transfer learning step, pre-training is performed on a public sewage pipeline detection dataset to learn general pipeline features and patterns; during fine-tuning, all weight parameters from the input layer to the Kth layer in the pre-trained model are fixed, and only the weight parameters from the K+1th layer to the output layer are updated; wherein, the value of K ranges from 70% to 90% of the total number of layers in the model.
[0031] Furthermore, the pruning uses L1 regularization, with the L1 regularization loss function being: ,in The regularization intensity coefficient is . Here, M represents the original loss function, and M is the number of model weight parameters. It is the j-th weight parameter.
[0032] Furthermore, the quantization converts the floating-point parameters in the model into 8-bit integer or 16-bit floating-point representations.
[0033] Furthermore, the parallel computing acceleration utilizes NVIDIA GPU's CUDA programming model to dynamically adjust GPU computing resource allocation.
[0034] Furthermore, the master clock includes a field-programmable gate array or an external synchronizer.
[0035] The beneficial effects of this invention are that by mapping data to a high-dimensional quantum feature space (Bloch representation), the model can more fully capture the potential and complex correlation patterns between different modalities. Simulation of quantum superposition states allows features to simultaneously contain multiple state information, greatly enriching the expressive capacity of the features, thereby enabling the identification of more hidden and complex pipe defects. The model is first pre-trained on a large public dataset, acquiring "general knowledge" about pipe defects. When applied to a new target environment, only a small amount of sample data from that environment needs to be used to fine-tune the last few layers of the model. This allows the model to quickly adapt to complex and changing environments such as different lighting conditions, pipe materials, and sediment types. Through L1 regularization pruning, a large number of unimportant connections and neurons are removed, significantly reducing the number of model parameters and computational load, ensuring that the optimized model can run at high speed and low latency on the robot's limited hardware resources. Attached Figure Description
[0036] Figure 1 A flowchart of a multi-sensor information fusion method for a sewage pipeline inspection robot;
[0037] Figure 2 This is a structural diagram of a sewage pipeline inspection robot that integrates multi-sensor information. Detailed Implementation
[0038] To enable those skilled in the art to better understand the technical solution, the present invention will be described in detail below with reference to embodiments. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.
[0039] Example 1
[0040] See attached document Figure 1 and attached Figure 2 A multi-sensor information fusion method for a sewage pipeline inspection robot includes the following steps:
[0041] S1 Sensor Synchronous Acquisition: Trigger signals are sent to multiple sensors of the sewage pipeline inspection robot via the master clock to control the multiple sensors to acquire multimodal data such as pipeline inner wall images, wall thickness distance, and harmful gas concentration at the same time;
[0042] S1 includes at least steps S110-S130:
[0043] S110: Perform master clock triggering and sensor control to generate synchronous multimodal data;
[0044] A master clock sends trigger signals to multiple sensors on the sewage pipeline inspection robot, controlling these sensors to simultaneously acquire multimodal data on the pipeline's inner wall image, wall thickness distance, and hazardous gas concentration. Specifically, a master clock synchronization link is first established at the robot's sensing layer, covering the image sensor, wall thickness distance sensor, and hazardous gas concentration sensor. On the image acquisition side, a high-resolution camera and illumination unit are configured to capture the visual features of the pipeline's inner wall; on the wall thickness measurement side, an ultrasonic or laser ranging sensor is configured to obtain the pipeline wall thickness distance sequence; and on the gas detection side, a multi-channel hazardous gas sensor is configured to acquire concentration time-series data. Upon receiving the master clock trigger signal, all sensors synchronously start data acquisition and record a unified timestamp.
[0045] The multimodal data enters a synchronization processing flow after acquisition. Specifically, timestamp alignment and frame synchronization processing are performed on the image data to ensure that each frame of the image is strictly consistent with the corresponding wall thickness distance and hazardous gas concentration data in time; sampling rate matching and interpolation processing are performed on the wall thickness distance data to align with the image frame rate; and instantaneous value capture and smoothing processing are performed on the hazardous gas concentration data to eliminate sampling jitter. After the above synchronization processing is completed, synchronized multimodal data is formed, which includes a time-aligned sequence of pipe inner wall images, a sequence of wall thickness distances, and a sequence of hazardous gas concentrations, with a unified time window index for subsequent steps.
[0046] S120. Extract multimodal features from the synchronous multimodal data, perform feature alignment and registration processing, and obtain a multimodal feature sequence;
[0047] The synchronous multimodal data is read and segmented according to fixed time windows. Multimodal features are extracted within each time window. Specifically, visual feature extraction is performed on the pipeline inner wall image sequence, including texture features, defect area contours, and color distribution, and these features are bound to time window indices. Statistical feature extraction is performed on the wall thickness distance sequence, including mean distance, fluctuation amplitude, and outlier markers, and the wall thickness change trend is recorded. Concentration feature extraction is performed on the harmful gas concentration sequence, including peak concentration, average concentration, and duration of exceeding the standard, and gas type identification is recorded. Preferably, cross-modal alignment is performed at the feature level: image texture features are spatially correlated with outlier markers of wall thickness distance, and gas concentration peaks are temporally correlated with suspicious areas in the image to form a unified multimodal feature description.
[0048] Once all the above features are generated within each time window, they are arranged in chronological order to form a multimodal feature sequence. This multimodal feature sequence includes image visual features, wall thickness statistical features, and gas concentration features. A time window index and a modal association index are established in the data structure to ensure accurate referencing by downstream steps.
[0049] S130. Perform data standardization and enhancement processing on the multimodal feature sequence to generate an enhanced multimodal dataset;
[0050] The multimodal feature sequence is read, and data standardization and augmentation operations are performed. Specifically, the image visual features are normalized, converting texture and contour features to a standard scale; the wall thickness statistical features are dimensionally unified, converting the mean distance and fluctuation amplitude into dimensionless indicators; and the gas concentration features are standardized by scaling based on a safety threshold. Simultaneously, data augmentation is performed: image data is augmented through rotation, scaling, and noise addition; wall thickness distance data is augmented through time-series distortion and jitter; and gas concentration data is augmented through concentration perturbation and temporal interpolation. After augmentation, an augmented multimodal dataset is formed, which includes the standardized and augmented multimodal feature sequences and retains the time window index and modality association index for subsequent model training.
[0051] S2 Data Preparation: Collect multi-sensor multi-modal data of various types of sewage pipes, construct a dataset, and divide the dataset into a training set, a validation set, and a test set, with the ratio of the training set, validation set, and test set being 7:2:1;
[0052] Step S2 includes at least steps S210-S230:
[0053] S210. Collect and organize multi-sensor, multi-modal data to construct the original dataset;
[0054] Multi-sensor, multi-modal data were collected from various types of sewage pipes, including images of the pipe inner walls, wall thickness distances, and hazardous gas concentrations. Specifically, data was collected from sewage pipes of different materials, diameters, and aging levels, covering common defect types such as cracks, corrosion, and blockages. After data collection, the data was processed and labeled: defect types and locations were labeled for image data; normal and abnormal ranges were labeled for wall thickness distance data; and safe and excessive states were labeled for hazardous gas concentration data. After processing, a raw dataset was constructed, and metadata on pipe type and environmental conditions was recorded.
[0055] S220. The original dataset is divided and preprocessed to obtain a training set, a validation set, and a test set.
[0056] The original dataset is read and randomly divided into training, validation, and test sets in a 7:2:1 ratio. Specifically, the distribution of various pipe types and defect patterns is ensured to be consistent across the training, validation, and test sets during the partitioning process. Preprocessing includes: resizing and grayscale normalizing the image data; denoising and normalizing the wall thickness distance data; and logarithmic transformation and normalization of the harmful gas concentration data. After preprocessing, the training, validation, and test sets are stored separately, along with data indexes and label information.
[0057] S230. Perform quality checks and consistency verifications on the partitioned dataset to generate a usable dataset;
[0058] Quality checks are performed on the training, validation, and test sets, including checks for data integrity, annotation consistency, and distribution balance. Specifically, the consistency of image annotations with wall thickness anomalies and gas exceedances is checked, and the temporal alignment of the multimodal data is verified. After passing consistency verification, a usable dataset is generated, which includes the training, validation, and test sets, along with a quality report and index mapping, for use in model training.
[0059] S3 Model Training: A quantum-inspired neural network fusion model is trained on the training set. This model uses the superposition and entanglement properties of qubits to extract multimodal data features. Quantum-inspired feature extraction maps samples to quantum features, which are generated by mapping samples to quantum features. A 2-qubit composite system was obtained. The purpose of this invention is to put the system in a direct product state and to keep the qubits in the composite system far apart from each other while satisfying the requirement of mutual independence as much as possible. Therefore, it is necessary to express and constrain the obtained quantum characteristics. The Bloch is represented as When a 2-qubit composite system is in a product state, the system's quantum state is uniquely determined by the positive and negative qubits. The quantum characteristics at this point can be simply represented as... ,in It is the Bloch representation of a positive quantum bit. is the Bloch representation of a negative qubit, and C is the correlation parameter. Initialize the mapping parameters to preserve the maximum feature information of the samples.
[0060] Quantum state obtained from mapping The maximum feature information of the sample should be preserved, meaning the original sample can be reconstructed from the quantum state. The mapping function τ makes the quantum state... To retain the maximum feature information of the sample data, the information loss between the new sample obtained from quantum state reconstruction and the original sample is calculated using the information loss function. Defined as:
[0061] Where N is the number of samples. It is the i-th original sample. It is the i-th new sample reconstructed from the quantum state.
[0062] S4 Model Optimization: The parameters of the quantum-inspired neural network fusion model were optimized using the Adam optimization algorithm;
[0063] Initialize model parameters: initial model weights First moment (momentum): Second moment (adaptive learning rate): Number of iterations: Hyperparameter: learning rate (Default is 0.001), attenuation coefficient Numerical stability term ;
[0064] Parameters are updated iteratively, and the following operations are performed during each iteration:
[0065] The iteration counter increments, and the value of t increases by 1: ;
[0066] Gradient calculation: based on current model parameters Calculate the loss function using randomly selected sample batches. For model parameters gradient ,in ;
[0067] Update the first moment: according to Update the first-order moment estimate;
[0068] Update the second moment: according to Update the second-order moment estimate, where The square of the gradient;
[0069] First-order moment deviation correction: Calculate the corrected first-order moment ;
[0070] Second-order moment deviation correction: Calculate the corrected second-order moment ;
[0071] Parameter update: based on Update all weight parameters of the model;
[0072] Check convergence conditions: If the gradient norm is less than the threshold or the number of iterations reaches the upper limit, stop optimization; otherwise, continue iteration. After optimization, generate the optimized model parameters and record the optimization trajectory.
[0073] S5 Transfer Learning: The quantum-inspired neural network fusion model is pre-trained and fine-tuned using transfer learning, wherein the pre-training is performed on a public sewage pipe inspection dataset and the fine-tuning uses a small amount of sample data from the target domain.
[0074] Step S5 includes at least steps S510-S530:
[0075] S510. Pre-train on a public sewage pipeline inspection dataset to learn general features;
[0076] A quantum-inspired neural network fusion model was pre-trained using a publicly available wastewater pipeline inspection dataset. Specifically, the publicly available dataset was loaded, and data preprocessing similar to that used for S200 was performed. The goal of the pre-training was to learn general patterns and features of pipeline defects, wall thickness variations, and gas concentrations. After pre-training, the pre-trained model parameters were saved.
[0077] S520. Collect a small amount of sample data in the target field and perform data adaptation processing;
[0078] Collect a small sample of data, including images, wall thickness, and gas concentration data, from the target area or specific type of sewage pipes. Standardize and augment the data to align it with the distribution of the pre-training data. This process creates the target domain dataset.
[0079] S530: Perform model fine-tuning, fix some layer parameters and update subsequent layers;
[0080] Load the pre-trained model and fix all weight parameters from the input layer to the Kth layer, where K is 70% to 90% of the total number of layers in the model. Update only the parameters from the (K+1)th layer to the output layer, and train using the target domain dataset. After fine-tuning, generate a model adapted to the target domain and record the fine-tuned parameters.
[0081] S6 model compression: Pruning and quantization are performed on the trained quantum-inspired neural network fusion model. Pruning uses L1 regularization to remove unimportant connections and neurons, and quantization converts the floating-point parameters in the model into low-precision fixed-point representations; specifically:
[0082] A pruning algorithm is used to prune the trained quantum-inspired neural network fusion model, removing unimportant connections and neurons to reduce the number of model parameters. L1 regularization is applied to constrain the model parameters, setting parameters below a certain threshold to 0 to achieve the pruning operation.
[0083] L1 regularization simplifies the model by penalizing the magnitude of model parameters, guiding the model to prioritize parameters with smaller absolute values, and even compressing some non-critical parameters to 0.
[0084] Original loss function: only measures the difference between the model's predicted value and the true value, such as mean squared error or cross-entropy;
[0085] Add L1 regularization loss function: ,in The regularization intensity coefficient is . Here, M represents the original loss function, and M is the number of model weight parameters. It is the j-th weight parameter. It degenerates into unregularization; The larger the value, the stronger the penalty, and the easier it is for the parameter to be compressed to 0.
[0086] The pruned model is quantized, converting the floating-point parameters in the model into low-precision fixed-point representations, such as 8-bit integers or 16-bit floating-point numbers, to further reduce computational complexity.
[0087] S7 Deployment and Operation: A parallel computing framework is deployed on the hardware platform of the sewage pipeline inspection robot. The parallel computing acceleration of the quantum-inspired neural network fusion model is realized by using the CUDA programming model of the graphics processing unit, and the allocation of computing resources is dynamically adjusted according to the robot's motion state and the complexity of the inspection task.
[0088] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the present invention. These examples are merely for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or variations without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of the present invention to other situations without modification, should all be considered within the scope of protection of the present invention.
Claims
1. A sewer pipeline inspection robot multi-sensor information fusion method, characterized in that, The method comprises the following steps: S1 sensor synchronous acquisition: sending a trigger signal to multiple sensors of a sewer pipeline detection robot through a master clock to control the multiple sensors to collect multi-modal data at the same time, the multi-modal data including pipeline inner wall images, wall thickness distances, and harmful gas concentrations; S2 data preparation: collecting multi-modal data of various types of sewer pipelines, constructing a data set, and dividing the data set into a training set, a validation set, and a test set, wherein the proportions of the training set, the validation set, and the test set are 7:2:1; S3 model training: training a quantum-inspired neural network fusion model on the training set, the quantum-inspired neural network fusion model using superposition and entanglement characteristics of quantum bits for multi-modal data feature extraction; S4 model optimization: optimizing parameters of the quantum-inspired neural network fusion model using an Adam optimization algorithm; S5 transfer learning: pre-training and fine-tuning the quantum-inspired neural network fusion model using transfer learning, wherein pre-training is performed on a public sewer pipeline detection data set, and fine-tuning uses sample data of a target domain; S6 model compression: pruning and quantizing the quantum-inspired neural network fusion model trained in S5, wherein pruning removes connections and neurons less than or equal to a threshold using an L1 regularization method, and quantization converts floating-point number parameters in the model to low-precision fixed-point number representations; S7 deployment and operation: deploying a parallel computing framework on a hardware platform of the sewer pipeline detection robot, using a CUDA programming model of a graphics processing unit to achieve parallel computing acceleration of the quantum-inspired neural network fusion model, and dynamically adjusting allocation of computing resources according to a motion state of the robot and complexity of a detection task.
2. The method according to claim 1, wherein, The feature extraction of the quantum-inspired neural network fusion model comprises mapping sample data to a quantum feature space to obtain a quantum state , the mapping function is , the sample data is , and the quantum state is .
3. The method according to claim 2, wherein, The quantum state is represented in the Bloch sphere as where is the Bloch representation of the positive qubit, is the Bloch representation of the negative qubit, and C is a correlation parameter; when the quantum state is in a direct product state, the quantum characteristic simplifies to .
4. The method according to claim 3, wherein, The mapping function τ is such that the quantum state retains maximum feature information of the sample data, an information loss function is defined as: ； where N is the number of samples, is the i-th original sample, is the i-th new sample reconstructed from the quantum state.
5. The method according to claim 1, wherein, The Adam optimization algorithm comprises the following sub-steps: Initialize model parameters: model initial weights First moment: Second moment: Number of iterations: Hyperparameters: learning rate Decay coefficient Numerical stability term ; Parameter iteration update, in each iteration process, the following operations are performed: The iteration counter is incremented, and the t value is increased by 1: ; Gradient calculation: based on current model parameters Calculate the loss function using randomly selected sample batches. For model parameters gradient ,in ; Updating the first moment: According to updating the first moment estimate; Updating the second moment: According to updating the second moment estimate, where is the square of the gradient; First moment bias correction: Calculate the first moment after correction ; Second moment bias correction: compute corrected second moment ; Parameter update: according to Update all weight parameters of the model.
6. The method of claim 1, wherein the method further comprises: In the transfer learning step, pre-training is performed on a public sewer pipeline detection data set to learn general pipeline features and patterns; during fine-tuning, all weight parameters from the input layer to the Kth layer in the pre-trained model are fixed, and only weight parameters from the K+1th layer to the output layer are updated; wherein K is 70% of the total number of layers of the model.
7. The method according to claim 1, wherein, The pruning uses L1 regularization, an L1 regularization loss function: wherein is a regularization strength coefficient, is an original loss function, M is the number of model weight parameters, is the jth weight parameter.
8. The method of claim 1, wherein the method further comprises: The quantization converts floating-point number parameters in the model to 8-bit integer or 16-bit floating-point number representations.
9. The method of claim 1, wherein, The parallel computing acceleration uses a CUDA programming model of an NVIDIA GPU to dynamically adjust allocation of GPU computing resources.
10. The method of claim 1, wherein, The master clock includes a field programmable gate array or an external synchronizer.